Process Systems Engineering Vol. 3 - Supply Chain Optimization, Part I (Wiley-VCH, 2008) PDF

Process Systems Engineering
Edited by
Efstratios N. Pistikopoulos,
Michael C. Georgiadis, and
Vivek Dua
Volume 3
Supply Chain Optimization, Part I
Volume Edited by
Lazaros G. Papageorgiou and
Michael C. Georgiadis
Related Titles
Haber, R., Bars, R., Schmitz, U. Bröckel, U., Meier, W., Wagner, G. (eds.)
Predictive Control in Process Product Design and Engineering
Engineering Best Practices
approx. 400 pages with approx. 760 pages in 2 volumes with 355 figures and
150 figures 60 tables
2007 2007
Hardcover Hardcover
ISBN: 978-3-527-31492-8 ISBN: 978-3-527-31529-1
Engell, S. (ed.) Ingham, J., Dunn, I. J., Heinzle, E.,

Logistic Optimization of Prenosil, J. E., Snape, J. B.
Chemical Production Chemical Engineering Dynamics

Processes An Introduction to Modelling and Computer
approx. 300 pages Simulation
2007
640 pages with 474 figures and 4 tables
Hardcover
2007
ISBN: 978-3-527-30830-9
Hardcover
ISBN: 978-3-527-31678-6
Dimian, A. C., Bildea, C. S.
Chemical Process Design Puigjaner, L., Heyen, G. (eds.)
Computer-Aided Case Studies Computer Aided Process and
approx. 500 pages with approx. 450 figures Product Engineering
2008 910 pages in 2 volumes with 281 figures
Hardcover and 77 tables
ISBN: 978-3-527-31403-4 2006
Hardcover
ISBN: 978-3-527-30804-0
Keil, F. J. (ed.)
Modeling of Process
Agachi, P. S., Nagy, Z. K., Cristea, M. V.,
Intensification Imre-Lucaci, A.
2007 Model Based Control
Hardcover Case Studies in Process Engineering
ISBN: 978-3-527-31143-9
2007
Hardcover
ISBN: 978-3-527-31545-1
Process Systems Engineering
Edited by
Efstratios N. Pistikopoulos, Michael C. Georgiadis, and
Vivek Dua
Volume 3: Supply Chain Optimization, Part I

Volume Edited by
Lazaros G. Papageorgiou and Michael C. Georgiadis
The Editors All books published by Wiley-VCH are
carefully produced. Nevertheless, authors,
Prof. Efstratios N. Pistikopoulos editors, and publisher do not warrant the
Centre for Process Systems Engineering information contained in these books,
Department of Chemical Engineering including this book, to be free of errors.
Imperial College London Readers are advised to keep in mind that
Roderic Hill Building statements, data, illustrations, procedural
South Kensington Campus details or other items may inadvertently be
London SW7 2AZ inaccurate.
United Kingdom
Library of Congress Card No.:
Dr. Michael C. Georgiadis applied for
Centre for Process Systems Engineering
Department of Chemical Engineering British Library Cataloguing-in-Publication Data
Imperial College London A catalogue record for this book is available
Roderic Hill Building from the British Library.
London SW7 2AZ
Bibliographic information published by
United Kingdom
the Deutsche Nationalbibliothek
Dr. Vivek Dua Die Deutsche Nationalbibliothek lists this
Centre for Process Systems Engineering publication in the Deutsche
Department of Chemical Engineering Nationalbibliografie; detailed bibliographic
University College London data are available in the Internet at
London WC1E 7JE <http://dnb.d-nb.de>.
United Kingdom
© 2008 WILEY-VCH Verlag GmbH & Co.
Dr. Lazaros G. Papageorgiou KGaA, Weinheim
Centre for Process Systems Engineering
All rights reserved (including those of
Department of Chemical Engineering
translation into other languages). No part of
University College London
this book may be reproduced in any form – by
London WC1E 7JE
photoprinting, microfilm, or any other
United Kingdom
means – nor transmitted or translated into a
machine language without written
permission from the publishers. Registered
names, trademarks, etc. used in this book,
even when not specifically marked as such,
are not to be considered unprotected by law.
Typesetting VTEX, Litauen

Printing Strauss GmbH, Mörlenbach
Binding Litges & Dopf buchbinderei
GmbH, Heppenheim
Cover Design Grafik-Design Schulz,
Fußgönheim
Printed in the Federal Republic of Germany

Printed on acid-free paper
ISBN: 978-3-527-31693-9
V
Contents
Preface – Volume 3: Supply Chain Optimization XIII

List of Authors XVII
1 Bridging the Gap Between Production, Finances, and Risk in

Supply Chain Optimization 1
L. Puigjaner and G. Guillén-Gosálbez
1.1 Introduction 1
1.2 Merging Process Operations and Finances: A Literature Review 1
1.3 Integrated Models for Tactical and Strategic Decisions in SCM 4
1.3.1 Financial Formulation: How Can we Take Optimal Financial
Decisions? 4
1.3.1.1 Budgeting Model 5
1.3.1.2 Objective Functions 8
1.3.2 Design – Planning Formulation 11
1.3.2.1 Mass Balance Constraints 11
1.3.2.2 Capacity and Facilities Location Constraints 12
1.3.3 Integration Between Models 13
1.3.4 Case Study 15
1.4 SCM Under Uncertainty 28
1.4.1 Mathematical Formulation 29
1.4.1.2 Capacity and Facilities Location Constraints 31
1.4.1.3 Objective Function Net Present Value (NPV) 31
1.4.2 Risk Management in SCM 31
1.4.2.1 Financial Risk 32
1.4.2.2 Downside Risk 34
1.4.2.3 Worst Case 35
1.4.3 Multiobjective Problem 35
1.4.4 Case Study 36
1.5 Conclusions 38
Supply Chain Optimization. Part I. Edited by Lazaros G. Papageorgiou and Michael C. Georgiadis
Copyright © 2008 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
ISBN: 978-3-527-31693-9
VI Contents
2 Multiagent Framework for Modeling and Design of Supply Chain

Networks 45
L. Puigjaner, G. Guillén-Gosálbez and F. D. Mele
2.1 Introduction 45
2.2 A Discrete Event Simulator for SCM 46
2.2.1 Simulation Model 46
2.2.2 The Demand Model 49
2.2.3 The Inventory Model 49
2.3 A Multiagent System for SCM 50
2.3.1 The Site Agent 53
2.3.2 The Sales Subagent 54
2.3.3 The Inventory Subagent 56
2.3.4 The Production Subagent 58
2.3.5 The Purchasing Subagent 59
2.3.6 The Transportation Subagent 61
2.4 Novel Business Aspects in SCM 62
2.4.1 Environmental Impact 62
2.4.1.1 The Environmental Module 63
2.4.2 Financial Concerns 67
2.4.2.1 Implementation: Financial Module 67
2.5 Multiagent System: Integrated Framework 68
2.6 Case Study 71
2.6.1 Objectives 78
2.6.2 Results 79
2.7 Concluding Remarks and Future Work 83
3 Environmentally Conscious Supply Chain Management 87

M. Türkay
3.1 Introduction 87
3.2 Closed-Loop Supply Chains 90
3.3 Environmentally Conscious Production 91
3.3.1 Process Models for Energy Production Systems 92
3.3.1.1 Boiler Models and Use of Biofuels 93
3.3.1.2 Turbine Models 94
3.3.1.3 Fuel Tank Models 94
3.3.1.4 Mixer Models 95
3.3.1.5 Environmental Considerations 95
3.3.1.6 Material Balance 96
3.3.1.7 Electricity Purchase 96
3.3.1.8 Operating and Startup Costs 96
3.3.1.9 Carbon Capture and Sequestering (CCS) Systems 96
3.3.1.10 Objective Functions 98
3.3.1.11 Illustrative Example 98
3.4 Sustainable Transportation 100
Contents VII
3.4.1 Intermodal Transportation 101

3.5 Conclusions 103
4 Optimal Design and Operational Planning of Responsive Process

Supply Chains 107
F. You and I. E. Grossmann
4.1 Introduction 107
4.2 Literature Review 108
4.3 Motivating Example 110
4.4 Problem Statement 111
4.5 Model 114
4.5.1 Network Structure Constraints 114
4.5.1.1 Supplier: Plant Site 114
4.5.1.2 Plant Site: Customer 115
4.5.1.3 Input and Output of a Plant 115
4.5.2 Operational Planning Constraints 117
4.5.2.1 Production Constraints 117
4.5.2.2 Sale Constraints 117
4.5.2.3 Upper Bound Constraints 117
4.5.3 Cyclic Scheduling Constraints 118
4.5.3.1 Assignment Constraints 119
4.5.3.2 Sequence Constraints 119
4.5.3.4 Timing Constraints 120
4.5.3.5 Cost Constraints 121
4.5.3.6 Upper Bound Constraints 121
4.5.4 Lead Time Definition 122
4.5.5 Nonnegative Constraints 125
4.5.6 Net Present Value 125
4.5.7 Solution Procedure 125
4.6 Numerical Example 126
4.7 Conclusions 128
5 Reverse Production Systems – Optimization Modeling to Support

Supply Chains for Product Recovery 135
T. Assavapokee, M. J. Realff and J. C. Ammons
5.2 Literature Review of Reverse Product System Modeling 136
5.3 Carpet and Carpet Recycling 137
5.4 Model Description 138
5.4.1 Objective Function 139
5.4.2 Mathematical Formulation 140
VIII Contents
5.5 Case Study Data and Scenarios 143

5.6 Results and Discussion 146
5.7 Conclusions 150
6 Optimal Design of Supply Chain Networks using Mathematical

Programming 157
P. Tsiakis, L. G. Papageorgiou and M. C. Georgiadis
6.2 Optimal Design of Production and Distribution Network 161
6.2.1 Notation 163
6.2.2 Mathematical Formulation of Deterministic Problem 163
6.2.2.1 Network Structure Constraints 163
6.2.2.2 Logical Constraints for Transportation Flows 164
6.2.2.3 Material Balances 164
6.2.2.5 Capacity of Distribution Centers 165
6.2.2.6 Objective Function 166
6.3 Distribution Center Location Case Study 168
6.3.1 Problem Description and Data Analysis 169
6.3.2 Production 169
6.3.3 Warehousing 170
6.3.4 Transportation 171
6.3.5 Demand 171
6.3.6 Results – Network Description 172
6.3.7 Comments 173
6.4 Optimal Production Allocation and Distribution Supply Chain
Networks 174
6.4.1 Problem Description – Data Analysis 175
6.4.1.1 Production Plants 175
6.4.1.2 Distribution Centers 175
6.4.1.3 Customer Zones 176
6.4.2 Network Optimization – Base Scenario 178
6.4.3 Fixed Product Allocation and Customer Assignment Scenario 180
6.4.4 Results Analysis 182
6.4.5 Comments 182
6.5 Conclusions 183
7 Design and Planning of Closed-Loop Supply Chains 187

A. P. Barbosa-Póvoa, M. I. Gomes Salema and A. Q. Novais
7.2 Reverse Logistics 188
7.2.1 Defining the Closed-Loop Supply Chain 189
Contents IX
7.3.1 Forward Supply Chain Models 192

7.3.2 Reverse Supply Chain Models 193
7.3.3 Closed-Loop Supply Chain Models 195
7.3.4 Research Agenda 197
7.4 Closed-Loop Supply Chains Modeling 198
7.4.1 Problem Description 198
7.4.2 Time Modeling 200
7.4.3 Model Formulation 201
7.4.3.1 Constraints 202
7.4.4 European Case 207
7.4.4.1 Results 208
7.5 Conclusions 212
8 Combining Strategic Design and Operative Planning in the Process

Industry 219
J. Kallrath
8.2 Combined Strategic and Operative Planning 222
8.2.1 Operative Planning 222
8.2.2 Multisite Design and Strategic Planning 223
8.2.2.1 The Need for Exact Optimization in Design Optimization 224
8.2.2.2 Design Planning under Uncertainty 226
8.2.3 Combined Approach: Design and Operative Planning 228
8.2.3.1 Motivation for the Combined Approach 228
8.2.3.2 The Choice of the Objective Function 229
8.2.3.3 Time Horizon and Time Resolution 230
8.2.3.4 Design Planning, Operative Planning, and Scheduling 230
8.2.4 The Need for an Integrated Tailor-Made Approach 230
8.2.4.1 Solution Approach 231
8.2.4.2 Software Design Issues 233
8.2.4.3 Limits of the Approach 233
8.3 Combined Strategic and Operative Planning – Commercial
Software versus Individual Models 234
8.3.1 Commercial Software for Operative Planning and Design
Planning 234
8.3.2 The Role of Individual Software for the CSOP Approach and
APS 236
8.3.3 When to Select Individual Tailor-Made Methods? 236
8.3.4 Developing and Maintenance Costs of the Inhouse Model 238
8.3.5 Dependences on Software Providers or Inhouse Modeling
Specialists? 239
8.4 Summary 239
X Contents
9 Chemical Supply Chain Redesign 245

P. K. Naraharisetti, I. A. Karimi and R. Srinivasan
9.2 Chemical Supply Chains 247
9.3 Capacity Management 249
9.3.1 Operations Management 250
9.3.2 Factors Influencing Investment Decisions 251
9.4 Supply-Chain Redesign 252
9.4.1 Nonmathematical Models 252
9.4.1.1 Strategic Planning 253
9.4.1.2 Tactical Planning 256
9.4.1.3 Operational Planning 257
9.4.2 Mathematical Models 258
9.4.2.1 Strategic Planning 258
9.4.2.2 Tactical Planning 259
9.4.2.3 Contracts for Material Supply 259
9.4.2.4 Remarks 259
9.4.3 This Chapter 260
9.5 Problem Description 261
9.5.1 Problem Statement 264
9.6 Problem Formulation 265
9.6.1 Capacity Change 265
9.6.2 Capacity Utilization 267
9.6.3 Material Balance 269
9.6.4 Supply Contracts 271
9.6.5 Loans and Bonds 272
9.6.6 Costs and Incomes 272
9.6.7 Remarks 276
9.7 Example 276
9.7.1 Results and Discussion 281
9.7.2 Future Work 291
9.8 Conclusion 292
10 A Hybrid Tabu-Branch and Bound Approach for the Solution of

Large-Scale Supply Chain Management Models 301
J. M. Natali, P. Chen and J. M. Pinto
10.2.1 Mathematical Structure of the CFPN Model 304
10.2.2 Preliminary Analysis of the CFPN Model Solution Approaches 304
10.3 Hybrid Tabu Search Procedure 305
10.3.1 Mixed Integer Linear Relaxation 306
10.3.2 Algorithmic Description 308
10.3.2.1 Unrestricted Fractional Solutions 309
Contents XI
10.3.2.2 Restricted Fractional Solutions 310

10.3.2.3 Tabu Search Guidance 310
10.3.3 General Algorithm 311
10.4.1 Algorithmic Analysis 314
10.4.2 Comparison to Full-scale Results 316
10.5 Conclusions 319
11 Enhancement of Oil and Gas Production via Network

Optimization 325
N. Thijssen, K.-J. Li and I. Mittendorff
11.2 Integrated Network Analysis 327
11.3 High-Level Network Model 328
11.3.1 Supply Module 328
11.3.2 Production Module 329
11.3.2.1 Separators 329
11.3.2.2 Compressors 331
11.3.2.3 Pumps 331
11.3.3 Pipelines Module 332
11.3.4 Demand Module 333
11.4 Workflow 333
11.5 Results 335
Index 339
XIII
Preface – Volume 3: Supply Chain Optimization
Modern industrial enterprises are typically multiproduct, multipurpose, and multi-

site facilities operating in different regions and countries and dealing with a global-
wide international clientele. In such enterprise networks, the issues of global enter-
prise planning, coordination, cooperation, and robust responsiveness to customer
demands at the global as well as the local level are critical for ensuring effective-
ness, competitiveness, business sustainability, and growth. In this context, it has
long been recognized that there is a need for efficient integrated approaches that
consider, in a systematic way, various levels of enterprise management, plant-wide
coordination and plant operation, in order to reduce capital and operating costs,
increase supply chain productivity and improve business responsiveness.
The supply chain concept has in recent years become one of the main approaches
to achieve enterprise efficiency. The terminology implies that a system view is taken
rather than a functional or hierarchical one. Enterprises cannot be competitive
without considering supply chain activities. This is partially due to the evolving
higher specialization in a more differentiated market. Most importantly, compe-
tition drives companies to reduced cost structures with lower inventories, more
effective transportation systems, and transparent systems able to support informa-
tion throughout the supply chain. A single company rarely controls the production
of a commodity as well as sourcing, distribution, and retail.
Many typical supply chains today have production that spans several countries
and product markets. The opportunities for supply chain improvements are large.
Costs of keeping inventory throughout the supply chain to maintain high customer
service levels are generally significant. There is a wide scope to reduce the inven-
tory while still maintaining the high service standards required. Furthermore, the
manufacturing processes can be improved so as to employ current working capital
and labor more efficiently.
It has widely been recognized that enhanced performance of supply chains ne-
cessitates: (i) appropriate design of supply chain networks and its components and
(ii) effective allocation of available resources over the network. Thus, in the last
few years, there has been a multitude of efforts focused on providing improve-
ments of supply chain management and optimization. These efforts include a wide
range of models: from commercial enterprise resource planning systems and so-
ISBN: 978-3-527-31693-9
XIV Preface – Volume 3: Supply Chain Optimization
called advanced planning systems to academic achievements (for example, linear

and mixed-integer programming, multiagent systems).
Management of supply chains is a complex task mainly due to the large size of
the physical supply network and inherent uncertainties. In a highly competitive
environment, improved decisions are required for efficient supply chain manage-
ment at strategic, tactical, and operational levels with time horizons ranging from
several years to a few days, respectively. Depending on the level, one or more of the
following decisions are taken:
• Number, size and location of manufacturing sites,
warehouses, and distribution centers.
• Network connectivity (e.g., allocation of suppliers to plants,
warehouses to markets, etc.).
• Production decisions related to plant production planning
and scheduling.
• Management of inventory levels and replenishment policies.
• Transportation decisions concerning mode of transportation
(e.g., road, rail, etc.) and also size of material shipment.
• Sustainability aspects (e.g., environmental impact
considerations, recycling policies etc.).
Most of the above challenging research issues are addressed in Volumes 3 and 4
of this book series.
Volume 3 of this book series focuses on strategic and tactical supply chain plan-
ning with emphasis on the process industries.
The need to incorporate financial considerations in the design and operation
of supply chains has long been recognized and it is addressed in Chapter 1 by
Puigjaner and coworkers. In the first part of their work, they present a systematic
mathematical programming framework that incorporates financial considerations
at the tactical and strategic supply chain decision-making levels. In the second part,
different techniques are described that aim to manage risk associated with the op-
eration of supply chain systems under uncertainty. In the subsequent chapter, the
same authors describe a new simulation-based modeling technique for taking sup-
ply chain management (SCM) decisions by utilizing software agents as building
blocks. The resulting agent-based system is a discrete event simulator that is able
to use a number of tools such as if-then rules and mathematical programming
algorithms. Furthermore, an alternative strategy for tackling SCM decision is de-
veloped, which is based on metaheuristic rules applied in the multiagent simulator.
The resulting tool is flexibile enough to accommodate further extensions and de-
tails of integrated supply chains.
Türkay in Chapter 3 presents an integrated approach to ensure an environmen-
tally conscious performance of supply chain systems in addition to the economi-
cal performance. Three different approaches are considered and modeled: (i) the
product-centric approach involves avoiding the use of environmentally harmful
materials in the product at the product design stage and the recovery and reuse
of the product after it has been consumed by the end user; (ii) the production-
system-centric approach considers the selection of raw materials and the design
Preface – Volume 3: Supply Chain Optimization XV
of the production systems for minimizing the environmental impact; and (iii) the
transportation-centric approach focuses on the use of different transportation sys-
tems that would reduce the environmental effects.
You and Grossmann consider in Chapter 4 the problem of optimal design and op-
erational planning of multiechelon, multisite process supply chain networks with
consideration of responsiveness and profitability. The problem is formulated as
a bicriterion optimization model in which the objectives are to maximize the net
present value and to minimize the lead time. This allows establishing trade-offs be-
tween the economics and responsiveness of the supply chain network. The model
produces a Pareto-optimal curve, which reveals how the optimal net present value,
and therefore the network structure of the supply chains, changes with different
specifications of lead time.
The work of Assavapokee, Realff, and Ammons described in Chapter 5 is moti-
vated by the increasing pressure to design large scale recovery and recycling sys-
tems for electronics, carpet, and other complex products with relatively low value
at their end-of-life. A supply chain mathematical programming model is proposed
for reverse production system design for carpet recycling. The overall modeling
and optimization approach can inform decision makers as to where to focus their
efforts in recruitment of sources of material and the overall target costs that are
achievable.
In Chapter 6, Tsiakis, Papageorgiou, and Georgiadis present the application of
a mixed-integer linear programming (MILP) model on two real examples where
the aim is to determine the optimal configuration of a production and distribution
network subject to operational and financial constraints. Operational constraints
include quality, production and supply restrictions, and are related to the allocation
of the production and the workload balance between production sites. Financial
constraints include production costs, transportation costs, and duties for the mate-
rial flowing within the network subject to exchange rates. As a business decision
the outsourcing of production is considered whenever the organization cannot sat-
isfy the demand.
The group of Barbosa and coworkers in Chapter 7 propose a generic modeling
approach for the design and planning of closed-loop supply chains. This considers
the simultaneous analysis of both strategic and tactical decisions where the system
structure is defined taking into account the planning of the related logistic activ-
ities. Several challenging problems in the area of reverse and closed-loop supply
chains are critically discussed.
Kallrath in Chapter 8 presents an integrated modeling framework addressing var-
ious conceptual issues in strategic, multisite design, and planning problems in the
process industry. The need for a combined design and operative planning model is
emphasized and efficient tailored-made modeling and solution techniques for real-
world problems are discussed. Finally, several problems for supply chain planning
and design in large companies focusing on the needs for appropriate modeling and
maintainable IT structures are highlighted.
Naraharisetti, Karimi, and Srinivasan in Chapter 9 present a novel MILP model
for making efficient capacity management and supply chain redesign decisions
XVI Preface – Volume 3: Supply Chain Optimization
for a multinational corporation. The model can provide the basis for obtaining the
best strategy for investment, involving a variety of real decisions such as facility
relocation, disinvestment, technology upgrade, raising capital through loans and
bonds, and hedging risk through signing contracts for material supply. The same
model can repeatedly be used to analyze various probabilistic scenarios and thus
make the choice of implementing one of the various resulting plans. A realistic
case study was used to illustrate the benefits that can be achieved by using the
developed model.
Pinto and coworkers in Chapter 10 present a hybrid algorithm for the solution
of mixed-integer linear programming problems in which a small subset of the dis-
crete variables account for most of the computational complexity of the model. The
algorithm builds upon the parametric tabu-search procedure proposed by Glover
and the use of a structured branch-and-cut (B&C) solver to harness the advantages
of both methods. The suggested procedure was applied to a continuous flexible
process network model and optimal solutions of small- to large-scale instances of
the model were obtained in a relatively small number of iterations.
In the final chapter, Thijssen, Li and Mittendorff present an integrated approach
for the optimal timing of developing new oil and gas fields. The overall framework
is capable of identifying future bottlenecks in the network and evaluating options
for removal of these bottlenecks. Several studies performed by the authors indicate
that such an integrated network analysis helps to bring production closer to the
technical potential of the wellhead platforms. It also provides significant insight
into the impact that all changes together have on the performance of the network
and the savings that can be achieved.
This collection represents a set of stand-alone works that captures recent re-
search trends in the development and application of techniques, methodologies,
algorithms, and tools for optimizing various aspects of supply chain systems. We
hope that by the end of the book, the reader will have developed a commanding
comprehension of the main aspects of integrated supply chains, the ability to criti-
cally access the key characteristics and elements related to the design and operation
of supply chains and the capacity to implement the new technology in practice.
We are extremely grateful to the authors for their outstanding contributions and
for their patience, which have led to a final product that far exceeded our expecta-
tions.
London, June 2007

L. G. Papageorgiou
M. C. Georgiadis
XVII
List of Authors
Jane C. Ammons Dr. Michael C. Georgiadis

Georgia Institute of Technology Imperial College London
School of Industrial and Systems Centre for Process Systems Engineering
Engineering Department of Chemical Engineering
Atlanta, GA 30332-0295 Roderic Hill Building
USA South Kensington Campus
London SW7 2AZ
Tiravat Assavapokee UK
University of Houston
Department of Industrial Engineering Maria Isabel Gomes Salema
Houston, TX 77204-4008 Centro de Matemática e Aplicações
USA FCT, UNL
Monte de Caparica
Professor Ana Paula Barbosa-Póvoa 2825-114 Caparica
Centro de Estudos de Gestão Portugal
IST, UTL
Professor Ignacio E. Grossman
Av. Rovisco Pais
Carnegie Mellon University
1049-101 Lisboa
Portugal
5000 Forbes Avenue
Pittsburgh, PA 15213-3890
Peter Chen
USA
Polytechnic University
Othmer-Jacobs Department of Chemical Dr. Gonzalo Guillén-Gosáblez
and Biological Engineering Universtitat Politècnica de Catalunya
6 Metrotech Center Chemical Engineering Department
Brooklyn, NY 11201 Escola Tècnica Superior d’Enginyeria
USA Industrial
Avda. Diagonal 647
E-08028 Barcelona
Spain
ISBN: 978-3-527-31693-9
XVIII List of Authors
Professor Josef Kallrath João M. Natali

BASF Aktiengesellschaft Othmer-Jacobs Department of Chemical
GVC/S-B009 and Biological Engineering
D-67056 Ludwigshafen Polytechnic University
Germany 6 Metrotech Center
and Brooklyn, NY 11201
University of Florida USA
Dept. of Astronomy
Gainesville, FL 32611 Dr. Augusto Q. Novais
USA Departmento de Modelação e Simulação
INETI
Professor Iftekhar A. Karimi Est. do Paço do Lumiar
National University of Singapore 1649-038 Lisboa
Department of Chemical and Portugal
Biomolecular Engineering
4 Engineering Drive 4 Dr. Lazaros G. Papageorgiou
Singapore 117576 University College London
Singapore Centre for Process Systems Engineering
Kuei-Jung Li
Torrington Place
Shell Global Solutions International BV
London WC1 7JE
P.O. Box 541
UK
2501 CM
The Hague Professor Jose M. Pinto
The Netherlands Polytechnic University
Fernando Daniel Mele Othmer-Jacobs Department of Chemical
Universtitat Politècnica de Catalunya and Biological Engineering
Chemical Engineering Department 6 Metrotech Center
Escola Tècnica Superior d’Enginyeria Brooklyn, NY 11201
Industrial USA
Avda. Diagonal 647
Professor Dr. Luis Puigjaner
E-08028 Barcelona
Universtitat Politècnica de Catalunya
Spain
Chemical Engineering Department
Ilse Mittendorff Escola Tècnica Superior d’Enginyeria
Shell Global Solutions International BV Industrial
P.O. Box 541 Avda. Diagonal 647
2501 CM E-08028 Barcelona
The Hague Spain
The Netherlands
Professor Matthew J. Realff
Pavan Kumar Naraharisetti
Georgia Institute of Technology
Institute of Chemical and Engineering
School of Chemical Engineering
Sciences
Atlanta, GA 30332-0100
1 Pesek Road, Jurong Island
USA
Singapore 627833
Singapore
List of Authors XIX
Professor Rajagopalan Srinivasan Dr. Panagiotis Tsiakis

Institute of Chemical and Engineering IBM UK Ltd
Sciences 76–78 Upper Ground, South Bank
1 Pesek Road, Jurong Island London SE1 9PZ
Singapore 627833 UK
Singapore
and Professor Metin Türkay
National University of Singapore Koc University
Department of Chemical & College of Engineering
Biomolecular Engineering Rumelifeneri Yolu
4 Engineering Drive 4 Sariyer 34450 Istanbul
Singapore 117576 Turkey
Singapore Dr. Fengqi You
Dr. Nort Thijssen Carnegie Mellon University
Shell Global Solutions International BV Department of Chemical Engineering
P.O. Box 541 5000 Forbes Avenue
2501 CM Pittsburgh, PA 15213-3890
The Hague USA
The Netherlands
1
1
Bridging the Gap Between Production, Finances, and Risk in
Supply Chain Optimization
Luis Puigjaner, Gonzalo Guillén-Gosálbez
1.1
Introduction
The tight profit under which chemical process industries (CPI) operate are forcing
companies to pay more and more attention to the design and operation of their
supply chains (SCs). Traditional approaches available in the process systems engi-
neering (PSE) literature to address the design and operation of chemical SCs focus
on the process operations side and neglect the financial part of the problem. This
simplification may have undesirable negative impacts in the firm, as it may lead to
suboptimal or even infeasible overall plans for the whole SC.
This chapter presents novel ideas and tools to enhance the SCM that rely on the
incorporation of financial concerns at the tactical and strategic SC decision-making
levels. Specifically, the integration between diverse functional disciplines in SCM
is carried out through the use of holistic models devised to optimize the combined
effects of process operations and finances. Furthermore, the numerous sources of
technical and commercial uncertainty affecting the SC operation further increases
the complexity of the problem. Thus, the second part of this chapter is devoted to
describe different techniques that aim to manage the risk associated with the SCM
under uncertainty. The main advantages of both approaches are illustrated through
a case study of a SC comprising several plants, warehouses and markets.
1.2
Merging Process Operations and Finances: A Literature Review
Recent advances in PSE have focused on devising enterprise wide modeling and
optimization strategies that integrate decisions of distinct functions of a business
into a global model. Nevertheless, despite the effort made in the area, almost all of
the models developed to date focus on the process operations side and neglect the
financial variables and constraints associated with cash flows.
ISBN: 978-3-527-31693-9
2 1 Bridging the Gap Between Production, Finances, and Risk in Supply Chain Optimization
However, the effective control of cash is one of the most important requirements
of financial management and its steady and healthy circulation throughout the en-
tire business operation has repeatedly been shown to be the basis of business sol-
vency [1]. In fact, the availability of cash governs the production decisions taken in
a company. A production plan cannot be implemented if it violates the minimum
cash flow imposed by the firm (i.e., if the amount of raw materials and/or utilities
required cannot be purchased due to a temporary lack of cash).
Moreover, assessing the feasibility of the scheduling/planning decisions from
a financial point of view may not be enough for companies that want to achieve
a competitive advantage in the marketplace. Fierce competition in today’s global
markets is forcing companies to perform further analyses in order to find the best
production-distribution decisions to be carried out in their SCs. If they wish to re-
main competitive, it is essential that they properly assess the different process op-
erations alternatives in terms of their ability to markedly improve the value of the
company. In fact, maximum-profit or minimum-cost decisions may lead to poor
financial results if their financial impact is not properly assessed prior to their
being implemented. Thus, managers should extend their analysis to include the
more general objective of maximizing the shareholder value (SHV) of the firm as
opposed to the common optimization of traditional myopic key performance indi-
cators (KPIs) such as cost or profit.
We believe that companies can create more value and achieve better performance
by devising integrated approaches for SCM. Such strategies should be capable of
holistically optimizing the combined effects of process operations and finances,
thus exploiting the synergy between different management disciplines.
The approaches that currently exist, however, still address the overall problem
in a sequential fashion through the optimization of partial KPIs. The use of these
traditional sequential procedures applied in the operation and design of chemical
processes is mainly motivated by the functional organizational structures of the
firms. Today, most companies have separate departments for production, supply,
logistics, service to customers, etc. In an environment of this type, each functional
area’s plan is sequentially considered as input to the others according to a hierarchy.
Thus, the models supporting the decision making in batch chemical processes op-
erate in an isolated way. They optimize partial sets of decision variables but they do
not lead to real integration relationships despite promoting the sharing of informa-
tion between different business entities [2]. This partitioning of decision making
in companies has been reflected in the goals of the studies and the optimization
models developed to support them.
Despite the fact that optimization models seem to offer an appealing framework
for analyzing corporate financial decisions and constraints, as well as for integrat-
ing them with process operation decisions and constraints [3], relatively few in-
tegrated corporate financial models have been implemented so far [4–6]. Thus,
although planning and scheduling models for SCM should improve an overall
business performance measure, they usually neglect financial issues and pursue
a myopic key performance indicator as the objective to be optimized.
1.2 Merging Process Operations and Finances: A Literature Review 3
Table 1.1 Supply chain models for the CPI.
Performance measures
Economic Customer Efficiency Environmental
basis service basis basis basis
Material Consumption
Energy Consumption
Facilities utilization
Waste Generation
Eco-indicator 99
Inventory Level
Financial Risk
Net Revenue
Makespan
Tardiness
Earliness
Cleanup
Authors
Equity
Setup
Profit
NPV
Cost
CSL
[11] X
[12] X
[13] X
[14] X
[15] X
[16] X X X X X X
[17] X
[18] X
[19] X X X
[20] X
[21] X
[22] X
[23] X
[24] X
[25] X X X X X X
[26] X
[27] X
[28] X
[29] X
[30] X
[31] X
[32] X
[33] X
[6] X
[34] X X X
[35] X
[36] X X
[37] X X
[38] X X
[39] X
[40] X X
[41] X
[42] X
[43] X
Table 1.1 shows the results of a review made concerning the type of approaches
and in particular the specific KPIs applied so far in SCM in the CPI. The objec-
tive functions have been classified into four main categories: (1) economic basis,
(2) customer service basis, (3) efficiency basis, and (4) environmental basis. The
review shows that profit and cost are by far the most exploited indicators. Thus,
it is obvious that an opportunity naturally exists for practitioners and researchers
that aim to contribute to the fulfillment of this important gap in the literature that
is currently waiting for further attention. In fact, with the recent advances in op-
timization theory and software applications there is no apparent reason why we
cannot contribute to this field by constructing models for SCM merging concepts
from diverse areas. This need to extend the studies and analysis of process oper-
ations to incorporate financial considerations has been widely recognized in the
literature [3, 7–10], and we believe that will be the focus of intensive research in the
future.
1.3
Integrated Models for Tactical and Strategic Decisions in SCM
As mentioned before, the application of integrated models for SCM is crucial to en-
hance the SCM. This section provides guidelines to construct holistic models for
SCM that merge concepts and ideas from process operations and finances. Specifi-
cally, this chapter focuses on developing quantitative tools for SCM based on mixed
integer modeling techniques. Thus, standard mathematical formulations for SCM
are enriched through the inclusion of financial variables and constraints.
We will first present a mathematical formulation suitable to optimize the fi-
nancial decisions in the context of SCM. Then, a detail description of the steps
required to connect such formulation with standard models that address tactical
and/or strategic decisions in SCM will be given.
1.3.1
Financial Formulation: How Can we Take Optimal Financial Decisions?
To make optimal financial decisions in the context of SCM we need to properly

manage the cash flows associated with the operation of the network. This can
be accomplished by furnishing practitioners with budgeting or cash management
models that act as decision-support tools in the financial area. Such formulations
allow all the variables impacting the cash flows of the company to be optimized si-
multaneously. Different budgeting models have been devised to date. Specifically,
a number of budgeting models appeared in the literature of the late 1950s, when
linear programming computation methods also emerged. Srinivasan [44] presents
a review of such deterministic cash management models. The impact that these
formulations have in the overall performance of the network can be indeed en-
hanced through their use in conjunction with process operations models. Indeed,
the combined use of both formulations results in integrated tools capable of judi-
ciously exploring and optimizing the combined effects of process operations and
finances.
Specifically, in our approach, the management of cash associated with the opera-
tion of the SC is analyzed by extending the mathematical formulation developed by
Orgler [45]. Such enhanced formulation is then connected to the process operations
variables and constraints through the purchases of raw materials and utilities from
suppliers and the sales of final products to customers. As a result of the application
of this integrated model, optimal process operations and financial decisions can be
computed simultaneously. Therefore, payments to providers, short-term borrow-
ing, pledging decisions and transactions of marketable securities are scheduled
along with the manufacturing and distribution tasks accomplished in the nodes of
the network. The specific set of variables and constraints of the financial model are
next described in detail.
1.3.1.1 Budgeting Model

The budgeting variables and constraints of the financial model should be deter-
mined according to the specific applicable rules (depreciation), legislation (taxes),
etc. This may lead to many different formulations depending on the specific case
being analyzed. To circumvent this issue, our model includes a set of general equa-
tions that can be easily adapted to any particular situation.
The budgeting model considers that the whole time horizon is divided into t time
periods or equal length. The cash balance for each planning period is the following:

t
Casht = Casht−1 + ECasht + NetCLine
t − Payet t
e t =1
(1.1)
−FCostt + NetMS
t − FPayt + Capitalt + NetLDebt
t + Othert ∀t.
Equation (1.1) describes the cash for each period t (Casht ). Here, the cash at
every time period t (Casht ) is a function of the previous cash (Casht−1 ), the exoge-
nous cash from the sales of products or, in general, from any other inflow of cash
(ECasht ), the amount borrowed or repaid to the short-term credit line (NetCLine
t ), the
raw materials, production and transport payments on accounts payable incurred in
any previous or actual period t (Payett ), the payments of the fixed cost (FCostt ), the
sales and purchases of marketable securities (NetMS t ), the amount invested on facil-
ities (FPayt ), the capital supported by the shareholders of the company (Capitalt ),
the amount borrowed or repaid to the long-term credit line (NetLDebt t ) and finally
other expected outflows or inflows of cash (Othert ).
A certain proportion of the accounts receivable may be pledged at the beginning
of a period. Pledging is the transfer of a receivable from the previous creditor (as-
signor) to a new creditor (assignee). When a firm pledges its future receivables, it
receives in the same period only a part, normally 80%, of their face value. Thus,
it can be assumed that a certain proportion of the receivables outstanding at the
beginning of a period is received during that period through pledge, as stated by

Eq. (1.2).

t−1
t−1
Pledtt ASalest t ∀t. (1.2)
t =t−α max t =t−α max
In this equation the variable Pledtt represents the amount pledged within pe-
riod t on accounts receivable maturing in period t , while ASalest t represents the
accounts receivable associated with the sales of products executed in period t and
maturing in t. Here, the parameter α max denotes the maximum maturing period
of the accounts receivable.
Let us note that pledging represents a very expensive way of getting cash that will
only be used when no more credit can be obtained from the bank.
max

t
t−1
t+α
ECasht = ASalest t − Pledt t + φt t · Pledtt ∀t.
t =t−α max t =t−α max t =t+1 (1.3)
Finally, the exogenous cash is computed by means of Eq. (1.3) as the difference
between the accounts receivable maturing in period t and incurred in previous
periods t (ASalest t ) minus the amount of receivables pledged in previous periods
on accounts receivable maturing in period t plus the amount pledged in the actual
period on accounts receivable maturing in future periods. In this expression, φt t
represents the face value of the receivables being pledged.
CLinet CLinemax ∀t, (1.4)
CLinet = CLinet−1 (1 + η) + Borrowt − Repayt ∀t, (1.5)
NetCLine
t = Borrowt − Repayt ∀t, (1.6)
Repayt η · CLinet−1 ∀t. (1.7)
A short-term financing source is represented by an open line of credit with a
maximum limit imposed by the bank (Eq. (1.4)). Under an agreement with the
bank, credits can be obtained at the beginning of any period and are due after one
year at a given interest rate (η) that depends on the specific agreement reached
with bank. Equations (1.5) and (1.6) make a balance on borrowings, considering
for each period the updated debt (CLinet−1 ) from the previous periods, the balance
between borrows and repayments (NetCLine t ) and the interest of the credit line (η ·
CLinet−1 ). The bank requires a compensating balance, normally higher than the
20% of the amount borrowed. Therefore, the minimum cash (MinCash) has to be
higher than the compensating balance imposed by the bank. Moreover, the bank
regularly requires a repayment greater than or equal to the interests accumulated
in previous periods, as it is stated by Eq. (1.7).
e
t+γ
Payett βett = EPurchet ∀e, t T − γe . (1.8)
t =t
With respect to the accounts payable, which are due to the consumption of raw
materials, production and transport services, let us note that our formulation as-
sumes that the financial officer, at his option, may stretch or delay payments on
such accounts. Discounts for prompt payment can be obtained if purchases are
paid in short time and cannot be taken if the payments are stretched. The model
considers a maximum delay in the payments associated with a external supplier e
(γe ). Thus, Eq. (1.8) forces the payments executed in period t on accounts payable
to supplier e incurred in period t to equal the total amount due. In this expression,
technical coefficients (βett ) that multiply the payments executed in periods t on
accounts payable incurred in t , are introduced in the formulation in order to take
into account the terms of the raw materials, production and transport credits (i.e.,
2 percent-one week, net-28 days).
e
t+γ
Payett βett EPurchet ∀e, t > T − γe . (1.9)
t =t
The payment constraints belonging to the last periods of time are formulated as
inequalities (Eq. (1.9)), as it is not reasonable to require that total accounts payable
be zero at the end of the planning period.
prod
EPurchet = Purchrm tr
et + Purchet + Purchet ∀e, t. (1.10)
The external purchases from supplier e at every period t (EPurchet ), which are
computed through Eq. (1.10), include the purchases of raw materials and transport
and production utilities.

T
T
NetMS
t = StMS − ϒtMS
t + ZtMS
t
t =t+1 t =t+1
t−1
(1.11)
t−1
+ 1 + DttMS
ϒttMS − 1 + EttMS
ZttMS
∀t,
t =1 t =1
−1

t t
ZttMS
1 + EttMS
StMS + 1 + DttMS
ϒttMS
∀t, t < t. (1.12)
t =1 t =1
Equation (1.11) makes a balance for marketable securities. The portfolio of mar-
ketable securities held by the firm at the beginning of the first period includes
several sets of securities with known face values maturing within the time horizon
(StMS ). All marketable securities can be sold prior to maturity at a discount or loss
for the firm, as stated by Eq. (1.11). Revenues and costs associated with the trans-
actions in marketable securities are given by technical coefficients (DttMS MS
and Ett ).
ϒt t is the cash invested in period t on securities maturing in period t . Zt t is the
MS MS
cash income obtained through the security sold in period t maturing in period t .
Equation (1.12) is applied to constraint in each period the total amount of mar-
ketable securities sold prior to maturity to be lower than the available ones (those
belonging to the initial portfolio plus the ones purchased in previous periods mi-
nus those sold before).
FAssett = LBorrowt + Capitalt ∀t. (1.13)
Equation (1.13) balances the investment with the capital supported by sharehold-
ers (Capitalt ) and the amount borrowed from banks as long term debt (LBorrowt )
at each time period t.
LDebtt = LDebtt−1 (1 + irt ) + LBorrowt − LRepayt ∀t, (1.14)
NetLDebt
t = LBorrowt − LRepayt ∀t, (1.15)
LRepayt irt · CLinet−1 ∀t. (1.16)
Equations (1.14) to (1.16) reflect the payment conditions associated with the long
term debt. Let us note that these constraints are similar to those associated with
the short term credit line, as in practice both debts can be treated in a similar way.
Nevertheless, notice that in the case of the long term debt, the amount repaid in
each time period (LRepayt ) remains usually constant in every time period.
Casht MinCash ∀t. (1.17)
Equation (1.17) limits the cash in each period (Casht ) to be larger than a mini-
mum value (MinCash). A minimum cash is usually required to handle uncertain
events, like delays in customer payments, thus ensuring enterprise liquidity. More-
over, this minimum cash is usually required by the banks, which often demand
compensating balances to the companies.
1.3.1.2 Objective Functions

The most common objectives used by the PSE community are makespan, due date
fulfillment, profit, net present value (NPV), internal rate of return (IRR), and pay-
back time. On the other hand, in the financial area, the financial business decisions
have been traditionally assessed in terms of other ratios such as market-to-book
value, liquidity ratios, leverage, capital structure ratios, return on equity, sales mar-
gin, turnover ratios and stock security ratios, among others.
Nevertheless, nowadays the maximization of the shareholder’s value (SHV) of
the firm seems to be the main priority of the firms and what really drives their
decisions. The use of SHV as the objective to be maximized is mainly motivated
by the fact that it reflects in a rather accurate way the capacity that the company
has to create value. The SHV of the firm can be indeed improved by maximizing
its corporate value (CV). Specifically, according to Weissenrieder [46], the market
value of a company is a function of four factors: (1) investment, (2) cash flows, (3)
economic life, and (4) capital cost. Specifically, our work applies the discounted-
free-cash-flow method (DFCF) to compute the corporate value of a company.
The Discounted-Free-Cash-Flow Method (DFCF) Our strategy applies the discounted-

free-cash-flow method (DFCF) to assess the decisions undertaken by a firm. This
method has recently become the most preferred approach for the valuation of com-
panies given its capacity of properly assessing the four main factors that contribute
to create the market value of a firm. In fact, the DFCF method is well entrenched in
finance theory and its use is gaining wider acceptance in industrial scenarios. The
DFCF method values a project or an entire company by determining the present
value of its future cash flows and discounting them taking into account the appro-
priate capital cost during the time horizon for which it is defined [47] (economic
life).
According to financial theory, the enterprise market value of a firm is given by the
difference between the discounted stream of future cash flows during the planning
horizon and the net total debt at the end of its life time (NetDebtT ), as it is stated
by constraint (1.18). The final total debt includes both, the short and the long term
debt and also the cash (Eq. (1.19)).
CV = DFCF − NetDebtT , (1.18)
NetDebtT = CLineT + LDebtT − CashT . (1.19)
In the calculation of the DFCF, one must discount the free cash flows of each
period t at a rate equivalent to the capital cost (see Eq. (1.20)). The capital cost
reflects the time value of the money and also the risk of the investment. In fact,
the capital cost can be regarded as the expected return required to attract funds to
a particular investment [48].

T
FCFt
DFCF = . (1.20)
(1 + WACCt )t
t=0
The capital cost can be determined through the weighted average method. This
method considers the total capital structure of the company, including the overall
equity and the debt, as it is shown in Eq. (1.21). In this expression, λt denotes the
proportion of equity over the total capital investment.
WACCt = λt · E(ROE) + irt · (1 − λt ) · (1 − trate) ∀t, (1.21)
E(ROE) = r 0 + rp. (1.22)
To compute the expected return on equity, which is denoted by E(ROE), we ap-

ply Eq. (1.22). In this expression, E(ROE) is computed as the sum of a risk free
rate r 0 and a risk premium (rp). The former term represents the rate of return of
an investment free of default risk available in the market and is usually equal to
the yield to maturity offered by a government security. The latter, represents the
expected amount of return above the risk-free rate in exchange for a given amount
of variance [7, 48]. One of the most commonly employed methods to estimate the
risk premium is the capital asset pricing model (CAPM) (for more details regarding
this topic the reader is referred to Sharpe [49]).
FCFt = Profitt (1 − trate) − NetInvestt − NWCt ∀t. (1.23)
Cash flows at every period t (FCFt ) are given by the profit after taxes, net change
in investments and change in net working capital. Specifically, the free cash flows
are the difference between the net operating profit after taxes (NOPAT) and the in-
crease in capital invested. From this definition it follows that there will be value cre-
ation if the incoming value (Profitt (1 − trate)) is greater than the consumed value
(NWCt − NetInvestt ) as shown in Eq. (1.23).

Profitt = ESalest − EPurchet + FCostt − Invt ∀t. (1.24)
e
Equation (1.24) is applied to compute the profit at each period t from the in-
comes associated with the sales of final products (ESalest ), the production costs,
the cost of the transport services (EPurchet ), fixed costs (FCostt ), and the change in
inventory(Invt ).
NetInvestt = FAssett − Dept ∀t. (1.25)
The net investment at each period t represents the monetary value of the fixed
assets acquired in that period minus the depreciation. As mentioned before, the
depreciation term should be computed according to the specific applicable rules
(Straight Line, Sum-of-years Digits, Declining Balance, etc.)

NWCt = ARect + Invt − APayt + FExt ∀t. (1.26)
The change in net working capital associated with period t (NWCt ) is computed
from the change in accounts receivables, plus the change in inventory, minus the
change in accounts payable, plus any other financial expenses or incomes (FExt ),
as stated by Eq. (1.26).

t t +α

max

t+α max

t
ARect = ASalest t − Pledt t ∀t, (1.27)
t =t−α max +1 t =t+1 t =t+1 t =t −α max
ARect = ARect − ARect−1 ∀t. (1.28)

Equation (1.27) computes the accounts receivables corresponding to period t
from the sales executed in the actual or any earlier period and maturing in periods
beyond the present one minus the receivables pledged in previous periods, includ-
ing those associated with the actual one. Equation (1.28) determines the change in
accounts receivable at time period t.

t
t
t
APayt = EPurchet − βet t · Payet t ∀t, (1.29)
e t =1 e t =1 t =t
APayt = APayt − APayt−1 ∀t. (1.30)

The accounts payable (APayt ) are determined as the difference between all the
purchases executed in previous periods (EPurchet ) minus all payments done until
the actual period (Payet t ), as stated in Eq. (1.29). This constraint also takes into
account the discounts for prompt payments (βet t ). The change in accounts payable
at time period t is thus represented by constraint (1.30).
max

t+α
t

FExt = (1 − φt t )Pledt t − Payet t Coefet t − 1
t =t+1
e t =1 (1.31)
+ EttMS
· ZttMS
− DttMS MS
· ϒtt ∀t.
t t
Finally, Eq. (1.31) computes other financial expenses and incomes (FExt ) associ-
ated with the SC operation at every time period t. This term includes pledging
costs

((1 − φt t )Pledt t ), discounts for prompt payments to suppliers (Payet t Coefet t − 1 ),
and also the earnings (DttMS · Ytt ) and expenses (Ett · Ztt ) associated with the
MS MS MS
transactions of marketable securities.
1.3.2
Design – Planning Formulation
As mentioned before, integrated models for SCM can be constructed by merging fi-

nancial and process operations variables and constraints. Thus, the integration be-
tween finances and process operations can be achieved by linking the cash manage-
ment formulation previously presented with a mathematical model reflecting the
operation of the SC. Particularly, in this section we connect the budgeting model
with a design/planning formulation based on the work of Hugo and Pistikopou-
los [37]. The structure of the SC taken as reference to develop such mathematical
formulation is illustrated in Fig. 1.1. The model has been adapted to allow the stor-
age of products and to include distribution center nodes in the SC network. We
first describe, in detail, the equations of the process operations model and then
illustrate the way in which it can be linked with the cash management formulation
presented in the previous section.
1.3.2.1 Mass Balance Constraints

The mass balance must be satisfied in each of the nodes that integrate the SC
network.
Manufacturing Sites Equation (1.32) represents the mass balance for each raw ma-
terial r consumed in each manufacturing site s in every time period t. Thus, this
equation states that the purchases of raw material r provided by suppliers e (Er )
plus the initial inventory kept at the site must equal the inventory at the end of the
period t plus the quantity consumed by the manufacturing tasks. Let us note that
RAW SITES & DISTRIBUTION

MARKETS
MATERIALS TECHNOLOGIES CENTERS
Site 1
Technology 1 Pij
st
DC
1 Market
1
st
hr
Technology j
rc
Pu
Raw
st
Qiw
Materials Site s
Technology 1
DC
w Salesiwmt Market
m
Technology j
Fig. 1.1 Supply chain structure taken as reference.

this equation should be only applied to those equipments consuming raw material
r (i ∈ (Ij ∩ Ir )).

Purchrm erst + SIrst−1 = SIrst + µrij · Pij st ∀r, s, t. (1.32)
e∈Er j i∈(Ij ∩Ir )
The mass balance for final products i in each manufacturing site s is enforced
via Eq. (1.33). This expression states that the amount of final product manufactured
at each site during a given time period t plus the initial stock of the product must
equal the final inventory of the product plus the amount transported from the site
to the distribution centers w.
fp
Pij st + SOist−1 = SOist + Purchiwst ∀i, s, t. (1.33)
j ∈Ji w
Distribution Centers Equation (1.34) expresses the mass balance for the distribution
centers w. Thus, this equation states that the total amount of final product i coming
from all the sites s plus the initial inventory of the product kept at the distribution
center must equal the final inventory plus the sales in the final markets m.

Qiwst + SWiwt−1 = SWiwt + Salesiwmt ∀i, w, t. (1.34)
s m
Unlike other works in the literature, our model assumes that part of the demand
can actually be left unsatisfied due to limited production capacity. Thus, Eq. (1.35)
forces the sales of product i carried out in market m during time period t to be less
than or equal to the demand.

Salesiwmt Demimt ∀i, m, t. (1.35)
w
1.3.2.2 Capacity and Facilities Location Constraints

Here we define two different variables, FSj st and FWwt , which represent the total
capacity of manufacturing sites s and distribution centers w, respectively, during
time period t. Let us note that our model is general enough to address not only the
design of a new SC, but also the retrofitting of an existing network. In the latter
case, the problem should be formulated by fixing at the initial time period (t = 0)
the value of the variables representing the capacity of the facilities according to
the topology of the original network. Furthermore, variables FSEj st and FWEwt
denote the expansion in capacity of different facilities of the network during time
period t. Thus, for those nodes in which the company have not invested so far, the
establishment of a new facility takes place in the first time period in which these
variables take a nonzero value.
Equations (1.36) and (1.37) are added to control the changes in the capacities of
the facilities over time. These constraints include the binary variables Vj st and Xwt ,
which take a value of 1 if the facility being represented (either the technology j at
manufacturing site s or the distribution center w) is expanded in capacity and zero
otherwise. Let us note that the increments in capacity are bounded in the range
[FSEL j st , FSEj st ], which represent the realistic physical interval in which they must
U
fall.
Vj st · FSEL U
j st FSEj st Vj st · FSEj st ∀j, s, t, (1.36)
Xwt · FWEL U
wt FWEwt Xwt · FWEwt ∀w, t. (1.37)
Equations (1.38) and (1.39) are added to update the total capacity by the increased
amount during time period t.
FSj st = FSj st−1 + FSEj st ∀j, s, t, (1.38)
FWwt = FWwt−1 + FWEwt ∀w, t, (1.39)
plant
νij · Pij st FSj st−1 ∀j, s, t, (1.40)
i∈Ij

ware
νiw · SWiwt FWwt−1 ∀w, t. (1.41)
i
Constraints (1.40) and (1.41) force the capacity utilized in each facility to be lower
plant ware reflect the amount of ca-
than the installed one. The parameters νij and νiw
pacity consumed per unit of product.
Finally, the model assumes a maximum availability of raw materials. Thus, Eq.
(1.42) forces the amount of raw material r purchased to supplier e at each time pe-
riod t to be lower than an upper bound given by physical limitations (Aert ). In this
expression, Re denotes the set of raw materials that can be provided by supplier e.

Purchrm
erst Aert ∀e, r ∈ Re , t. (1.42)
s
1.3.3
Integration Between Models
The integration between both formulations is carried out through the sales of
products, the purchases of raw materials, transport services and utilities to final
providers, the fixed cost associated with the operation of the network, and the total
investment in capital. Thus, the accounts receivable incurred in any period t and
maturing in period t can be easily computed from sales of products executed in
period t, the fraction of these sales that will be collected in period t and the prices
of the products sold, as it is stated in Eq. (1.43). Here, δmtt denotes the fraction of
sales carried out in market m in period t that will be paid in period t .

ASalestt = Salesiwmt · δmtt · P riceimt ∀t, ∀t > t, (1.43)
i w m
prod
EPurchet = EPurchrm tr
et + EPurchet + EPurchet ∀e, t. (1.44)
The external purchases from supplier e at every period t (EPurchet ), which are
computed through Eq. (1.44), include the purchases of raw materials and transport
and production utilities.
The external purchases of raw materials to supplier e at every period t (EPurchet )

can be then computed through Eq. (1.45).

EPurchrm
et = Purchrm
erst · πert ∀e, r, t. (1.45)
r s
Furthermore, constraints (1.46) to (1.49) can be added to model the quantity dis-
counts, i.e., price reductions offered by the suppliers to induce large orders. The
inclusion of these constraints allows the potential benefits of reduced purchase
prices and fewer orders to be traded-off against the increase in inventory costs.
Specifically, we define a set of discounts intervals Der for each raw material r for
which supplier e offers quantity discounts (r ∈ DRe ). Each interval corresponds
to a different discount factor ψerd . The limits of interval d ∈ Der are denoted as
rm
Purchertd−1 and Purchertd . We introduce a new set of binary variables Fertd that
take the value of 1 if the amount of raw material r purchased from supplier e in
period t by all the sites falls into discount interval d, and 0 otherwise:
rm rm
1 if Purchrmert ∈ [Purchertd−1 , Purchertd ]
Fertd =
0 otherwise
To enforce the above definition, the following linear constraints are applied [33]:

Fertd = 1 ∀e ∈ DRe , r, t, (1.46)
d∈Der
rm rm
Fertd−1 · Purchertd−1 Purchrm
ertd Fertd · Purchertd (1.47)
∀e ∈ DRe , r, t, d ∈ Der ,

Purchrm
ert = Purchrm
ertd ∀e ∈ DRe , r, t. (1.48)
d∈Der
Constraint (1.46) forces each order to fall into a single interval, i.e., only one
of the variables Fertd (say, for d = d ∗ ) takes a value of 1, with all others being
zero. Constraint (1.47) allocates each of the orders to its corresponding interval
using the defined binary variable Fertd . Such equation forces the auxiliary contin-
uous variable Purchrm ∗
ertd to equal 0 for all d = d , while also bounding Purchertd ∗
rm rm
in the range [Purchertd ∗ −1 , Purchertd ∗ ]. Finally, constraint (1.49) expresses the con-
dition for which the summation of the auxiliary variable Purchrm ertd over d must
equal the variable Purchert . This implies that Purchrm rm
ert = Purchertd ∗ and, there-
rm rm rm
fore, Purchert ∈ [Purchertd ∗ −1 , Purchertd ∗ ], as desired. Let us note that similar con-
straints could be easily derive to account for other types of quantity discounts, i.e.,
in utilities, transportation services, and so forth.
Taking into account the discounts offered by the suppliers e for raw material r
for a set of discount intervals d, which are denoted by ψerd , the total amount of
money to be invested in raw materials is the following:

EPurchrm et = r
rm
d∈Der Purchertd · πert · Fertd · (1 − ψerd ) (1.49)
∀e ∈ DRe , t.
On the other hand, the external purchases of transport services and production
tr1 and ρ tr2 de-
utilities are determined through Eqs. (1.50) and (1.51). Here, ρeiws eiwm
note the unitary transport cost associated with sending products from plants to
warehouses and from warehouses to markets, respectively. Furthermore, τijut1 se rep-
ut2 τ ut3 τ ut4
resents the unitary production cost associated to the plants, whereas τrse ise iwe
represent the inventory costs.

Purchtret = tr1
Qiwst · ρeiws + tr2
Salesiwmt · ρeiwm
i j s i w m (1.50)
∀e, t,
prod

Purchet = Pij st · τijut1
se
i j ij s (1.51)
ut2 ut3 ut4
+ SIrst · τrse + SOist · τise + SWiwt · τiwe ∀e, t.
r s i s i w
The total fixed cost of operating a given SC structure in every time period t can
be computed by means of Eq. (1.52) as a sum of the fixed costs associated with the
current plant equipment j already installed in every site s (FCFSj st · FSj st−1 ) plus
the sum of the fixed costs of each distribution center w (F CF Wwt · FWwt−1 ).

FCostt = FCFSj st · FSj st−1 + F CF Wwt · FWwt−1 ∀t. (1.52)
j s w
Finally, the total investment in capital or fixed assets is computed through Eq.
(1.53). This term includes the investment made to expand the capacity of equip-
ment j in manufacturing site s in period t (PriceFS j st · FSEj st ), plus the investment
required to open a manufacturing plant, in case it is opened at period t (IstS · SBst ),
plus the investment required to support distribution center w capacity increase
wt · FWEwt ), plus the investment required to set a distribution center if it is
(PriceFW
opened at period t (IstS · SBst ).
 

FAssett =  PriceFS S
j st · FSEj st + Ist · SBst

s (1.53)

j
FW W
+ Pricewt · FWEwt + Iwt · SWwt ∀t.
w
The overall problem can be therefore mathematically posed as follows:
maximise CV
subject to
Eqs. (1.1)–(1.53).
1.3.4
Case Study
The capabilities of the proposed approach are illustrated by solving a retrofitting

problem of a SC comprising several manufacturing sites, distribution centers, and
Manufacturing
sites
Distribution
centres
Markets
W3
W2
S2
S3
W1
S1
W4
Fig. 1.2 Supply chain structure of the case study.
markets located in different European countries. A set of potential technologies are

assumed to be available in the manufacturing sites. Furthermore, several potential
locations for the manufacturing sites and the distribution centers, from which the
products should be transported to the final markets, are also considered, as is de-
picted in Fig. 1.2. The potential locations for the plants embedded in the SC (S1,
S2, and S3) are Barcelona (B), Budapest (Bu), and Milan (Mi). These plants can
manufacture three different products (P 1, P 2, and P 3) with four different tech-
nologies (TA to TD). These final products must be transported to the distribution
centers prior to being sent to the final markets (M1 to M5), where they become
available to customers. We assume an existing installed capacity of TA in S1 and
S3 of 10,000 cu and 50,000 cu, respectively. The investment cost associated with
the establishment of a manufacturing site is equal to 8,800,000 mu. The potential
locations for the distribution centers (W 1 to W 4) are B, Bu, Manchester (M), and
Lisbon (Li), whereas for the final markets are B, Bu, London (L), Li, and Toulouse
(T). We assume that at time zero W 1 has an installed capacity of 4,000 m3 . The
investment needed to open a distribution center is equal to 2,500,000 mu.
The specific volumes of the products and raw materials are shown in Table 1.2.
The amount of each type of raw material required to manufacture each product de-
pends on the specific technology being applied in each case (see Table 1.3). More-
over, the capacity coefficients for each technology and product are shown in Table
1.4. Table 1.6 gives the costs of the raw materials and their maximum availability at
Table 1.2 Specific volume of products i (υi (10−5 m3 /kg)).
Product υi
P1 4.3
P2 8.0
P3 5.5
Table 1.3 Mass fractions for consumption of raw materials by plant equipment j (πrij (adim.)).
Product Raw material

R1 R2 R3 R4 R5 R6
j = TA
P1 0.59 0.35 0.18
P2 0.47 0.24 0.30
j = TB
P2 0.47 0.24 0.30
P3 0.24 0.71 0.30
j = TC
P1 0.59 0.30 0.18
P2 0.53 0.24 0.30
P3 0.24 0.77 0.30
j = TD
P1 0.59 0.35 0.12
P3 0.24 0.71 0.24
Table 1.4 Capacity consumption of plant equipment j by each product i (νij (c.u./kg)).
Product Equipment
TA TB TC TD
P1 8.00 7.20 8.40 7.60

P2 7.00 8.40 7.35 8.82
P3 7.00 8.40 7.35 7.70
each period, which is assumed to remain constant within the whole planning hori-
zon. The demand data are listed in Table 1.5. The initial inventories are supposed
to be equal to zero for all products and raw materials. The upper bound imposed
to the increment in the capacity of the technologies at each manufacturing site is
equal to 500,000 cu and the lower bound is 50,000 cu. Table 1.7 shows the fixed and
investment costs associated with each technology. The upper and lower bounds im-
posed to the increment in capacity of the distribution centers are 2,000 and 30,000
m3 . The rest of the data associated with the DCs can be found in Table 1.8. The
capacities of the facilities can only be increased every two years. The availability of
utilities is assumed to be unlimited.
With respect to the financial matters, it is assumed that the firm has at the begin-
ning of the planning horizon an initial portfolio of marketable securities. Specifi-
cally, the firm owns 15,000 mu in marketable securities maturing in period 2 and
18,000 mu maturing at period 3. The initial cash is assumed to be equal to the
minimum cash allowed, which is 125,000 mu. Under an agreement with a bank,
the firm has an open line of short term credit at a 15% annual interest with a maxi-
mum debt allowed of 4,000,000 mu. The initial debt is assumed to be equal to zero
and the prices of the materials kept as inventories at the end of the time horizon are
Table 1.5 Demand of product i at market m at each period t (Demimt (103 kg)).
Product i t
1–7 8–20 21–23 24–32 33–47 48–56 57–60
m = M1
P1 105.0 105.0 105.0 150.0 105.0 150.0 105.0
P2 50.8 50.8 50.8 72.5 50.8 72.5 50.8
P3 43.8 43.8 43.8 62.5 43.8 62.5 43.8
m = M2
P1 105.0 600.0 105.0 150.0 105.0 150.0 105.0
P2 119.0 680.0 119.0 170.0 119.0 170.0 119.0
P3 133.0 760.0 133.0 190.0 133.0 190.0 133.0
m = M3
P1 35.0 35.0 35.0 50.0 35.0 50.0 35.0
P2 220.5 220.5 220.5 315.0 220.5 315.0 220.5
P3 150.5 150.5 150.5 215.0 150.5 215.0 150.5
m = M4
P1 70.0 400.0 70.0 100.0 70.0 100.0 70.0
P2 294.0 1680.0 294.0 420.0 294.0 420.0 294.0
P3 147.0 840.0 147.0 210.0 147.0 210.0 147.0
m = M5
P1 140.0 140.0 140.0 200.0 140.0 200.0 140.0
P2 171.5 171.5 171.5 245.0 171.5 245.0 171.5
P3 49.0 49.0 49.0 70.0 49.0 70.0 49.0
Table 1.6 Cost (m.u./kg) and maximum availability

(Aert (103 kg)) of raw material r at each period of time t.
Raw material
R1 R2 R3 R4 R5 R6
Availability 150.0 125.0 335.0 60.0 60.0 40.0

Cost 10.10 6.60 6.10 8.10 12.1 10.1
Table 1.7 Equipment j fixed cost (FCFSj st (m.u./Mg)) and investment (PriceFS
j st (m.u./kg)).
Equipment j
TA TB TC TD
Fixed cost 25.0 37.0 50.0 45.0

Investment 3.26 3.91 2.93 4.04
assumed to be a 85% of their market prices for final products and a 100% for raw
materials. The market prices of the final products are assumed to remain constant
during the whole planning horizon and are provided in Table 1.9.
Table 1.8 Distribution centers fixed cost (FCFWwt (m.u./m3 ))

and investment (PriceFS 3
wt (m.u./m )).
Distribution center w
W1 W2 W3 W4
Fixed cost 17.5 15.0 15.0 13.75

Investment 100.0 80.0 75.0 70.0
Table 1.9 Price of product i at each market m and period t(Priceimt (m.u./kg)).
Product
P1 P2 P3
m = M1, M3, M4, M5

16.15 17.51 18.39
m = M2
16.43 17.81 18.71
Table 1.10 Transportation cost of product i from site s to

tr1 (m.u./m3 )).
distribution center w (ρeiws
Distribution center w Site s

S1 S2 S3
i = P1
W1 0.02 0.48 0.27
W2 0.48 0.02 0.18
W3 0.19 0.36 0.24
W4 0.21 0.60 0.24
i = P2
W1 0.04 0.90 0.49
W2 0.90 0.04 0.34
W3 0.36 0.67 0.45
W4 0.38 1.12 0.45
i = P3
W1 0.03 0.62 0.34
W2 0.62 0.03 0.23
W3 0.25 0.46 0.31
W4 0.26 0.77 0.31
The SC under study has three external suppliers, the first providing raw mate-
rials, the second transportation services, and the third labor utilities. Liabilities in-
curred with the raw materials supplier must be repaid within one month according
to the terms of the credit (2 percent-same period, net-28 days for the raw materials
supplier). The supplier of raw materials offers discounts for large orders. Thus, a
3% discount is applied for orders of raw material R1 higher than 45,000 kg and a
Table 1.11 Transportation cost of product i from distribution
tr2 m.u./m3 )).
center w to market m (ρeiwm
Distribution center w Market m

M1 M2 M3 M4 M5
i = P1
W1 0.02 0.48 0.21 0.15 0.11
W2 0.48 0.02 0.39 0.60 0.22
W3 0.21 0.39 0.04 0.18 0.22
W4 0.21 0.60 0.17 0.02 0.22
i = P2
W1 0.04 0.90 0.38 0.27 0.20
W2 0.90 0.04 0.72 1.12 0.40
W3 0.38 0.72 0.07 0.34 0.40
W4 0.38 1.12 0.32 0.04 0.40
i = P3
W1 0.03 0.62 0.26 0.19 0.14
W2 0.62 0.03 0.49 0.77 0.28
W3 0.26 0.49 0.05 0.23 0.28
W4 0.26 0.77 0.22 0.03 0.28
5% for orders above 80,000 kg. The payments associated with the transport services
and labor tasks cannot be stretched and must be fulfilled within the same time pe-
riod in which the purchase incidence takes place. Transportation costs are given
in Tables 1.10 and 1.11, whereas production costs are shown in Table 1.12. The
technical coefficients associated with the set of marketable securities that the firm
has agreed to purchase and sale have been computed by considering a 2.8% annual
interest for purchases and a 3.5% for sales. We consider outflows of cash equal to
5.0, 7.5, and 10.0 millions of m.u. in periods 13, 15, and 17 due to wages, rents, and
dividends. We also assume that the ratio between the long term debt and the equity
must always be kept equal to 0.41. With respect to the long term debt, let us note
that the firm can access a long term credit at a 10% annual interest. Shareholders
expect an annual ROE equal to 30%. The taxes rate is equal to 30%. Depreciation
is calculated by means of the straight line method applied over a time horizon of
ten years. Finally, we assume that receivables on sales in any period are paid with
a delay according to the proportions given in Table 1.13 and may be pledged at a
80% of their face value regardless their maturing period.
Sixty one planning periods with a length of one month each are considered.
The implementation in GAMS [50] of the integrated formulation leads to an MILP
model with 40,306 equations, 46,916 continuous variables, and 252 discrete vari-
ables. It takes 185 CPU seconds to reach a solution with a 0% integrality gap on a
AMD Athlon 3000 computer using the MIP solver of CPLEX (10.0).
The integrated model is first solved by maximising the corporate value of the
firm. To explicitly show the trade-off between the corporate value and the NPV and
profit we next add a new constraint to force the model to seek solutions with higher
Table 1.12 Production cost of product i manufactured in plant equipment j at site s (m.u./kg).
Plant equipment j Site s

S1 S2 S3
i = P1
TA 0.63 0.59 0.71
TB 0.51 0.48 0.58
TC 0.72 0.69 0.82
TD 0.77 0.73 0.88
i = P2
TA 0.55 0.52 0.62
TB 0.59 0.56 0.67
TC 0.63 0.60 0.72
TD 0.90 0.85 1.02
i = P3
TA 0.55 0.52 0.62
TB 0.59 0.56 0.67
TC 0.63 0.60 0.72
TD 0.78 0.74 0.89
Table 1.13 Fraction of sales that are receivable at n time periods after sales (δmtt ).
Market m Time periods between execution and maturing period of sales (n = t − t)

0 1 2 3 4 5 6
M1, M3, M5 0.05 0.10 0.25 0.60 0.00 0.00 0.00

M2 0.00 0.00 0.00 0.00 0.05 0.20 0.75
M4 0.00 0.05 0.05 0.15 0.75 0.00 0.05
NPVs and profits, respectively. This is indeed equivalent to apply the -constraint
method proposed by Haimes et al. [51] to our problem, taking into account in each
case the above-commented objectives, CV and NPV and CV and profit, respectively,
at the same time.
The SC network configurations obtained by following this procedure are sum-
marized in Tables 1.14–1.16. The Pareto curves obtained by applying the afore-
mentioned procedure are shown in Figs. 1.3 and 1.4. Numerical results show that
the solutions computed by maximising profit or NPV as single objectives are far
away from the optimal one (i.e., the best solution in terms of CV). Certainly, the
maximum corporate value solution is almost 300% higher than the one computed
by maximising profit and 267% higher than the one accomplished when maximis-
ing NPV. On the other hand, the maximum profit and NPV solutions are quite
similar. Moreover, from these results, it is clear that in both the cases a conflict
exists between different objectives (i.e., maximum corporate value and maximum
profit or NPV). Thus, numerical results show that an improvement in the profit or
NPV is only possible if the decision maker is willing to compromise the corporate
Table 1.14 Maximum profit network design.
Total profit 121,653,714.87 m.u.

Net present value (NPV) 47,476,865.35 m.u.
Corporate value (CV) 36,392,463.16 m.u.
Time period (t)

0 1 24 48
Manufacturing sites
s j Capacity increment (c.u.)
S1 TA 10,000.0 0.0 0.0 0.0
TB 0.0 0.0 0.0 0.0
TC 0.0 0.0 0.0 0.0
TD 0.0 0.0 0.0 0.0
S2 TA 0.0 500,000.0 0.0 0.0
TB 0.0 407,796.6 0.0 0.0
TC 0.0 500,000.0 0.0 0.0
TD 0.0 500,000.0 0.0 0.0
S3 TA 50,000.0 0.0 0.0 0.0
TB 0.0 0.0 0.0 0.0
TC 0.0 496,610.2 0.0 0.0
TD 0.0 478,813.6 0.0 0.0
Distribution centers
w Capacity increment (m3 )
W1 4,000.0 0.0 0.0 0.0
W2 0.0 2,000.0 0.0 0.0
W3 0.0 0.0 0.0 0.0
W4 0.0 0.0 0.0 0.0
7
x 10
16
14
12
Corporate value (m.u.)
10
2
1.06 1.08 1.1 1.12 1.14 1.16 1.18 1.2 1.22
Profit (m.u.) x 10
8
Fig. 1.3 Corporate value – profit Pareto curve.
value of the firm. Certainly, SC configurations with better profits or NPVs can only
be achieved at the expense of a reduction in the corporate value of the firm.
Table 1.15 Maximum net present value network design.
Total profit 121,464,749.10 m.u.

Time period (t)

0 1 24 48
Manufacturing sites
S1 TA 10,000.0 0.0 0.0 0.0
TB 0.0 0.0 0.0 0.0
TC 0.0 0.0 0.0 0.0
TD 0.0 0.0 0.0 0.0
S2 TA 0.0 500,000.0 0.0 0.0
TB 0.0 0.0 0.0 0.0
TC 0.0 500,000.0 0.0 0.0
TD 0.0 500,000.0 0.0 0.0
S3 TA 50,000.0 329,830.5 0.0 0.0
TB 0.0 0.0 0.0 0.0
TC 0.0 496,610.2 0.0 0.0
TD 0.0 478,813.6 0.0 0.0
W1 4,000.0 0.0 0.0 0.0
W2 0.0 2000.0 0.0 0.0
W3 0.0 0.0 0.0 0.0
W4 0.0 0.0 0.0 0.0
7
x 10
16
14
12
Corporate value (m.u.)
10
2
4.2 4.3 4.4 4.5 4.6 4.7 4.8
Net present value (m.u.) x 10
7
Fig. 1.4 Corporate value – net present value Pareto curve.
It is worth to mention that this case study represents a very specific situation
where there is one market (M2) in which the product prices are higher in com-
parison with the others (see Table 1.9). At such market, accounts receivable are due
within a large time period. Under this assumption, the design-planning model that
Table 1.16 Maximum corporate value network design.
Total profit 105,955,397.81 m.u.

Time period (t)

0 1 24 48
Manufacturing sites
S1 TA 10,000.0 500,000.0 0.0 0.0
TB 0.0 395,796.6 0.0 0.0
TC 0.0 500,000.0 0.0 0.0
TD 0.0 500,000.0 0.0 0.0
S2 TA 0.0 0.0 0.0 0.0
TB 0.0 0.0 0.0 0.0
TC 0.0 0.0 0.0 0.0
TD 0.0 0.0 0.0 0.0
S3 TA 50,000.0 0.0 0.0 0.0
TB 0.0 0.0 0.0 0.0
TC 0.0 496,610.2 0.0 0.0
TD 0.0 478,813.6 0.0 0.0
W1 4,000.0 2,000.0 0.0 0.0
W2 0.0 0.0 0.0 0.0
W3 0.0 0.0 0.0 0.0
W4 0.0 0.0 0.0 0.0
Table 1.17 KPI values of optimal configurations for second case study.
Network configuration KPI

Total profit NPV CV
Optimal total profit 116,597,744.59 45,374,556.23 89,386,052.53

Optimal NPV 116,377,423.52 45,516,838.30 110,018,635.40
Optimal CV 105,955,397.81 42,068,183.09 145,023,155.58
accounts for the maximization of a myopic KPI (either profit or NPV) that neglects
the financial side of the problem decides to configure a supply chain network ca-
pable of easily fulfilling the demand of market M2 as much as possible (see Fig.
1.5).The profit and the NPV are indeed blind KPIs in the sense that they are not
capable of properly assessing the financial costs associated with the shortages of
cash. As a result, accounts receivables increase in certain periods of time in which
the firm has to face important payments. The budgeting model is then forced to
pledge them, mainly during month 18 to 36 (quarters 6 to 12). Pledging is indeed a
8
x 10
4
Maximum profit
Maximum net present value
3.5
Maximum corporate value
2.5
Sales (m.u.)
1.5
0.5
0
M1 M2 M3 M4 M5 Total
Markets
Fig. 1.5 Sales carried out in each market for each optimal SC network configuration.
7
x 10
3
0
m.u.
−1
−2
−3 Profit after taxes

Change in capital invested
Discounted free cash flows
−4
0 2 4 6 8 10 12 14 16 18 20
Quarters
Fig. 1.6 Value accumulation at each quarter for maximum profit SC network structure.
7
x 10
1
0.5
0
m.u.
−0.5
−1 ∆ accounts receivables
∆ accounts payables
Net investment
∆ other expenses
−1.5
2 4 6 8 10 12 14 16 18 20
Quarters
Fig. 1.7 Capital investment needed composition at each

quarter for maximum profit SC network structure.
very expensive way of getting cash and because of that the firm reduces its capacity
of creating value when applying it.
On the other hand, the integrated approach that accounts for the maximization
of a suitable financial objective function (i.e., corporate value) is able to properly as-
7
x 10
3
0
m.u.
−1
−2

−4
0 2 4 6 8 10 12 14 16 18 20
Quarters
Fig. 1.8 Value accumulation at each quarter for maximum net

present value SC network structure.
7
x 10
1
0.5
0
m.u.
−0.5
−1
∆ accounts receivables
Net investment
∆ other expenses
−1.5
2 4 6 8 10 12 14 16 18 20
Quarters

quarter for maximum net present value SC network structure.
7
x 10
3
0
m.u.
−1
−2

−4
0 2 4 6 8 10 12 14 16 18 20
Quarters
Fig. 1.10 Value accumulation at each quarter for maximum

corporate value SC network structure.
7
x 10
1.5
0.5
m.u.
−0.5
∆ accounts receivables
−1
Net investment
∆ other expenses
−1.5
2 4 6 8 10 12 14 16 18 20
Quarters

quarter for maximum corporate value SC network structure.
Fig. 1.12 Optimal profit financial Gantt chart.
Fig. 1.13 Optimal NPV financial Gantt chart.
sess the trade-off between the increment in profit that can be achieved by fulfilling
demand at market M2 and the increment in net working capital that is required
to carry out this decision. Hence, the integrated approach computes a SC configu-
ration that does not fulfill the demand in M2 due to the poor payment conditions
associated with its customers. Consequently, the net working capital needed is re-
duced significantly thus increasing the value accumulated during the whole plan-
ning horizon. This can be observed in Figs. 1.6 to 1.11, which depict the behavior
of the corporate value and the structure of the consumed value during the planning
Fig. 1.14 Optimal CV financial Gantt chart.
horizon for each optimum SC network configuration. For the purpose of facilitat-
ing their interpretation, the planning periods (months) have been aggregated into
quarters in these figures. In Figs. 1.12 to 1.14, financial Gantt charts for each op-
timal SC configuration are shown. They describe how the cash is composed and
utilized during the first 25 planning periods.
To demonstrate advantages of using CV and its robustness to assess strategic de-
cisions under different scenarios, the previous case study has been modified and
solved. Two modifications have been done: (1) prices payed by market M2 are equal
to those payed by rest of markets and (2) due time of account receivables in mar-
ket M2 has the same conditions offered by market M4. The optimal SC network
configurations obtained are quite similar to those resulting in the previous case
study. In Table 1.17, the values of each KPI are shown for each optimal configura-
tion. Even for this second case, it can be observed that the maximum CV solution
is 62% better than the one computed by maximising profit and 32% superior than
the one accomplished when maximising NPV. It is clear from these two examples
that the integrated model introduced in this work has a great potential when ad-
dressing the challenge of designing a SC capable to preserve and improve the value
of the corporation.
1.4
SCM Under Uncertainty
The complexity of the SCM problem is further increased by a high degree of un-
certainty brought about by external factors, such as continuously changing market
conditions and customer expectations, and internal parameters, such as product
yields, qualities, and processing times. Although it has been widely recognized the
importance of incorporating uncertainties in SCM, most of the models developed
so far in the literature are deterministic [41, 52–54], i.e., they assume that all the
problem data is known in advance. Thus, the accuracy of the solutions generated
by using deterministic models may depend on the degree of uncertainty.
Specifically, a literature survey reveals that the most important and extensively
studied source of uncertainty has been demand [55, 56]. The emphasis on incor-
porating demand uncertainty into the planning decisions is appropriate given the
fact that meeting effectively customer demand is what mainly drives most planning
initiatives.
The SCM problem under demand uncertainty has emerged only recently as an
area of research. The prevalent approach in the area of optimization under uncer-
tainty prior to the realization of uncertainty has been the multistage stochastic pro-
gramming, which deals with problems involving a sequence of decisions that react
to outcomes that evolve over time. In these kinds of approaches, at each stage, one
makes decisions based on currently available information, i.e., past observations
and decisions, prior to the realizations of future events. Stochastic approaches dif-
fer primarily in the selection of the decision variables and the way in which the
expected value term, which involves a multidimensional integral accounting for
the probability distribution of the uncertain parameters, is computed. Specifically,
two distinct methodologies for representing uncertainty can be identified within
probabilistic methods. These are the scenario-based approach and the distribution-
based approach.
In the first approach [33, 39–41, 57, 58], the uncertainty is described by a set of
discrete scenarios capturing how the uncertainty might play out in the future. Each
scenario is associated with a probability level representing the decision maker’s ex-
pectation of the occurrence of a particular scenario. The scenario-based approach
avoids the problem of multivariate integration when the random variables follow
multidimensional continuous distributions. This is achieved by generating a finite
set of scenarios, from sampling or a discrete approximation of the given distribu-
tions, to represent the probability space. With the scenarios or scenario tree speci-
fied, the stochastic program becomes a deterministic equivalent program.
In cases where a natural set of discrete scenarios cannot be identified and
only a continuous range of potential futures can be predicted, the distribution-
based approach is used. By assigning a probability distribution to the continuous
range of potential outcomes, the need to forecast exact scenarios is obviated. The
distribution-based approach is also adopted in several works [59, 60], in which de-
mand is modeled as normally distributed with a specified mean and standard devi-
ation.
1.4.1
Mathematical Formulation
This section illustrates the way in which a stochastic programming model address-
ing the design of chemical SCs under demand uncertainty can be derived. Thus,
a recourse model with two stages is constructed to incorporate the uncertainty as-
sociated within the demand. In a two-stage stochastic optimization approach the
uncertain model parameters are considered random variables with an associated
probability distribution and the decision variables are classified into two stages.
The first-stage variables correspond to those decisions that need to be made here-
and-now, prior to the realization of the uncertainty. The second stage or recourse
variables correspond to those decisions made after the uncertainty is unveiled and
are usually referred to as wait-and-see decisions. After the first-stage decisions are
taken and the random events realized, the second-stage decisions are made subject
to the restrictions imposed by the second-stage problem. Due to the stochastic na-
ture of the performance associated with the second-stage decisions, the objective
function consists of the sum of the first-stage performance measure and the ex-
pected second-stage performance (for an overview of stochastic techniques refer to
[61]).
In our problem the uncertainty associated with the demand is represented by
a set of scenarios with given probability of occurrence. Thus, in the deterministic
equivalent of the original stochastic formulation, these scenarios along with their
associated probabilities must be provided as input data into the model. In cases
where the demand follows a specific type of probability distribution, this can be
discretized by applying Monte Carlo sampling, thus generating a set of explicit
scenarios.
Specifically, the decision variables representing the network configuration are
considered as first-stage variables as it is assumed that they have to be taken at the
design stage before the demand uncertainty is unveiled. On the other hand, the
amount of products to be produced and stored in the nodes of the SC, the flows of
materials transported among the entities of the network, and the product sales are
considered as second-stage variables.
At the end of the design horizon, a different value of the objective function is ob-
tained for each particular realization of demand uncertainty. The proposed model
accounts for the maximization of the expected value of the objective function dis-
tribution.
We next present a simplified version of the mathematical model described in
the previous section and illustrate the way to derive the stochastic version of such
model. Thus, for the sake of simplicity, in the new version of the model the financial
variables and constraints of the problem have been removed. Furthermore, the
model assumes that the structure of the network must be fixed at the beginning
of the time horizon and cannot be further modified in subsequent periods of time
(i.e., capacity expansions are not allowed).
Thus, to construct the deterministic multiscenario equivalent of the stochastic
model, we need to introduce a superscript n in all the second stage variables of
the problem. In our case, the second-stage variables are the amount of materials
produced and stored in different sites, the flows of materials between nodes, and
sales. By doing so, we make sure that they can take different values depending
on the specific realization of the uncertain events. Furthermore, the mass balance
and capacity constraints must be defined over all the scenarios. Then, the new
constraints have the following form:

Purchrm
erstn + SIrst−1n = SIrstn + αrij · Pij stn
e∈Er j i∈(Ij ∩Ir )
∀r, s, t, n, (1.54)

Pij stn + SOist−1n = SOistn + Qiwstn ∀i, s, t, n, (1.55)
j ∈Ji w

Qiwstn + SWiwt−1n = SWiwtn + Salesiwmtn ∀i, w, t, n, (1.56)
s m

Salesiwmtn Demimtn ∀i, m, t, n. (1.57)
w
1.4.1.2 Capacity and Facilities Location Constraints

plant
ϕij · Pij stn FSj st−1 ∀j, s, t, n, (1.58)
i∈Ij

ware
ϕiw · SWiwtn FWwt−1 ∀w, t, n. (1.59)
i
1.4.1.3 Objective Function Net Present Value (NPV)

The new objective function is the maximization of the expected NPV, which is
computed by making the average of the NPV over all the scenarios.

E[NPV] = probn · NPVn . (1.60)
n
The NPV in each scenario is computed through the following equations:

T
Profittn − FAssett
NPVn = ∀n, (1.61)
(1 + RR)t
t=0

Profittn = Salesiwmtn · Priceimt − Purchrm erstn −
i w m r s
− Qiwstn · λtr1
eiws − Salesiwmtn · λtr2
eiwm −
s

i
j iw m
− Pij stn · τijut1
se −
ut2
SIrstn · τrse − (1.62)
s r s

i
j
ut3 ut4
− SOistn · τise − SWiwtn · τiwe −
i s i w
−FCostt − Invtn .
1.4.2
Risk Management in SCM
Although stochastic models optimize the total expected performance measure, they
usually do not provide any control on its variability over different scenarios, i.e.,
they assume that the decision maker is risk neutral. Although the SC decisions
obtained could be considered more robust than the deterministic ones based on
nominal parameter values, by taking a purely expected profit maximization per-
spective, the model assumes that the decision maker is risk neutral or indifferent
to profit variability. However, different attitudes toward risk may be encountered.
In general, most decision makers are risk averse implying a major preference for
lower variability for a given level of return [62]. Therefore, there is no guarantee
that the process will perform at a certain level over all of the uncertain parameters
space. The only guarantee is that the average is optimized.
The extension of the traditional objective based on simple expectations was pre-
sented by Ahmed and Sahinidis [56], who argued that robustness should also be
sought. They penalized downside risk, defined here as costs above the expected
cost. [7] also recognized that simply optimizing expected returns can lead to higher
risk solutions. [63] presented some techniques to manage financial risk in schedul-
ing problems similarly to the way it was done by Barbaro and Bagajewicz [62] for
planning problems. Some of these techniques were also used by [38, 39] for ma-
nipulating the financial risk in the area of design and operation of supply chains
under uncertainty.
The idea underlying risk management is incorporation of the trade-off between
risk and profit within the decision-making process, thus leading to a multiobjective
optimization problem in which the expected performance and the risk measure are
the objectives. The next section describes different risk measures proposed so far
in the literature to control the variability of the objective function over the entire
range of scenarios.
1.4.2.1 Financial Risk

Barbaro et al. [62] defined the financial risk associated with a plan under uncer-
tainty as the probability of not meeting a certain target profit (maximization) or
cost (minimization) level referred to as [62]. For the two-stage stochastic prob-
lem, the financial risk associated with a design x and target profit is therefore
expressed by the following probability:
FR = P (Profit < ), (1.63)
where Profit is the profit after the uncertainty has been unveiled and a scenario
realized. The definition of FR can be rewritten with the help of binary variables
as follows:

FR = probs · Zs , (1.64)
s
where Zs is a new binary variable which equals 1 in the case Profits < and 0
otherwise.
In a discrete scenario case, financial risk is given by the cumulative frequency
obtained from the Profit histogram. A more straightforward way of assessing and
understanding the trade-offs between risk and profit is to use the cumulative risk
curve as depicted in Fig. 1.15.
Barbaro et al. [62] proved that minimization of risk at some profit levels renders
a trade-off with expected profit. A risk-averse decision maker will feel comfortable
with low risk at low values of , while a risk taker will prefer to lower the risk at
high values of . The trade-off lies in the fact that minimizing risk at low values of
(e.g., a loss) is in conflict with the minimization of risk at high values of (e.g.,
large profits) and vice versa. From a mathematical programming point of view,
minimizing FR for a continuous range of profit targets results in an infinite

multiobjective optimization problem (see Fig. 1.16). Even though this model would
be able to reflect the decision-maker’s intention, having an infinite optimization
problem is computationally prohibitive.
The authors, however, suggested to approximate the ideal infinite optimization
approach by a finite multiobjective problem that only minimizes risk at some finite
number of profit targets and maximizes the expected profit. This approach gives
rise to the following finite multiobjective formulation:

maximise E[Profit] = probs · Profits ,
s
MinFR1 = probs · Zs1 ,
s
MinFRi = probs · Zsi ,
s
...
0.9
0.8
0.7
0.6
Risk
0.5
0.4
0.3
0.2 FR Ω
1
0.1
0
0 20 Ω1 40 60 80 100
DR Profit (m.u.)
Ω
Fig. 1.15 Cumulative risk curve.
1
Min FR
0.9 Ω
4
0.8 Min FR Ω
3
0.7
0.6
Risk
0.5
0.4
Min FR Ω
2
0.3
Max E[profit]
0.2 Min FR
Ω
1
0.1
Ω Ω Ω3 Ω
1 2 4
0
0 20 40 60 80 100
Profit (m.u.)
Fig. 1.16 Multiobjective problem.

subject to
Profits i − Us · Zsi ∀i, s, (1.65)
Profits i + Us · (1 − Zsi ) ∀i, s, (1.66)
Zsi ∈ {0, 1} ∀i, s, (1.67)
where ys represents the optimal second-stage solution associated with the schedule
x that corresponds to scenario s. In the above formulation, constraints (1.65) and
(1.66) force the new integer variable zsi to take a value of zero if the profit for sce-
nario s is greater or equal than the target level (i ) and a value of one otherwise. To
do this, an upper bound of the profit of each scenario (Us ) is used. The value of the
binary variables is then used to compute and penalize financial risk in the objective
function. Such procedure generates a set of Pareto optimal solutions behaving in
dissimilar ways under the uncertain environment from which the decision maker
should choose the best one according to his/her preferences. Nevertheless, the in-
clusion of new integer variables represents a major computational limitation of
the resulting formulation. By using only one target level (), one can reduce the
number of integer variables required, which leads to the following multiobjective
problem:
maximise {E[Profit], −FR }.
However, in those cases where a high number of scenarios and thus binary vari-
ables are required, the resulting formulations are still in general computational
prohibitive. In view of these limitations, two alternative approaches for integrating
risk management have been considered in this work and appended as a second cri-
terion to the objective function: the downside risk definition proposed by [64] and
the worst-case performance.
1.4.2.2 Downside Risk

A possible way of avoiding the use of binary variables to determine the risk con-
sists of reformulating the problem without explicitly using this definition. For this
purpose, the use of the concept of downside risk, in the way introduced by [64], is
applied. The DR is calculated with the help of the following constraints:

DR = probs · δs , (1.68)
s
δs − Profits ∀s,
δs 0 ∀s,
where δs is a continuous variable. Therefore, downside risk is defined as the area

under the risk curve from profit ∞ to , where is the profit target (see Fig.
1.15). The DR can be utilized to control financial risk at different Profit targets by
varying from small values up to higher values and obtaining a full spectrum of
solutions to be used by the decision maker as a support-decision tool.
It is worthwhile to note that, whereas financial risk is defined as a probability,

downside risk is an expected value. The incorporation of this metric within the
modeling framework gives rise to a multiobjective optimization problem:
maximise {E[Profit], −DR }
1.4.2.3 Worst Case

The worst-case profit is also adopted as an alternative metric to control or reduce
the probability of meeting unfavorable scenarios. A major difference with respect
to the other approaches is that the probability information of the problem cannot
be used. Moreover, both the expected profit and the profit in the worst-case scenario
are to be maximized.
WC Profits . (1.69)
The worst case is computed through Eq. (1.69), and the inclusion of this term
as an alternative objective to be maximized leads to the following multiobjective
formulation:
maximise {E[Profit], W C}.
1.4.3
Multiobjective Problem
The incorporation of any of the risk measures mentioned above into the objective
function would give rise to the following multiobjective mixed integer linear prob-
lem (moMILP):
maximise {E[NPV], W C}
subject to
Eqs. (1.36)–(1.39), (1.54)–(1.62),
where RiskMeasure denotes either the risk, the downside risk, or the worst case.
Specifically, the worst case metric has the advantage, compared with the others, of
having an easy implementation and also a good numerical performance in two-
stage stochastic models [63]. For this reason, it has been chosen as the risk metric
to be incorporated into our formulation.
The solution of this multiobjective formulation should provide a set of Pareto
efficient solutions of the problem, that is to say, a set of solutions that properly
trade-off one of the objectives against the other. Such efficient solutions have the
property of not being dominated by any of the others (i.e., an improvement in one
of the objectives can only be achieved at the expense of a decrease in the other
one). Diverse standard multiobjective optimization methods developed so far in
the literature can be used to solve the afore-mentioned problem. Specifically, we
apply in this case the -constraint method which consists of maximizing one of the
objectives and constraint the other one to be lower than a minimum desired level.
7
x 10
2.46
2.44
2.42
2.4
E[NPV] (m.u.)
2.38
2.36
2.34
2.32
1.925 1.93 1.935 1.94 1.945 1.95 1.955 1.96 1.965
7
Worst case (m.u.) x 10
Fig. 1.17 Optimal NPV financial Gantt chart.
Thus, the Pareto solutions of the problem are computed by solving a set of single-
objective problems that account for the maximization of the expected NPV and
must satisfy a constraint regarding the NPV achieved in the worst case scenario.
maximise E[NPV]
subject to
Eqs. (1.36)–(1.39), (1.54)–(1.62)
W C .
Therefore, by changing the values of , the set of Pareto solutions representing
the compromise between E[NPV] and WCase can be obtained. Each of these solu-
tions implies a SC configuration and a set of planning decisions. Such solutions
should finally be used by the decision maker responsible of finding the one rep-
resenting the right compromise between E[NPV] and risk according to his/her
preferences.
1.4.4
Case Study
We next apply the formulation shown above to the same case study previously pre-
sented. We consider a time horizon of 6 years. The uncertainty associated with the
demand is represented by 50 scenarios which are generated by applying a Monte
Carlo sampling to the normal probability functions that are assumed to character-
ize it. The means of these probability functions are a quarter of the deterministic
demands, while the standard deviation is equal to 20% in all the cases.
The problem is first solved by maximizing the expected NPV and neglecting the
worst case. This provides the first Pareto solution of the curve. The worst case is
next gradually constrained to generate the entire Pareto frontier (see Fig. 1.17). As
can be observed, a conflict naturally exists between solutions that maximizes the
expected NPV and the worst case. Thus, an improvement in the worst case can
1
WCase = 19,262,416 m.u.
WCase = 19,362,416 m.u.
0.9 WCase = 19,462,416 m.u.
WCase = 19,562,416 m.u.
0.8 WCase = 19,622,416 m.u.
0.7
0.6
Risk
0.5
0.4
0.3
0.2
0.1
0
1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9
NPV (m.u.) x 10
7
Fig. 1.18 Optimal CV financial Gantt chart.
Table 1.18 Maximum profit network design.
Manufacturing sites Capacity increment (c.u.)

s j Maximum E[NPV] Maximum worst case
S1 TA 500,000 374,996
TB 0 50,000
TC 273,870 242,173
TD 351,454 312,483
S2 TA 0 0
TB 0 0
TC 0 0
TD 0 0
S3 TA 0 0
TB 0 0
TC 0 0
TD 0 0
Distribution centers Capacity increment (c.u.)

w Maximum E[NPV] Maximum worst case
W1 2,000 2,000
W2 0 0
W3 2,000 2,000
W4 0 0
only be achieved at the expense of a decrease in the expected NPV. Each Pareto
solution involves a supply chain structure. Table 1.18 shows the SC structures as-
sociated with the maximum expected NPV and the worst case solutions. As can
be observed, the model is forced to seek supply chain configurations with lower
capacities in order to reduce the probability of lower benefits. This can also be ob-
served in Fig. 1.18, which depicts the cumulative probability curves associated with
different solutions of the problem. As can be observed, the probability of having
low earnings gradually decreases when the worst case is forced to take low values.
For instance, the maximum expected NPV solution has a 15% probability of a NPV
lower than 21,000,000 m.u., whereas the maximum worst case solution has only
a 10%. On the other hand, the former solution leads to a 10% of a NPV above
21,000,000 m.u., whereas in the solution latter this probability is reduced to 0%.
1.5
Conclusions
This chapter has addressed the design and retrofit of chemical SCs taking into
account financial concerns. The proposed framework applies mixed integer mod-
eling techniques to develop holistic mathematical models for SCM that are able to
optimize the process operations decisions in conjunction with the finances.
The main advantages of our holistic approach have been highlighted through a
case study, in which the integrated model has been compared with the traditional
method. The former approach pursues the maximization of a suitable financial key
performance indicator that is able to properly assess the expenses associated with
the shortages of cash (i.e., the corporate value of the firm at the end of the time
horizon). On the other hand, the latter strategy accounts for the maximization of a
myopic key performance indicator that is unable to assess the costs of the financing
sources. Numerical results show that the integrated solution not only ensures the
feasibility of the strategic decisions from the financial viewpoint but also leads to a
superior economic performance given its higher capacity of creating value for the
firm.
Furthermore, we have also investigated the impact that different sources of un-
certainty affecting the SCM problem has in the performance of the network. For
this purpose, we have investigated the use of multiscenario two-stage stochastic
programming techniques accounting for the maximization of the expected NPV.
To control the variability of the objective function over the entire range of scenar-
ios, a risk measure has been appended to the objective function as an additional
objective to be optimized along with the expected NPV. The resulting multiobjec-
tive stochastic formulation has been applied to the same case study used before,
for which a set of Pareto optimal solutions has been obtained. Numerical results
have shown that a trade-off naturally exists between both objectives and that the
variability of the objective function can be controlled through structural changes
in the configuration of the SC. These solutions, which behave in dissimilar ways
under the uncertain environment, should be used by the decision maker to find
the best one according to his/her preferences.
Acknowledgments
The authors wish to acknowledge support of this research work from the European
Community (Contract Nos. GIRD-CT-2001-00466 & MRTN-CT-2004-512233), the
CICyT-MEC (project No DPI2003-0856), and the CIRIT-Generalitat de Catalunya
References 39
(project No I-898). Contribution from Mariana Badell and José Miguel Laínez,
members of the research group CEPIMA is also much appreciated (Chemical En-
gineering Department, Universitat Politècnica de Catalunya).
References
1 Howard, B. B., Upton, M., Introduc- 17 Türkay, M., Oruç, C., Fujita, K.,
tion to Business Finance, McGraw-Hill, Asakura, T., Comput. Chem. Eng. 28
New York, 1953. (2004), pp. 985–992.
2 Romero, J., Badell, M., Bagajewicz, 18 Jun-Hyung, R., Dua, V., Pistikopou-
M., Puigjaner, L., Ind. Eng. Chem. los, E. N., Comput. Chem. Eng. 28
Res. 42 (2003), pp. 6125–6134. (2004), pp. 1121–1129.
3 Shapiro, J. F., Modeling the Supply 19 Chen, C., Lee, W., Comput. Chem.
Chain, Duxbury, 2001. Eng. 28 (2004), pp. 1131–1144.
4 Yi, G., Reklaitis, G. V., AIChE J. 49 20 Seferlis, P., Giannelos, N., Comput.
(2003), pp. 3084–3094. Chem. Eng. 28 (2004), pp. 799–809.
5 Yi, G., Reklaitis, G. V., AIChE J. 50 21 Badell, M., Romero, J., Puigjaner,
(2004), pp. 2849–2865. L., Comput. Chem. Eng. 28 (2004), pp.
6 Guillén, G., Badell, M., Espuña, A., 45–61.
Puigjaner, L., Comput. Chem. Eng. 30 22 Gupta, A., Maranas, C. D., Comput.
(2006), pp. 421–436. Chem. Eng. 27 (2003), pp. 1219–1227.
7 Applequist, G. E., Pekny, J. F.,
23 Mokashi, S. D., Kokossis, A. C., Com-
Reklaitis, G. V., Comput. Chem. Eng.
put. Chem. Eng. 27 (2003), pp. 927–
24 (2000), pp. 2211–2222.
949.
8 Shapiro, J. F., Comput. Chem. Eng. 28
24 Gupta, A., Maranas, C. D., MacDon-
(2004), pp. 855–861.
ald, C. M., Comput. Chem. Eng. 24
9 Grossmann, I. E., Comput. Chem. (2000), pp. 2613–2621.
Eng. 29 (2004), pp. 29–39.
25 Zhou, Z., Cheng, S., Hua, B., Com-
10 Shah, N., Comput. Chem. Eng. 29 put. Chem. Eng. 24 (2000), pp. 1151–
(2005), pp. 1225–1235. 1158.
11 Shulz, P., Diaz, M., Bandoni, J. A.,
26 Bok, J. W., Grossmann, I. E., Park,
Comput. Chem. Eng. 29 (2005), pp.
S., Ind. Eng. Chem. Res. 39 (2000), pp.
1305–1316.
1279–1290.
12 Wan, X., Pekny, J. F., Reklaitis, G. V.,
27 Gjerdrum, J., Shah, N., Papageor-
giou, L. G., Ind. Eng. Chem. Res. 40
1317–1328.
(2001), pp. 1650–1660.
13 Jung, J. Y., Blau, G., Pekny, J. F.,
28 Lababidi, H. M., Ashmed, M. A., Ala-
Reklaitis, G. V., Eversdyk, D., Com-
tiqi, I. M., Al-Enzi, A. F., Ind. Eng.
put. Chem. Eng. 28 (2004), pp. 2087–
Chem. Res. 43 (2004), pp. 63–73.
2106.
14 Lasschuit, W., Thijssen, N., Comput. 29 Papageorgiou, L. G., Rotstein, G. E.,
Chem. Eng. 28 (2004), pp. 863–870. Shah, N., Ind. Eng. Chem. Res. 40
(2001), pp. 275–286.
15 Neiro, S. M. S., Pinto, J. M., Comput.
Chem. Eng. 28 (2004), pp. 871–896. 30 Sundaramamoorthy, A., Karimi,
I. A., Ind. Eng. Chem. Res. 43 (2004),
16 Berning, G., Brandenburg, M.,
pp. 8293–8306.
Gürsoy, K., Kussi, J. S., Mehta, V.,
Tölle, F.-J., OR Spektrum 28 (2004), 31 Oh, H., Karimi, I. A., Ind. Eng. Chem.
pp. 913–927. Res. 43 (2004), pp. 3364–3380.
32 Perea-López, E., Ydstie, E., Gross- 47 Grant, J. L., Foundations of Economic

mann, I., Comput. Chem. Eng. 27 Value Added, Wiley, New York, 2003.
(2003), pp. 1201–1218. 48 Pratt, S. P., Cost of Capital: Estima-
33 Tsiakis, P., Shah, N., Pantelides, tion and Applications, Wiley, New York,
C. C., Ind. Eng. Chem. Res. 40 (2001), 2002.
pp. 3585–3604. 49 Sharpe, W. F., Portfolio Theory and
34 Guillén, G., Mele, F. D., Bagajew- Capital Markets, McGraw-Hill, New
icz, M. J., Espuña, A., Puigjaner, L., York, 1999.
Chem. Eng. Sci. 60 (2005), pp. 1535– 50 Brooke, A., Kendrik, D., Meeraus,
1553. A., Raman, R., Rosenthal, R. E.,
35 Fandel, G., Stammen, M., Int. J. Prod. GAMS – A User’s Guide, GAMS De-
Econ. 89 (2004), pp. 293–308. velopment Corporation, Washington,
36 Cakravastia, A., Toha, I. S., Naka- 1998.
mura, N., Int. J. Prod. Econ. 80 (2002), 51 Haimes, Y. Y., Lasdon, L. S., Wismer,
pp. 231–248. D. A., IEEE Trans. Syst., Man Cybern. 1
37 Hugo, A., Pistikopoulos, E. N., En- (1971), pp. 296–297.
vironmentally conscious planning 52 Geoffrion, A. M., Graves, G. W., Lee,
and design of supply chain networks. S. J., “Strategic distribution system
In Process Systems Engineering 2003, planning: A status report”, Techni-
Chen, B., Westerberg, A. W., eds., cal Report, Hax, A. C., ed., North-
Elsevier, Amsterdam, 2004. Holland, Amsterdam, 1978.
38 Guillén, G., Mele, F., Bagajewicz, 53 Brown, G. G., Graves, G. W., Hon-
M., Espuña, A., Puigjaner, L., Chem. czarenko, M., Manage. Sci. 33 (1987),
Eng. Sci. 60 (2005), pp. 1535–1553. pp. 1469–1480.
39 Guillén, G., Bagajewicz, M., Se- 54 Yan, H., Yu, Z., Cheng, T. E., Com-
queira, S. E., Espuña, A., Puigjaner, put. Oper. Res. 30 (2003), pp. 2135–
L., Ind. Eng. Chem. Res. 44 (2005), pp. 2155.
557–575.
55 Petkov, S. B., Maranas, C. D., AIChE
40 Guillén, G., Pina, C., Espuña, A., J. 44 (1998), p. 896.
Puigjaner, L., Ind. Eng. Chem. Res. 44
56 Ahmed, S., Sahinidis, N. V., Ind. Eng.
(2005), pp. 7405–7419.
Chem. Res. 37 (1998), pp. 1883–1892.
41 Guillén, G., Mele, F., Espuña, A.,
57 Cohen, M. A., Lee, H. L., Manufac-
Puigjaner, L., Ind. Eng. Chem. Res.
turing strategy: Concepts and Meth-
45(22) (2006), pp. 7566–7581.
ods, in: The Management of Productiv-
42 Guillén, G., Espuña, A., Puigjaner, ity and Technology in Manufacturing,
L., AIChE J. 52(11) (2006), pp. 3864– Kleindorfer, P. R. (ed.), Plenum,
3881. New York, 1985.
43 Mele, F., Guillén, G., Espuña, A., 58 Balasubramanian, J., Grossmann,
Puigjaner, L., Ind. Eng. Chem. Res. I. E., Ind. Eng. Chem. Res. 43 (2004),
45(9) (2006), pp. 3133–3148. pp. 3695–3713.
44 Srinivasan, V., Omega 14 (1986), pp. 59 Petkov, S. B., Maranas, C. D., Ind.
145–166. Eng. Chem. Res. 36 (1997), pp. 4864–
45 Orgler, Y. E., Manage. Sci. 20(10) 4881.
(1969), p. 1350. 60 Gupta, A., Maranas, C. D., Ind. Eng.
46 Weissenrieder, F., “Value Based Chem. Res. 39 (2000), pp. 3799–3813.
Management: Economic Value Added
61 Birge, Z., Louveaux, S., Principles
or Cash Value Added?”, Technical
on Stochastic Programming, Springer,
Report, Department of Economics,
Berlin, 1997.
Gothenburg University, Sweden,
1998.
Appendix 41
62 Barbaro, A. F., Bagajewicz, M. J., 64 Eppen, G. D., Martin, R. K.,

AIChE J. 50 (2004), pp. 963–989. Schrage, L., Oper. Res. 37 (1989),
63 Bonfill, A., Bagajewicz, M., Espuña, pp. 517–527.
A., Puigjaner, L., Ind. Eng. Chem. Res.
43 (2004), pp. 741–750.
Appendix
Notation
Indices
e suppliers
d discount interval
i products
j plant equipment
m markets
r raw materials
s manufacturing sites
t planning periods
w distribution centers
Sets
DRe raw materials for which supplier e offers quantity discounts
Edisc set of suppliers that offer quantity discount
Er set of suppliers e that provide raw material r
Ij products that can be processed in plant equipment j
Ir products that consumed raw material r
Ji equipment that can process product i
Re set of raw materials provided by supplier e
Parameters
Aert maximum availability of raw material r in period t associated with
supplier e
CLinemax upper bound of short term credit line
DttMS technical coefficient for investments in marketable securities
DFerd discount factor associated with discount interval d for raw material
r offered by external supplier e
Demimt demand of product i at market m in period t
EttMS technical coefficient for sales of marketable securities
E[ROE] expected return on equity
FCFSj st fixed cost per unit of capacity of plant equipment j at site s in
period t
FCFWwt fixed cost per unit of capacity of distribution center w in period t
irt interest rate of long term debt
IstS investment required to open site s in period t

W
Iwt investment required to open distribution center w in period t
MinCash lower bound of cash
Othert other expected outflows or inflows of cash in period t
Priceimt price of product i at market m in period t
PriceFS
j st investment required per unit of capacity of equipment j increased
at site s in period t
PriceFW
wt investment required per unit of capacity of distribution center w
in-creased in period t
r0 risk free rate of return
rp risk premium rate
RR return rate
StMS marketable securities of the initial portfolio maturing in period t
trate tax rate
T length of planning horizon
Binary variables
Fertd 1 if the amount of raw material r purchased from supplier e in
period t
is within discount interval d, 0 otherwise
SBst 1 if site s is opened in period t, 0 otherwise
Vj st 1 if the capacity of equipment j is increased at site s in period t, 0
otherwise
W Bwt 1 if distribution center w is opened in period t, 0 otherwise
Xwt 1 if the capacity of distribution center w is increased in period t, 0
otherwise
Continuous variables
AP ayt amount of accounts payable in period t
ARect amount of accounts receivable in period t
ASalestt sales executed in period t and receivable in period t
Borrowt total amount borrowed from the short term credit line in period t
Capitalt capital supported by shareholders in period t
Casht cash in period t
CLinet short term debt in period t
CV corporate value at the end of the planning horizon
Dept depreciation in period t
DFCF sum of discounted free cash flows at the end of the planning hori-
zon
ECasht exogenous cash in period t
EPurchet economic value of purchases executed in period t to supplier e
EPurchrm
et amount of money payable to supplier e in period t associated with
consumption of raw materials
EPurchtr
et amount of money payable to supplier e in period t associated with
consumption of transport services
Appendix 43
pr
EPurchet amount of money payable to supplier e in period t associated with
consumption of production utilities
ESalest economic value of sales carried out in period t
FAssett increment in fixed assets in period t
FCFt free cash flows in period t
FCostt fixed cost in period t
FExt other financial expenses and incomes in period t
FSj st total capacity of plant equipment j during period t at site s
FSEj st capacity increment of plant equipment j during period t at site s
FWj st total capacity at distribution center w during period t
FWEj st capacity increment at distribution center w during period t
LBorrowt total amount of money borrowed from he long term credit line in
period t
LDebtt long term debt in period t
LRepayt total amount repaid to the long term credit line in period t
NetCLine
t total amount of money borrowed or repaid to the short term credit
line in period t
NetLDebt
t total amount of money borrowed or repaid to the long term credit
line in period t
NetMS
t total amount received or paid in securities transactions in period t
NetDebtT net total debt at final period T
NetInvestt net investment in period t
NPV net present value computed for whole planning horizon
Pij st production rate of product i in equipment j at site s in period t
Paytt payments to external supplier e executed in period t on accounts
payable incurred in period t
Pledtt amount pledged within period t on accounts receivable maturing
in period t
Profitt profit achieved in period t
Purchrm
erst amount of raw material r
purchased to supplier e at site s in period t
Purchrm
et amount of raw material r
purchased in period t from external supplier e
qerdt amount of material r within discount
interval d purchased to supplier e in period t
Qiwst amount of product i sent from site s to distribution center w in
period t
Repayt total amount repaid to the short term credit line in period t
Salesiwmt amount of product i sold from distribution center w in market m
in period t
SIrst amount of stock of raw material r at site s in period t
SOist amount of stock of product i at site s in period t
SWiwt amount of stock of product i at distribution center w in period t
TProfit total profit achieved at end of planning horizon
WACCt weighted average cost of capital in period t
ϒttMS
cash invested in period t in marketable securities maturing in pe-
riod t
ZttMS
security sold in period t maturing in period t
APayt change in amount of accounts payable in period t
ARect change in amount of accounts receivable in period t
I nv change in inventory value in period t
NWCt change in net working capital in period t
Greek symbols
γe maximum delay on payments of supplier e
α max maximum delay in receivables at market m
η interest of short term debt
βett technical discount coefficient for payments to external supplier e
executed in period t
ψerd discount factor associated with discount interval d for raw material
r offered by external supplier e
πert cost of raw material r offered by external supplier e in time period
t
δmtt fraction of sales carried out in period t that are receivable in period
t in market m
plant
νij capacity consumption of plant equipment j by product i
φtt face value of accounts maturing in period t pledged in period t
ψert price of raw material r offered by external supplier e in time pe-
riod t
λt proportion of equity over total capital investment in period t
ware
νiw specific volume of product i in warehouse w
tr1
ρeiws unitary transport costs of product i from plant s to warehouse w
payable to external supplier e
tr2
ρeiwm unitary transport costs of product i from warehouse w to market
m payable to external supplier e
υi specific volume of product i
τijut1
se cost of the utilities associated with product i manufactured with
equipment j in site s and payable to external supplier e
ut2
τrse cost of the utilities associated with handling the inventory of raw
material r in site s and payable to external supplier e
ut3
τise cost of the utilities associated with handling the inventory of final
product i in site s and payable to external supplier e
ut3
τiwe cost of the utilities associated with handling the inventory of final
product i in warehouse w and payable to external supplier e
Superscripts
L lower bound
U upper bound
45
2
Multiagent Framework for Modeling and Design of
Supply Chain Networks
Luis Puigjaner, Gonzalo Guillén-Gosálbez, Fernando Daniel Mele
There is a large body of work on supply chain (SC) optimization in the chemical
process industry (CPI). However, some of the basic aspects of the optimization
problem are not adequately handled by the models and solution strategies devel-
oped so far in the literature. Specifically, the approaches that currently exist for
supply chain management (SCM) have a rather narrow scope and neglect impor-
tant business aspects such as financial and environmental concerns. Furthermore,
they usually apply fairly simple time representations and are unable to capture the
complex dynamics that characterizes the SCM problem. This chapter describes an
integrated solution strategy for SCM, which covers the aforementioned aspects and
implements the ideas and concepts developed in our research.
2.1
Introduction
Many of the approaches developed in the past in the area of SCM are not able to
consider the delays associated with the information that flows through a real SC.
These approaches, which usually involve using mixed integer modeling techniques
to address SC planning decisions, assume that the orders placed by customers are
immediately known by the entire SC and that the demand is satisfied as it arrives
to the system. It is believed that these kinds of models are not as insightful as those
in which the dynamic features of the SCs are explicitly taken into account. Specifi-
cally, at the operational level, models that capture the SC dynamics are particularly
appealing because they assist managers in understanding complex interactions be-
tween system states, data, and decision variables.
Another limitation of the tools that are available for SCM is that they have tradi-
tionally focused on myopic methods for manufacturing sites, logistics, and distri-
bution tasks rather than on approaches that consider all of these methods simul-
taneously. Specifically, a large body of literature in process systems engineering
(PSE) is concerned with the analysis of manufacturing plants and the computation
of optimal scheduling plans in terms of predefined criteria (usually makespan, cost
ISBN: 978-3-527-31693-9
46 2 Multiagent Framework for Modeling and Design of Supply Chain Networks
or profit) that satisfy a demand that comes directly from a customer [1] offer a de-
tailed review of these types of models. These approaches devised to date do not
consider important aspects of the problem, such as the buffering effect, demand
distortion, and storage costs, which are all due to the distribution channels through
which the materials flow. Neglecting these aspects may result in either infeasible
or suboptimal solutions for the entire SC.
2.2
A Discrete Event Simulator for SCM
Discrete event simulators offer an appealing framework for capturing the dynam-
ics of the SCM problem. These tools represent and model interactions between
the components of a generic SC network in a functional way, with the aim of fur-
ther obtaining a decision-making tool for SCM. In a discrete event-driven system
model of a SC, all the entities belonging to the SC are regarded as independent and
well-defined objects that are represented by a collection of states and transitions. A
simulation tool of this type can be used to model the behavior of decentralized sys-
tems, in which there is no global coordinator among the parties involved and each
of its entities makes decisions locally. The framework for the stochastic discrete
event simulator is shown in Fig. 2.4. It includes the following elements:
• State variables that describe a system at given points in time,
i.e., the simulation outputs.
• The simulation model, which includes the equations or other
relationships describing how the state variables change over
time as a function of decisions or external events.
• The sampling tool, which provides a representative sample
from the multivariate probabilistic distribution that
characterizes the uncertain parameters.
2.2.1
Simulation Model
This section describes how to derive a generic discrete event simulation model for
SCM. As mentioned before, the tool envisaged considers the SC as a decentralized
system where there is no global coordinator and every entity in it makes decisions
locally. The discrete event simulator can be described by means of states and tran-
sitions diagrams, as a state machine. Specifically, the way of implementing such
model in MATLAB [2], taking full advantage of two of its toolboxes – Stateflow and
Simulink – is next illustrated. Stateflow allows us to represent the system by means
of states and transitions diagrams, and Simulink offers the simulation environ-
ment.
The system can be constructed by putting together generic units that map the op-
eration of each of the entities embedded in the SC. The basic model for the generic
entity comprises a set of four states that work in a parallel manner: Order_reception,
Fig. 2.1 Scheme of a SC generic unit.
Fig. 2.2 Generic unit in Stateflow.
Material_reception, Delivering, and Ordering, as it is outlined in Fig. 2.1. Likewise,

Fig. 2.2 shows a more detailed representation of the generic unit as it can be seen
in the Stateflow environment.
Figure 2.3 shows the working mode of the state Order_reception, which is acti-
vated whenever a request-for-material (RFM) arrives to the system. The boxes rep-
resent each of the two states among which the system switches. In the figure, each
circle is a transition node where logical conditions are evaluated, and every line rep-
resents a logical condition that must be satisfied to go from one state to another.
The default or starting state for the plant is the Idle state. The plant will move from
the Idle state if the logical condition “ORin{demand}/a = b” becomes true. Further-
more, any new event ORin that arrives to the system makes the state of the system
change from Idle to Inv1 and generates also a new event called demand. The quan-
tity associated with a new order, b, is assigned to the internal variable a. The transi-
tion to the state Inv1 is effective when the condition “[(IF >= 10)&&(IG >= 10)]”
Fig. 2.3 Substates and transitions for the state Order_reception

(from the example with nine entities).
Fig. 2.4 Framework for stochastic simulation.
is satisfied. If the latter condition is not satisfied (i.e., “[(IF < 10)|(IG < 10)]”), the
system will remain in state Idle, and will also store in the variable O the quantity
required. O denotes the backlogged orders, whereas IF and IG are the inventory
levels of products F and G, respectively.
When the system switches to state Inv1, the counter of the time elapsed during
the inventory depletion, Inv1_c, is updated taking into account the current time t.
When this state is abandoned, the inventory levels of the different products, F, IF,
and G, IG, are depleted accordingly by 0.6 a and 0.4 a, respectively. The material
ready for transportation is then increased by a units (H+ = a). Furthermore, the
amount of material associated with any event ORin that may arrive to the system
while it is in in the state Inv1 will be stored in the variable O.
The condition “[t >= Inv1_c + 5]/a = b” is indeed a safety logical condition that
allows the system to leave state Inv1 if a maximum computation time is reached.
If there is not any backlogged order, “[O < 5],” the system returns to the state Idle.
Otherwise, the system enters again in state Inv1.
The generic unit described above is general enough to represent any of the
generic nodes embedded in a real SC, namely, a plant, a warehouse, or a retailer.
Furthermore, it can be connected with other units to model an entire SC. Some
key issues regarding the practical implementation of the discrete event simulation
model are next discussed in detail.
2.2.2
The Demand Model
The discrete event simulator requires the definition of a demand pattern, which
must be provided as input data into the system. The demand is represented by a
set of orders distributed over the time horizon of the study. Each of these orders
is characterized by a certain amount of material and time of occurrence. Different
demand models can be adopted to describe the size of the orders and the time
elapsed between two orders: (1) the sizes of the orders and the interarrival times
are constant; (2) the interarrival intervals are constant whereas the sizes follow a
known distribution law (usually a normal distribution); (3) both parameters are as-
sumed to follow known probability distributions (usually the sizes follow a normal
distribution whereas the interarrival intervals are modeled by a Poisson process
[3]).
2.2.3
The Inventory Model
The inventory policies are strategies that are implemented to give a quick answer
to questions such as: “when should a replenishment order be placed?” and “how
large should the replenishment order be?”. There are basically two main types of
strategies: continuous and periodic. In a continuous review strategy, an analysis of
the inventory status is carried out every time a transaction (shipment, receipt, etc.)
occurs. In periodic review, the stock status is only checked every τ time units re-
gardless of the events coming to the system. τ represents the time period between
two consecutive inventory reviews. The main inventory control strategies inside
each group are as follows.
1. Order-Point, (s, k) System. The (s, k) System is a continuous
review system in which the replenishment is made
whenever the inventory position drops below the reorder
point s. The replenishment quantity is proportional to the
difference between the reorder point and the current
inventory level. In this case, u = k(s − I ), where u is the
replenishment quantity, I is the current inventory position1) ,
and k is a parameter.
2. Order-Point, Order-Quantity, (s, Q) System. The (s, Q)
system is also a continuous review strategy, in which the
replenishment quantity is a fixed quantity Q; then u = Q.
3. Periodic-Review, Order-Up-To-Level, (τ, S) System. In the
(τ, S) system, an order is placed every τ units to raise the
1) Inventory position is usually understood as the On-hand stock, i.e., the stock physically on the shelf,
plus the On-order stock, minus the Backorders and minus the Committed stock, which is set aside by
virtue of contractual obligations.
inventory position to the level S. The replenishment quantity

is then u = S − I .
4. (τ, s, S) System. The (τ, s, S) system is a combination of a
(τ, S) system and a reorder point strategy. Every τ time
units, the inventory position I is checked. If I is below the
reorder point s, a replenishment quantity u = S − I is
ordered to raise the stock level to S. If the position is above s,
nothing is done until the next review.
5. (τ, s, k) System. Here, the replenishment quantity is equal to
u = k(s − I ), and the inventory position is checked only
every τ time units.
Let us notice that the parameters of each of these strategies, namely k, s, Q, τ

and S, should be tuned so as to minimize the inventory costs.
2.3
A Multiagent System for SCM
The capabilities of the discrete event simulator can be extended through the use
of multiagent systems. Certainly, the effectiveness of multiagent systems for solv-
ing complex problems has recently led to software agents being used to model and
optimize the SC operation. SC networks are truly complex systems whose good per-
formance relies on coordinating a number of entities. For this reason, approaches
that consider the SC in terms of agents and discrete event-based simulation may be
better suited to this kind of problem [4, 5]. These approaches model not only the dy-
namics of the individual components but also the interactions between them and,
since they mimic the natural structure of a SC and the interacting mechanism of
its entities, they have the additional advantage of being easily reconfigurable when
the chain structure changes. Furthermore, agents are a very effective technique
for designing distributed SC systems over the Internet, which is a very important
channel for doing business and sharing business-related information in a seamless
manner.
Given these advantages, multiagent systems have in the last few years become
a promising tool for solving SC problems [5–9], as they are a highly suitable way
of driving dinamic simulations. Compared with the many efforts to design other
kinds of approaches to SCM, research on agent-based systems is still very limited,
especially in the CPI sector, and usually focuses on partial problems, such as coop-
erative decision support or distributed simulation. Very few studies have addressed
both these problems simultaneously. Furthermore, many agent-based systems suf-
fer from a lack of optimization skills. To avoid this, external optimization tools must
be used to eliminate the need for a random trial-and-error search and to allow the
optimization procedure to be automated.
During these last years, part of our research effort has been devoted to developing
our own multiagent framework for SCM. The structure of the proposed system is
Fig. 2.5 Multiagent framework.
shown in Fig. 2.5. It consists of a system of real agents, a central agent, a simulator,
and a set of modules. This is an ambitious framework that aims to cover the entire
range of management tasks that must be accomplished by a SC network. A brief
overview of the multiagent system is provided next. Its main parts are the following.
• Real agents. They are software agents that work in a
computer network. Each maps one real-world entity or node
in the SC network, communicates with it, and stores all the
data concerning this entity. Depending on the SC manager
or decision maker, the real-agent system can solve particular
problems, update its database, suggest improvements, and
so forth.
• Real central agent. It can make decisions based on certain SC
models and built-in procedures, and then test the
conclusions over the simulation model. One of the
procedures for the central agent to make decisions is to run
experiments using the simulation model. The central agent
manages the information stored in the network and makes
decisions over the real-SC network. It is also through the
central agent that the external customers and suppliers
communicate with the system and negotiation processes are
carried out. This exchange of information may be done on
the Internet.
• Simulated agents. They make up a SC simulator that
emulates the behavior of the real-world SC and therefore the
real-agent system. The central agent can use the simulator to
test the effects of the decisions over the entire SC before the
latter is implemented in the real environment. The simulator
can be used by the central agent or by managers, who can

use the simulator as a stand-alone application. The
simulated agents have an operational model that emulates
the behavior of the entities in a real-world SC.
• Modules. Both the simulator and the real agents may use
some of the modular tools that the framework provides to
perform specific tasks. The modules may be demand
forecasters, negotiation algorithms, optimization packages,
scheduling software, performance index evaluators, and so
on.
Both the system of real agents and the system of simulated agents comprise two
categories of objects: software agents and messages. Of the agents that take part
in this system (see Fig. 2.6), the emulation agents represent the physical entities,
facilities, or nodes in the real-world SC, such as suppliers, customers, manufac-
turing plants, distribution centers, etc. Each of these agents includes a number
of subagents, namely inventory, sales, production, purchasing, and transportation.
Moreover, the system has agents other than emulation agents, such as the central
agent, which does not have any correspondence with a real entity in the chain. The
messages are objects that are exchanged between the agents (internal messages)
and also with other entities, such as the real entities of the chain, the user, and/or
the external customers (external messages). All of these messages represent mate-
rial, information, or cash flows.
The agent-based simulator provides an environment in which all business
processes can be emulated, such as sales, purchases, negotiation with customers
and suppliers, production and logistics planning at different levels, investment
Fig. 2.6 The multiagent system.

analysis, transportation and production scheduling, inventory management, de-

mand forecasting, etc. Agents model all the SC entities, which include manufac-
turers, third-party logistics, the internal departments of a company, etc. Each agent
may have one or more subagents The subagents represent internal departments
that have generic functions in all companies, which can be customized according
to the practices of a company. A general overview of the different agents embedded
in the multiagent framework is next given.
2.3.1
The Site Agent
Figures 2.7 and 2.8 represent site agents and their corresponding subagents and
messages. Particularly, Fig. 2.7 shows a storage agent and Fig. 2.8 depicts a fac-
tory agent. This last agent has one more subagent, which is called the production
subagent. The storage agent can be used to represent, for example, warehouses,
distribution centers, retailers, or any other nonproductive entity in the SC.
One of the processes that can take place in the factory agent is next described, as
it is depicted in Fig. 2.8.
When a request-for-quote (RFQ) is received (message ORD_in in Fig. 2.8), the
agent forwards the message to its sales subagent (message RFQ). Then, it creates
an internal request for information and sends it to the inventory and the produc-
tion subagents (messages RFQ between the sales and the inventory subagents and
Fig. 2.7 Storage agent.

Fig. 2.8 Factory agent.
between the sales and the production ones, respectively). Both subagents, the in-
ventory and the production ones, answer to the sales subagent through messages
RRFQ, which are broadcasted to the customer. The customer can then send a mes-
sage Confirm that arrives to the sales subagent as a message RFM. This message
is next sent to the inventory subagent. The inventory subagent depletes then its
level by the required quantity or asks for material to the production subagent, and
finally sends a material message to the transportation subagent, which delivers the
required quantity to the customer through a message MAT_out. The transportation
subagent also informs the sales subagent about the delivery through the message
Notice_deliver. In Fig. 2.8, the relationships between the production subagent and
the other subagents are illustrated in a quite simplified manner. Figure 2.14 shows
a more realistic representation of the activity of this subagent.
2.3.2
The Sales Subagent
Figure 2.9 depicts the relationships between a subagent of a storage of end products
of a factory and the other SC agents, whereas Figs. 2.10 and 2.11 represent the
internal states an transitions of this subagent. Note that the messages exchanged
between agents may change slightly depending on the type of facility in which the
subagent is located, i.e., storage of raw materials, storage of end products, etc.
Fig. 2.9 Sales subagent (a).
Fig. 2.10 Sales subagent (b).

Fig. 2.11 Sales subagent (c).
Figures 2.10 outlines how the sales subagent respond to the arriving messages
RFQ or RRFQ. The former messages represent queries regarding the availability
of a specific product and are sent by customers. The latter are the messages that
the inventory subagent sends in response to the queries made by the customers.
The boxes in this figure (Idle, Managing RFQ, Managing RRFQ, and Choosing RFQ)
represent states, whereas the arrows denote transitions. When the conditions asso-
ciated with the arrows are satisfied, the system is allowed to switch from one state
to another.
Furthermore, Fig. 2.11 depicts how the sales subagent reacts against messages
RFM and Notice_deliver. The first message, which can only be sent by the cus-
tomers, is a direct RFM. The second one is the notification message that the trans-
portation subagent sends after sending material to an external entity.
2.3.3
The Inventory Subagent
Figure 2.12 shows the links between an inventory subagent of a warehouse and the
remaining agents. Figure 2.13 describes the way in which the inventory subagent
reacts against the incoming messages, namely, RFQ, RFM, and MAT_in. The first
one is a query sent by the Sales subagent regarding the availability of material. The
second one represents a RFM made also by the sales subagent. The third one is
sent when a certain amount of material sent from an external supplier arrives to
the system to replenish the inventory level.
Fig. 2.12 Inventory subagent (a).
Fig. 2.13 Inventory subagent (b).

Fig. 2.14 Production subagent (a).
2.3.4
The Production Subagent
Each manufacturing plant of the multiagent system includes three different areas:
the storage of raw materials, the production, and the storage of final products (see
Fig. 2.14). The two storage areas can be represented using a generic site agent
scheme, as depicted in Fig. 2.7. On the other hand, the production area is repre-
sented by a production subagent that does not have any correspondence with the
agents or subagents described so far.
The production subagent models the production department of a SC node. Even-
tually, it can have additional subagents to model the various production lines or
facilities that can be encountered in a plant. This subagent calls for bids whenever
it receives a RFQ from the sales subagent of the storage of end products (message
RFQSP in Fig. 2.14). These bids are sent from the line subagents that represent the
production facilities. On the basis of these bids, it then generates a response to the
sales subagent (message RRFQPS). The production subagent also procures materi-
als from the storage of raw materials (message ORD_out) and allocates them to the
lines based on their needs. The model for the production subagents can be either
as simple as a capacity factor per unit of manufactured product or as complex as a
local optimization tool to schedule the activities in the production lines.
Figure 2.15 depicts the way in which the production subagent behaves on arriv-
ing messages RFQ, ORD_in, RORD_out, and MAT_in. RFQ represents a request
for information regarding the availability of a given amount of product. ORD_in is
Fig. 2.15 Production subagent (b).
a request for a given quantity of product. RORD_out is a message communicating

the unavailability of some products, and it is sent from the storage of raw materials.
Finally, MAT_in represents the arrival of raw materials required to start production
at the production lines.
2.3.5
The Purchasing Subagent
This subagent manages the purchasing tasks of the site agents. Specifically, it pe-
riodically asks for materials by sending requests to all the possible suppliers. Fur-
Fig. 2.16 Purchasing subagent (a).
Fig. 2.17 Purchasing subagent (b).
thermore, it is also responsible for choosing the best offer according to a given
criterion.
Thus, the purchasing subagent attaches a site to a network of suppliers. Figure
2.16 represents the environment of a purchasing subagent belonging to a retailer
facility. Figure 2.17 shows how the subagent reacts to different arriving events.
Fig. 2.18 Transportation subagent.
In this figure, message RORD_out is a set of offers sent by the suppliers to the
purchasing subagent. After receiving these offers, the subagent has to choose one
option and elaborate a message Confirm, which should be sent to the supplier that
has been selected. Notice_purch represents the message that the inventory subagent
sends to the purchasing subagent after receiving the materials sent by the suppli-
ers. Review is the message that the inventory subagent sends when a depletion of
the inventory level occurs. This message is used by the purchasing subagent if the
inventory control policy is continuous.
2.3.6
The Transportation Subagent
Figure 2.18 shows the transportation subagent of a retailer site. This subagent can
exchange three main messages: RRFM, MAT_out, and Notice_deliv. RRFM is the
RFM sent from the sales subagent. MAT_out represents the delivery of material
to the customers. Finally, Notice_deliv is a message sent by the transportation sub-
agent to the sales subagent upon delivery of the required products. The internal
behavior of this subagent depends on the kind of transportation strategy used in
the network (e.g., third-party logistics, own fleet).
One of the main shortcomings of the multiagent system previously described is
the lack of optimization skills, i.e., although system managers could play what-if
scenarios with input data and simulation models to evaluate alternative solutions,
there is still a need to systematize the optimization procedure and avoid a random
trial-and-error search. Moreover, the scope of the multiagent system is limited, as it
focused on process operations and neglected other important business aspects. In

the sections that follow, we describe how to extend the capabilities of the multiagent
framework to overcome the above-commented drawbacks.
2.4
Novel Business Aspects in SCM
This section deals with the efforts made to extend the scope of the SC analysis
with the aim of obtaining a holistic representation of the system. Specifically, it
focuses on two major areas: environmental impact and finances. To achieve inte-
gration between diverse areas of a company, a set of software modules must be
developed and incorporated into the overall multiagent system. The reasons for in-
cluding these business aspects in the SC analysis and a detailed explanation of the
associated software modules are given below.
2.4.1
Environmental Impact
Traditionally, the optimization models devised by the PSE community to assist op-
eration and design in the CPI industry have concentrated on finding the solution
that maximizes a given economic performance indicator while satisfying a set of
mass balances, assignment, and capacity constraints imposed by the topology of
the plant. In recent years, however, there has been a growing awareness of the
importance of including environmental aspects, along with traditional economic
criteria, in the optimization procedure.
The first attempts to incorporate environmental aspects applied fairly myopic
strategies that focused on minimizing the emissions from a plant [10–13]. Due to
their limited scope, they sometimes lead to solutions that reduced these emissions
at the expense of increasing burdens elsewhere in the lifecycle, in such a manner
that overall environmental impacts increased [14].
Lifecycle assessment (LCA) arose in response to this situation. LCA is an ob-
jective process for evaluating the environmental loads associated with a product,
process, or activity. It identifies and quantifies the energy and materials used and
the waste released to the environment, and evaluates and implements opportuni-
ties for effecting environmental improvements. The assessment covers the entire
lifecycle of the product, process, or activity, including extracting and processing
raw materials; manufacturing, transportation, and distribution; reuse and mainte-
nance; and recycling and final disposal [15]. The essence of LCA is that it considers
all material and energy flows from the “cradle” of primary resources (such as oil or
ore deposits) to the “grave” of final disposal (such as stable inert material in a land-
fill). Today, LCA can be said to be the main instrument in environmental supply
chain management (ESCM) as it can be effectively used to restructure SCs in order
to improve their environmental performance [14, 16–20].
Fig. 2.19 Framework for the Lifecycle Assessment.
2.4.1.1 The Environmental Module

Environmental concerns can be incorporated into the multiagent system through
the insertion of a special software module. This module allows the computation
of several environmental impact indexes according to the principles of the LCA.
The calculation of environmental impact indexes is based on the second and third
phases of the LCA methodology, as defined in the ISO standards [21, 22] (see Fig.
2.19). The module uses the information associated with the SC simulation, i.e.,
raw material selection and resources consumed by the labor tasks, to perform the
inventory analysis phase and then to evaluate a global environmental impact index
according to the lifecycle inventory assessment phase. Thus, the LCA-based module
constitutes a novel approach to consider environmental aspects in the SCM field.
It is able to perform a study that is fully adapted to the SC network, taking into
account features that include the material and energy links between facilities and
the processes the product undergoes once it leaves the company.
Let us note that there are several software programs, apart from the environmen-
tal module of the multiagent system itself, that implement the LCA methodology
following the ISO 14040. Some of them are PEMS 4 Database (1998)2) , TEAM and
DEAM (1998)3) , and SimaPro 6 LCA software (2004)4) . All of these programs are
able to compute material and energy balances and quantify the burdens and im-
pacts along the lifecycle of a process or product. The material and energy balances
for the process itself can also be performed by means of specific software programs
2) PIRA International PIRA, Leatherhead, UK, 31 Oct. 2005,

<www.pira.co.uk/pack/lca_software.htm>.
3) Ecobalance, UK, The Ecobilan Group, Arundel, UK, 31 Oct. 2005,
<www.ecobalance.com/uk_team.php>.
4) PRé Consultants bv., The Netherlands, 31 Oct. 2005, <www.pre.nl/simapro/default.htm>.
originally devised to assist in the design and operation of chemical processes. Nev-
ertheless, these general-purpose software programs are not capable of properly as-
sessing the environmental impact of a SC. This has motivated the development of
a specific environmental module for the multiagent system.
The application of LCA in the multiagent system only considers the first three
phases of the LCA methodological framework (ISO 14040 series). Figure 2.19
shows these phases schematically, as well as the inputs and outputs required by
each phase and the relationships between them.
• Goal and scope definition. In this phase, the system
boundaries and the object of study are set to ensure that the
relevant parts of the system are included.
• Inventory analysis, in which mass and energy balances are
performed to quantify all the material and energy inputs,
waste and emissions from the system, i.e., the
environmental burdens.
• Impact assessment. The environmental burdens quantified
in the inventory analysis are added to a limited set of
recognized environmental impact categories.
• Interpretation. This is the final step, in which the results
suggest policies aiming to reduce the environmental impacts
associated with the product or process.
Selecting the environmental categories that are to be considered in the impact

assessment phase is a particularly important point. The following list gives the
definitions of the environmental impacts that have been applied in the multiagent
framework.
• The resource depletion potential (RDP) describes the
depletion of nonrenewable resources, e.g., fossil fuels,
metals, and minerals, in relation to the world’s estimated
reserves.
• The global warming potential (GWP) is attributed to the
emissions of greenhouse gases, e.g., CO2 , N2 O, CH4 , and
other volatile organic compounds (VOC). GWP factors for
different greenhouse gases are expressed relative to the GWP
factor of CO2 , which is therefore defined to be unity.
• The ozone depletion potential (ODP) indicates the potential
of chlorofluorocarbons and chlorinated hydrocarbons for
depleting the ozone layer. The ODP factors of each of the
ozone-depleting substance are expressed relative to the
ozone depletion potential of CFC-11.
• The acidification potential (AP) is based on the contributions
of SO2 , NOx , HCl, NH3 , and HF to potential acid deposition,
i.e., on their potential to form H+ ions.
• The eutrophication (or nutrification) potential (EP) is defined
as the potential to cause over-fertilization of water and soil,
Pw
Pr R ecycle
Fss
Ps11
Super S upplier Wr
Ps12
Supplier Pss 1 Pm1 Po Pu
Manufacturing Other W as te
Use
W ss Ws1 process Manag-
ement
S upplier Ps2 Wo Wu
Fs2 2 Pm2 Ww
Wm
W s2
Fig. 2.20 Elementary supply chain representation utilized in the

environmental module case study.
which can result in increased growth of biomass. Emissions

of species such as NOx , NH+ 3−
4 , PO4 , P4 , together with the
chemical oxygen demand (COD), are considered to be
responsible for eutrophication. EP is expressed relative to
PO3−4 .
• The photochemical oxidants creation potential (POCP) is
also known as photochemical smog. It is thought to be
caused primarily by VOCs, including alkanes, halogenated
hydrocarbons, alcohols, ketones, esters, ethers, olefins,
alkynes, aromatics, and aldehydes. POCPs of these species
are expressed relative to the POCP of ethylene.
• The human toxicity potential (HTP) is related to releases to
air, water, and soil that are toxic to humans. The toxicological
factors are calculated using the acceptable daily intake or the
tolerable daily intake of toxic substances. Human
toxicological factors are still at an early stage of development,
so the HTP can only be taken as an indication and not as an
absolute measure of the toxicity potential.
The main features and advantages of the environmental module are next high-
lighted through a small case study, the structure of which is given in Fig. 2.20. This
simple network, which contains all the basic elements of a generic SC, will be used
to provide insights into the model contained in the environmental module.
For the sake of simplicity, a high degree of aggregation has been assumed in
this example. Thus, the case study only considers a reduced number of energy
and material streams. In fact, this SC representation can be considered to be the
result of a detailed modeling of a manufacturer agent representing a physical plant
embedded in the SC.
The study described in this section aims to obtain an ecolabel for product Pm1 .
This can be done by applying the LCA methodology guidelines to the SC system
under study. This procedure provides the environmental burden emissions asso-
ciated with the manufacturing process of the product. The way in which the LCA
framework can be applied to this specific example is next described in detail.
Fr
Pr R ecycle
Fss Pw
Ps11
Wr
S uper S upplier
S upplier Pss 1 Ps12
Pm1 Po Pu W as te
Other U se
Manufacturing
Wss Ws1 process Manage -
me nt
S upplier
2
Ps2 Wo Wu Ww
Fs2 Wm Pm2
Ws2
Fig. 2.21 Backward allocation.
In the first phase of the LCA, the functional unit (product or process) must be
identified. This “functionality” can always be expressed as an equivalent product
amount (in kg or MJ, according to the nature of the product), which aims to facili-
tate later calculations. The system boundaries for this example are depicted in Fig.
2.20. In some cases it will be necessary to go deeper inside the boundaries to find
environmental values for the global streams across the boundaries.
The second phase of the LCA, the inventory analysis phase, is next applied to the
manufacturing process, which is represented by the source block of product Pm1 .
This stage requires input data concerning the input streams Ps12 and Ps2 , the emis-
sions of the process (Wm ), and the products Pm1 and Pm2 . The downstream inputs
and emissions of the manufacturing process (Wo , Wu , and Ww ) are also required
in this second stage. A key issue in the inventory calculation is that of determining
the proper allocation policy to be applied for allocating the environmental loads as-
sociated with each product. An allocation procedure is required only if the causal
relationship between inputs, outputs, and emissions is not known with certainty.
If this is not the case, the following general expression can be applied to allocate
the environmental loads:

Pk · vk = fk · (−W · vw + Fs · vs ) − Fop · Pp · vp . (2.1)
s p=k
In this expression, Pk represents the stream of product k and vk is the corre-

sponding ecovector associated with product k. W · vw is the waste stream weighted
by its ecovector, Fs · vs is an input stream multiplied by its ecovector, and Pp · vp
is the corresponding output stream weighted by its ecovector p. Finally, fk and fop
are allocation factors that depend on the allocation policy selected in each case (e.g.,
mass allocation, energy allocation).
The allocation to the chain placed on the left side of the Manufacturer, which
comprises the nodes labeled as Super Supplier, Supplier 1, and Supplier 2, is next
analyzed in the same manner done in the Manufacturing case. This procedure is
called forward allocation (Fig. 2.21) because the environmental load is carried from
the left to the right, that is to say, in the same direction as the material flow in the
SC.
Following the LCA philosophy, the environmental module also considers a back-
ward allocation (see Fig. 2.22), which is carried out in the direction opposite to the
SC material flow. Thus, this procedure reflects the fact that the manufacturer is also
Fr
Pr R ecycle
Fss Pw
Ps11
Wr
S uper S upplier
S upplier Pss 1 Ps12
P m1 Po Pu Was te
Other Us e
Ws1 Manufacturing Manage-
W ss proces s
ment
S upplier
Ps2 Wo Wu Ww
Fs2 2
Wm Pm2
Ws2
Fig. 2.22 Forward allocation.
responsible for the environmental impact generated by their products after leaving
the manufacturing stage. That is to say, the analysis must include the environmen-
tal impact associated with all the processes in which the products are involved,
including their use, and finally, during the waste management and treatment of
the generated residues.
Finally, the environmental load associated with the recycle processes and streams
is assumed to be included in the LCA assessment carried out by the supplier. Thus,
the model (Fig. 2.20) cuts stream Pw to generate streams wr and Pr . These new
streams include the inputs and the emissions for the recycling process and the
inputs and emissions for Supplier 1, respectively. In this case, the environmental
load associated with stream Pw is considered to be zero.
In the third phase of the analysis, the environmental loads are translated into
environmental impact indexes. This part of the analysis will be further discussed
in the case study section.
2.4.2
Financial Concerns
The importance of incorporating financial concerns in the operation of chemical

SCs has been already highlighted in the previous chapter. Thus, we next concen-
trate on the description of the financial module developed to incorporate such con-
cerns into the multiagent system.
2.4.2.1 Implementation: Financial Module

In a similar spirit as the environmental module, the financial module allows the
incorporation of financial aspects within the multiagent framework. The module
implements the mathematical formulation presented in the previous chapter to
address the short-term management of cash. The model accounts for the maxi-
mization of the change in equity of the company, rather than the corporate value
of the firm. This has been motivated by the fact that the latter metric is more suit-
able for assessing long-term decisions whereas the former performance measure
is more appropriate for a medium-term time horizon, which turns out to be the
case of the multiagent system.
The connection with the multiagent simulator is made through the payments of
raw materials, production, and transportation utilities (i.e., dates and sizes of pur-
chases of raw materials and utilities) and the sales of products carried out by the
agents in the system. As such, this formulation optimizes the cash flows associ-
ated with the SC operation and also provides a financial key performance indicator
(KPI) to assist the central manager in finding the best planning decisions in terms
of their ability to increase the shareholder’s value (SHV) of the firm. Thus, the
cash management formulation allows payments to providers, short-term borrow-
ing, pledging decisions, and the buying/selling of securities to be scheduled in line
with production tasks. The set of constraints representing balances of cash, debt,
securities, and so on, which were described in the previous chapter, is then applied
to accommodate the aforementioned issues.
2.5
Multiagent System: Integrated Framework
In this section, the overall solution strategy for SCM, which is based on the multi-
agent system previously described is presented. The focus is on incorporating the
SC dynamics and also on enlarging the scope of the analysis. The main aim is to
achieve a realistic system for holistically optimizing the SC operation. Thus, this
framework must be able to reproduce in a realistic manner the SC operation while
also considering a variety of objectives given by the firm’s interests.
To achieve this goal, the multiagent system is taken as a basis for further de-
velopments. The first step is to include objectives other than profit or cost. Thus,
the environmental and financial modules previously described are added to the
multiagent framework. Inserting these novel modules into the multiagent system
provides a way to further explore the necessary tradeoffs upon which the decision-
making procedure should be based. However, the lack of optimization skills, which
is the main disadvantage of this sort of approach, still represents a major shortcom-
ing. Thus, to go beyond the mere modeling of a generic SC, a hybrid simulation-
based optimization solving strategy is incorporated. In PSE literature, simulation-
based optimization approaches to SCM have received scant attention and require
further study. However, some authors have made highly regarded contributions to
the field [23–26].
Thus, the logical rules implemented in the agents to drive their operation and
response against the arriving events are parameterized and an optimization algo-
rithm is then invoked to act over these parameters to improve an objective function
defined beforehand. Figure 2.23 shows the way in which the simulation model is
coupled with the optimization algorithm. Within this framework, the optimization
algorithm works iteratively in a loop, taking the values of the objective function,
the variables, and the constraints at each iteration and then computing new values
for the decision variables, which are further evaluated using the simulator.
Let us note that this approach involves different simulation runs over the plan-
ning horizon. Each of these simulations considers a different Monte Carlo sample
of the probability distributions that characterize the uncertain parameters ω. Each
performance measure fp (η, ω) is a function of the set of decision variables η and
2.5 Multiagent System: Integrated Framework 69
Fig. 2.23 Framework for stochastic simulation-based optimization.
Fig. 2.24 Multiobjective approach.
also of the specific realization of the uncertain parameters ω. A set of decisions

is made periodically in each of these simulations to replenish the inventories at
the different SC entities. The simulator has the outstanding advantage of being
capable of representing the real-world SC operation as fitted as desirable. The ex-
pected value of each performance measure p can then be computed by repeating
the procedure mentioned above for a sufficient number n of Monte Carlo samples
and making then the average of the performance results obtained in each scenario
Agent-based
simulator
Fig. 2.25 Implementation of the proposed framework.
(Fp (η) = E[fp (η, ω)]). The accuracy of this expected value is indeed given by the
number of simulation runs carried out.
Schematic representations of the specific implementation of the multiobjective
multiagent framework are depicted in Figs. 2.24 and 2.25. As shown, the agent-
based simulator module still acts as the core of the methodology by responding
to uncertainties through local dispatching rules, invoking local optimization mod-
ules when necessary, and solving conflicts through the exchange of messages. The
simulator has a number of agents that map each entity of a real-world SC network
and mimic its behavior. This simulator receives a set of input values (decision vari-
ables), η , and then, by emulating the system dynamics, provides valuable informa-
tion for calculating p performance measures, which have been previously defined
by the analyst. A multiobjective optimization algorithm is coupled with the multi-
agent system in order to look for the nondominated solutions of the SCM problem
being analyzed, i.e., those solutions that properly assess the tradeoff between the
different objectives.
As a result, an approximation of the Pareto set of solutions is obtained. Therefore,
the decision makers get not only one solution, but also a set of alternatives from
which they can then further explore interesting tradeoffs.
The software prototype built based upon the research previously described was
tested on a variety of SC problems, including real-world case studies. In addition,
the software design was carried out using best practices in object-oriented design,
such as UML (unified modeling language) modeling and design patterns (see Fig.
2.25). This will make the tool’s design suitable for future extensions, such as the
incorporation of new decision-making algorithms and the inclusion of new objects.
The software was implemented in C using Microsoft’s .NET framework. The real
agents are .NET web services that communicate among themselves by means of
XML (extended markup language) messages, under the SOAP protocol. Specifi-
2.6 Case Study 71
Fig. 2.26 SC network.
cally, in the case of the simulated agents, all the data required to define a SC case
are contained in an XML file whose extension is .scm.
2.6
Case Study
In this section, the main advantages of the integrated framework for SCM in the
CPI will be highlighted through a case study. A SC network consisting of 12 inter-
connected entities i, the structure of which is depicted in Fig. 2.26, is used in this
section to illustrate the capabilities of the multiagent system. The SC under study
comprises three plants (F 1, F 2, and F 3), three distribution centers (D1, D2, and
D3), five retailers (R1 to R5), and one supplier S.
The SC network operates as a pull system (make-to-order system) with a central-
ized view and a complete degree of information sharing. A set of inventory control
policies is implemented at the nodes of the network. Specifically, a periodic revision
strategy is applied at the distribution centers whereas a continuous one is used at
the retailers. The former strategy implies that for every τij time unit, the inventory
position Iij for product j at node i is checked.
If Iij is below the reorder point sij , a replenishment quantity uij = Sij − Iij is
ordered to raise the stock level to Sij . If the position is above sij , nothing is done
until the next review. In the latter strategy, the revision is made continuously (i.e.,
0 τij ∞; τij = constant). In both strategies, sij is known as the reorder point
and Sij is the storage capacity of each node i for each product j .
The demand is assumed to be uncertain and it is modeled as a set of scenarios
with given probability of occurrence. Each scenario comprises a set of events dis-
tributed over the time horizon of the study. Each of these events has an associated
amount of material and time of occurrence. Moreover, the amounts of materials
are assumed to be normally distributed around a mean value whereas the interar-
rival time intervals are considered to be uniformly distributed. Table 2.1 shows the
values for the customer demand.
Table 2.2 shows for each product the initial values of the different inventory para-
meters applied at each SC node. τij is the review period for the distribution centers,
sij is the reorder point, Inv0ij is the inventory level at the starting point of the sim-
ulation, and Sij is the storage capacity at each node. The upper and lower bounds
of the decision variables have been set taking as reference the initial values with
which the SC operates. For the review period, we assume that 0 τij 1.2 · τij0 . In
this expression, τij0 represents the initial value of each parameter, which is shown
in Table 2.2. For the rest of the parameters, a similar procedure has been applied:
0 sij 1.2 · sij0 and 0 Sij 1.2 · Sij0 .
The unit production and inventory costs (vpcij and vicij t ) are given in Tables
2.3 and 2.4. The production time is equal to 1 min per unit of product j for all
the products. Furthermore, for the sake of simplicity, the transportation costs and
times are assumed to be the same for all the products and nodes. The variable
transportation cost per unit, vtcij , is equal to 0.10 m.u./u. The transportation time
is considered to be uncertain and it is sampled from a normal distribution with a
mean of 1 h and variance of 2 h2 . Table 2.5 shows the product prices (priceij ).
The main activity of this SC is the electroplating of three different metallic ob-
jects (A, B, and C). These electroplating plants have the following relevant inputs:
electricity, water, the raw objects to be treated, and chromium and nickel salts for
the electrochemical treatment. The SC has also another plant that produces nickel
Table 2.1 Demand data at each sales region.
R1 R2 R3
A B C A B C A B C
Mean (u) 20 80 45 12 89 47 80 24 120

Variance (u2 ) 5 20 5 3 32 12 4 5 50
λij 111 111 123 123 111 124 105 150 140
R4 R5
A B C A B C
Mean (u) 35 80 84 90 110 80

Variance (u2 ) 15 15 15 20 23 25
λij 140 140 150 111 150 140
2.6 Case Study 73
Table 2.2 Inventory parameters before the optimization process

([τij ] = h, [sij ] = u, [Sij ] = 103 u, [Inv0ij ] = 103 u).
Supplier
S
τij sij Sij Inv0ij
P 7 800 8.0 6.0
Manufacturing plants
F1 F2 F3
τij sij Sij Inv0ij τij sij Sij Inv0ij τij sij Sij Inv0ij
P 7 800 8.0 6.0 7 750 8.0 6.0 7 750 8.0 6.0

A 7 800 8.0 6.0 7 750 8.0 6.0 7 750 8.0 6.0
B 7 700 8.0 7.0 7 850 8.0 7.0 7 750 8.0 6.0
C 7 850 8.0 3.5 7 800 8.0 6.5 7 750 8.0 6.0
D1 D2 D3
τij sij Sij Inv0ij τij sij Sij Inv0ij τij sij Sij Inv0ij
A 7 800 7.0 6.0 7 750 7.0 6.0 7 750 7.0 6.0

B 7 700 7.0 7.0 7 850 7.0 7.0 7 750 7.0 6.0
C 7 850 7.0 3.5 7 800 7.0 6.5 7 750 7.0 6.0
Retailers
R1 R2 R3
sij Sij Inv0ij sij Sij Inv0ij sij Sij Inv0ij
A 20 0.50 0.20 50 0.50 0.23 40 0.50 0.26

B 15 0.50 0.30 36 0.50 0.25 50 0.50 0.29
C 20 0.50 0.25 40 0.50 0.28 35 0.50 0.20
Retailers
R4 R5
sij Sij Inv0ij sij Sij Inv0ij
A 50 0.50 0.24 40 0.50 0.20

B 35 0.50 0.25 50 0.50 0.30
Table 2.3 Unit production costs at each plant, vpcij (m.u./u).
F1 F2 F3
A 1.0 3.0 2.0

B 2.0 2.0 3.0
C 3.0 1.0 3.0
Table 2.4 Variable inventory costs at each node, per product
and simulation step, vicij t (103 m.u./u/t).
Supplier Manufacturing plants

S F1 F2 F3
P 0.10 0.15 0.15 1.00

A – 0.15 0.15 1.00
B – 0.20 0.20 1.00
C – 1.00 0.40 1.00
Distribution centers Retailers

D1 D2 D3 R1 R2 R3 R4 R5
A 0.10 0.15 0.15 1.00 1.50 1.00 0.15 0.15

B 0.20 0.20 0.20 1.00 0.50 1.00 0.15 0.20
C 0.30 1.00 0.40 1.00 0.50 1.00 0.15 0.20
Table 2.5 Product prices, priceij (m.u./u).
R1 R2 R3 R4 R5
A 800 110 95 100 120

B 900 210 200 202 220
C 1500 305 290 309 300
(II) chloride (NiCl2 ). This plant also provides nickel (II) sulfate (NiSO4 ), which is
a byproduct that contributes to increase the SC earnings. Moreover, the plant pro-
ducing NiCl2 consumes coal, fuel, and nickel ore.
In the phase of Goal and Scope of the LCA study, the boundaries of the system
are set, as it is depicted in Fig. 2.26. This work assumes that chromium salts, tubes,
and water are elemental flows (i.e., they come directly from the environment and
do not exhibit any associated burden other than the resource depletion itself). On
the other hand, electricity comes from a generation process that uses mainly coal
and fuel as raw materials. The functional unit has been referred to product B.
The data required to compute the environmental indexes associated with the SC
operation are the consumption of utilities and raw materials and the emissions of
each process. This analysis covers all the main production/distribution tasks car-
ried out in the SC, including the extraction of raw materials and the manufacturing
process itself. Nevertheless, for the sake of simplicity, the use and disposal phases
have not been considered in this example although the environmental module is
indeed able to deal with them. This specific study can thus be seen as a “cradle to
gate” analysis instead of a “cradle to grave” one.
The data corresponding to the functional unit, inputs of materials and energy,
and outputs of products, subproducts and wastes from the system under study can
be found in Table 2.6. The first column of this table indicates whether the inputs
2.6 Case Study 75
Table 2.6 Aggregated data for all the electroplating plants, the
distribution centers, and the retailers.
Elemental flow? Inputs Emissions Products

Name Value Unit Name Value Unit Name Value Unit
No NiCl2 125 kg to the air: A 100 u

Yes Cr salts 200 kg CO 0.0025 kg B 50 u
Yes Brass 125 kg CO2 0.0025 kg C 25 u
Yes Water 2000 kg SOx 0.0025 kg
No Electricity 230 MJ to the water:
Cr 0.0025 kg
Ni 0.0250 kg
Table 2.7 Data for the NiCl2 production process.

Yes Coal 20 kg to the air: NiCl2 120 kg

Yes Fuel 40 kg CO 0.0250 kg NiSO4 100 kg
Yes Ni ore 50 kg CO2 0.0025 kg
NOx 0.0200 kg
to the water:
Ni 0.0250 kg
Table 2.8 Data for the process supplying electricity.

Yes Coal 45 kg to the air: Electricity 4500 MJ

Yes Fuel 50 kg CO 0.0250 kg
are elemental (“yes”) or not (“not”). There are two flows that are not elemental and
must then be computed: NiCl2 and electricity. Thus, additional data for these two
inputs are required (Tables 2.7 and 2.8). Let us note that in Table 2.6, the data are
referred to the functional unit, 50 u of product B, whereas in Tables 2.7 and 2.8,
the data are expressed according to an arbitrary calculation basis. The calculation
basis for data concerning NiCl2 production is 120 kg whereas for the electricity
production it is 4500 MJ.
The aforementioned data must be loaded in the system database. This informa-
tion will then be retrieved by the LCA module, which will start with the inventory
calculations. The module performs material and energy balances taking into ac-
count the structure of the network. An allocation method is applied to allocate the
environmental loads of the two output streams (NiCl2 and NiSO4 ) that are gener-
Table 2.9 Allocation to the NiCl2 .
Products Elemental flow? Inputs Emissions

NiCl2 120 kg Yes Coal 10.9091 kg to the air:

Yes Fuel 21.8182 kg CO 0.013636 kg
Yes Ni ore 27.2727 kg CO2 0.001364 kg
NOx 0.010909 kg
to the water:
Ni 0.013636 kg
Table 2.10 Material balance over 50 u of product B.
Inputs Emissions Products

Brass items 125 kg to the air: A 100 u

coal 13.66364 kg CO 0.017982 kg B 50 u
Cr salts 200 kg CO2 0.003920 kg C 25 u
Fuel 25.28283 kg NOx 0.062475 kg
Ni ore 28.40909 kg SOx 0.002500 kg
water 2000 kg to the water:
Cr 0.002500 kg
Ni 0.039205 kg
ated in the production of NiCl2 . In this case the mass allocation method [27] has
been applied. This method distributes the burdens (inputs and emissions) of the
manufacturing process between the two products. To do so, the loads are multi-
plied by a factor that, in this case, is equal to 60/(60 + 50) = 0.545 for NiCl2 and
50/(50 + 60) = 0.455 for NiSO4 (see Table 2.7). The results of this allocation pro-
cedure can be seen in Table 2.9. Let us note that the production of electricity does
not require any allocation method as the electricity is the only product generated
by the process. The results of the mass balance associated with the main product
are shown in Table 2.10.
The avoided burden method is next applied to allocate the environmental burdens
to the functional unit (product B). This method requires data concerning the stan-
dard manufacturing processes of products A and C. The results of this phase are
shown in Table 2.11. This table, which is known as the inventory table, provides
the environmental burdens expressed as the amount of resources consumed and
wastes released per functional unit.
In the Impact Assessment phase, which is next applied, the environmental im-
pact indexes are computed from the inventory table using the potential factors pro-
vided in [15]. In the multiagent framework, the environmental burdens are not
static, instead, they change dynamically according to the operation of the SC. To
2.6 Case Study 77
Table 2.11 Inventory table.

Yes Brass 125 kg to the air: B 50 u

Yes Coal 125 kg CO 0.002982 kg
Yes Cr salts 200 kg CO2 0.002420 kg
Yes Fuel 125 kg NOx 0.042475 kg
Yes Ni ore 125 kg SOx 0.000500 kg
to the water:
Cr 0.002300 kg
Ni 0.037205 kg
Table 2.12 Initial portfolio of marketable securities 10−3 (StMS

∗ (m.u.)).
t∗ = 1 2 3 4 5 6 7 8 9 10 11 12
150 150 70 50 40 75 60 100 100 25 60 80
compute the inputs and emissions dynamically, the following equation can be ap-
plied:
me = ϕeP · t P + ϕeI · t I + ϕeT · t T , (2.2)
where me is the mass (in kg) of environmental issue e (e.g., resource, emission).
Superscripts P, I, and T stand for the three main activities in a SC, i.e., production,
storing, and transportation, respectively. ϕeP is the contribution to the environmen-
tal burden e by the activity P expressed in kg per time unit of activity P. Finally, t P
accounts for the total time during the simulation in which activity P is active, ex-
pressed in time units. This equation allows us to compute the environmental loads
associated with the SC as a function of the production/distribution tasks carried
out in its nodes.
With regard to the financial matters, it is assumed that the firm has an initial
portfolio of marketable securities given by Table 2.12 at the beginning of the time
horizon. Twelve financial periods with a length of 1 week are considered. The initial
cash is equal to the minimum cash (5,000,000 m.u.). Under an agreement with a
bank, the firm has an open line of credit at an 8% annual interest with a maximum
debt allowed of 2,500,000 m.u., the initial debt being equal to 1,000,000 m.u. The
change in assets is computed considering that the raw materials and final products
kept as final inventories at the end of the time horizon can be sold at a 50% of their
market value. On the other hand, the value of the final inventories of intermediate
products is equal to zero.
Three external suppliers are considered: the first one provides raw materials, the
second one production utilities, and the last one transportation services. Liabilities
incurred with the raw materials supplier and the supplier of production utilities
have to be repaid within 4 weeks according to the terms of the credit (2%-7 days,
net-28 days for the raw material supplier, and net-28 days for the second one). The
payments associated with the transportation services cannot be stretched and must
be fulfilled within the same period of time in which the purchase incidence takes
place. The technical coefficients for the marketable securities purchased and sold
by the firm are computed assuming an annual interest rate of 2.8% for purchases
and 3.5% for sales. At the end of the time horizon (week 12) 500,000 m.u. are
withdrawn from the company as dividends. There are also outflows of cash equal
to 2.25, 1, 0.75, and 1.25 millions of m.u. in periods of 4, 6, 8, and 12 due to wages,
rents, changes in fixed assets, and the repayment of the long-term debt, which
remains constant during the whole time horizon. Finally, the receivables on sales
executed in any period are paid with a 28-day delay and may be pledged at an 80%
of their face value.
The problem then is to find the values of the parameters associated with the
inventory control policies implemented at the nodes of the SC that maximize the
financial performance and minimize the environmental impact.
2.6.1
Objectives
As stated before, two objectives are considered in this study: economic performance
and environmental impact. Then, the multiobjective (MO) problem can be stated
as follows:

−F1 (η) = −E[E(η, ω)]
min U = , (2.3)
η F2 (η) = E[EnvIndex(η, ω)]
where U is the set of objective functions. In Eq. (2.3), Fp (η) is a function of the in-
ventory parameters η that is evaluated by using the results of multiple Monte Carlo
samplings with embedded discrete event simulations. The economic performance
(i.e., change in equity) achieved by the SC in each simulation run is computed by
the financial module whereas the environmental impact index is determined by
the environmental module. These values are then used to compute the expected
values of the objectives by making the average of the performances achieved in
each scenario.
The environmental impact indexes to be minimized -EnvIndex(η, ω)-, can be
based on either the second or third phase of the LCA methodology. In this case
study, the environmental impact calculation is based on the “problem-oriented”
approach to impact assessment proposed by [15]. This work applies a set of coef-
ficients associated with material or energy flows (continuous variables) that repre-
sent the relative contribution of each burden to an impact factor.
Specifically, in this case, the global warming potential (GWP) has been chosen
as the environmental objective function to be minimized. To compute the GWP
factors associated with different greenhouse gases, the GWP of CO2 is taken as
reference. The GWP of CO2 is thus defined to be unity. Our model assumes that
the environmental burdens and impact functions are linear (i.e., they are directly
2.6 Case Study 79
proportional to the functional unit(s) and there is no synergistic or antagonistic

effect between them). The necessary steps to calculate the environmental impact
index are as follows.
• Step I: Perform the material and energy balances associated
with all the SC entities, following the principles of the ISO
14040 series. This step provides the inventory table with all
the environmental loads: input flows coming from the
environment and output flows. The outputs, in turn, can be
classified as outputs economically valuable (products and
byproducts) and emissions to the environment.
• Step II: Use the following equation to translate the SC
emissions into an indicator of the Greenhouse Effect (GHE)
enhancement:

GHE = GWPe · me , (2.4)
e
where GHE is the GHE enhancement indicator expressed in
kg of CO2 , and GWPe is the GWP corresponding to
emission e according to the model of Heijungs [15]. Finally,
me is the mass (in kg) of emission e released to the air.
The choice of the environmental objectives to be optimized depends on the goal

and scope of the study. In this work, an environmental impact has been selected
as the objective to be minimized. Nevertheless, as was mentioned before, the op-
timization strategy could be applied either at the inventory or impact assessment
levels. The environmental objectives would therefore be defined accordingly, as ei-
ther burdens or impacts.
2.6.2
Results
The multiagent system is applied to solve the problem previously outlined. The
financial and environmental modules are used to assess the economic and envi-
ronmental performances of the SC, respectively. A multiobjective GA (MOGA) is
utilized to seek the best values of the decision variables taking into account the
objectives computed by the aforementioned modules. Moreover, the performance
of the MOGA can be improved by applying a tuning strategy for determining the
optimal parameters of the GA, namely population size, probabilities of crossover,
and mutation.
Each simulation run entails a planning horizon of 3 months. The GA handles 96
decision variables that are associated with the inventory replenishment strategies
applied in the SC entities: the inventory parameters sij , Sij , and τij at the factories
and distribution centers (66 variables), and the parameters sij and Sij at the retail-
ers (30 variables). Real-valued encoding for the variables and maximum number of
generations as termination criterion have been used.
Figures 2.27, 2.28, and 2.29 show some snapshots of the graphical user interface
(GUI) of the modules developed to perform MO optimization and LCA calcula-
tions. Table 2.13 shows the settings used for the MOGA. In Fig. 2.30, two approx-
imations of the Pareto front showing the tradeoff between the economic perfor-
mance index (change in equity) and the GHE index are depicted. The curve on the
left is the initial approximation of the Pareto front, before starting the optimization
process. The curve on the right is the approximation of the Pareto front obtained
at the end of the optimization process when the maximum number of generations
is reached. Figures 2.31 and 2.32 show the evolution of the inventory level during
the simulation run at the distribution center D1 for the two extreme points of the
Pareto front, as it appears in the GUI of the multiagent system.
The points located on the right-hand side of the Pareto front correspond to so-
lutions that represent a very responsive SC. The operation of such SCs is charac-
terized by low inventory levels. This operation policy yields bigger profits, since in
Fig. 2.27 GUI for the MO optimization: Settings and numerical results.
Fig. 2.28 GUI for the MO optimization: Graphical results.

2.6 Case Study 81
Fig. 2.29 GUI for the LCA module.
Fig. 2.30 Pareto front approximation between the expected

change in equity and the greenhouse effect index.
Table 2.13 Settings for the GA-based strategy.
Algorithm used NSGAII

Number of decision variables of the outer loop 96
Number of timelines, n 90
Population size, N 20
Maximum number of generations, MaxGen 200
Crossover probability, Px 1.0
Mutation probability, Pm 0.1
Fig. 2.31 Inventory profiles of product B at distribution center

D1: On the left-hand side of the Pareto front.
Fig. 2.32 Inventory profiles of product B at distribution center

D1: On the right-hand side of the Pareto front.
this example high inventory costs have been assumed. On the other hand, it leads
to high environmental impacts, which are mainly due to the very busy transporta-
tion tasks required by these solutions, for which very high contributions to the
environmental index have been defined. In other words, these solutions achieve a
superior economic performance at the expense of higher GHE levels. The points
on the left-hand side of the front, in contrast, have lower values of the profit but
also lower values of the GHE index.
The CPU time required for the algorithm execution ranges from 20 min to sev-
eral hours in the cases tested. The simulation runs were carried out on an AMDK6
2.7 Concluding Remarks and Future Work 83
computer, 2.16 GHz, 512 MB. Let us note that although it is possible to improve the
quality of the solutions by increasing the number of iterations, the proposed strat-
egy is able to provide feasible and acceptable solutions in reasonable CPU times.
This CPU time depends on the number of simulation runs to be made (i.e., num-
ber of Monte Carlo samples to be explored) during each algorithm execution, and
also on the tuning parameters of the GA-based strategy.
2.7
Concluding Remarks and Future Work
This chapter has described the research carried out in the last few years, which
has mainly been devoted to developing a new modeling technique for SCM based
on a simulation model that uses software agents as building blocks. The resulting
agent-based system is a discrete event simulator that is able to use a number of
different tools such as if-then rules and mathematical programming algorithms.
An appropriate approximate strategy for tackling SCM problems has been devel-
oped from the viewpoints of analysis, improvement, and optimization. This strat-
egy relies on the use of metaheuristics over the multiagent simulator. The resulting
tool is flexible and extensible and allows new properties in the objects defined and
new objects to be defined. Moreover, this framework includes environmental and
financial considerations that enlarge the scope of the SC analysis and assist in as-
sessing the tradeoff between diverse environmental and economic concerns.
The application of dynamic simulation and optimization to the SCM problem
constitutes a wide-ranging, appealing field, in which there is much work to be
done. Integration between levels continues to be an unresolved matter, and there
are many areas in which there are opportunities for improvement, such as data
management and connecting different tools. The research carried out so far can
thus be taken as a basis for future enhancements. We propose that future research
be undertaken on the following topics:
• Simulation-based optimization methodologies: further
improvements are required in the form of better methods,
metamodels, and filters to accelerate the algorithm’s
convergence.
• Optimization under uncertainty: further work is needed to
develop new strategies to deal with the curse of
dimensionality featuring the stochastic approaches.
• MO optimization: the general framework developed so far
could be extended to incorporate others KPIs (social impacts,
safety impacts, etc.) as objectives to be optimized within the
decision-making process.
• Interaction with upper and lower decision levels: there is still
much work to be done to effectively integrate the present
approach with the higher business level and the lower
production scheduling and process control levels.
• Incorporation of negotiation abilities: this topic concerns the

consideration of the negotiation process between customers
and suppliers that takes place in SCs.
Acknowledgments
The authors wish to acknowledge support of this research work from the European
Community (Contract Nos. GIRD-CT-2001-00466 and MRTN-CT-2004-512233), the
CICyT-MEC (project No. DPI2003-0856), and the CIRIT-Generalitat de Catalunya
(project No. I-898).
References
1 Méndez, C. A., Cerdá, J., Gross- Electronic Commerce and Managing the
mann, I. E., Harjunkoski, I., Fahl, Internet-Enabled Supply Chain, 1999.
M., Comput. Chem. Eng. 30(6–7) 9 Julka, N., Karimi, I., Srinivasan, R.,
(2006), pp. 913–946. Comput. Chem. Eng. 26 (2002), pp.
2 Math, W. T., MATLAB 7.0 (R14), 1771–1781.
Simulink 6.0, Stateflow 6.0. User’s Man- 10 El-Halwagi, M. M.,
ual, The Math Works Inc., 2004. Manousiouthakis, V., Chem. Eng.
Sci. 45 (1990), pp. 2813–2831.
3 Law, A. M., Kelton, W. D., Simula-
tion Modeling & Analysis, 3rd edn., 11 Ciric, A. R., Jia, T., Comput. Chem.
McGraw-Hill, New York, 2000. Eng. 18 (1994), pp. 481–495.
12 Wang, Y. P., Smith, R., Chem. Eng.
4 Banks, J., Handbook of Simulation:
Sci. 49 (1994), pp. 981–1006.
Principles, Methodology, Advances, Ap-
plications, and Practice, Wiley, New 13 Linninger, A. A., Stephanopoulos,
E., Ali, S. A., Han, C., Stephanopou-
York, 1998.
los, G., Comput. Chem. Eng. (S) 19
5 Julka, N., Karimi, I., Srinivasan, R., (1995), pp. S7–S13.
Comput. Chem. Eng. 26 (2002), pp. 14 Azapagic, A., Clift, R., Comput.
1755–1769.
Chem. Eng. 10 (1999), pp. 1509–1526.
6 Goodwin, R., Keskinocak, P., 15 Guinée, J. B., Gorrée, M., Heijungs,
Murthy, S., Wu, F., Akkiraju, R., R., Huppes, G., Kleijn, R., de Kon-
Intelligent decision support for the ing, A., van Oers, L., Sleeswijk,
e-supply chain, in: Artificial Intelli- A. W., Suh, S., de Haes, H. A. U.,
gence for Electronic Commerce, AAAI de Bruijn, H., van Duin, R., Hui-
Workshop 99, 1999. jbregts, M. A. J., Handbook on Life
7 Sauter, J. A., Parunak, H. V. D., Cycle Assessment. Operational Guide to
Goic, J., ANTS in the supply Chain, the ISO Standards, Kluwer, Dordrecht,
in: Workshop on Agents for Electronic 2002.
Commerce at Agents 99, 1999. 16 Petrie, B. A. G. B. J., Romagnoli, J.,
8 Chen, Y., Peng, Y., Finin, T., Labrou, Comput. Chem. Eng. 24(2Ů7) (2000),
Y., Cost, S., Chu, B., Yao, J., Sun, R., pp. 1195–1200.
Wilhelm, B., A negotiation-based 17 Hoffmann, V. H., Hungerbuhler,
multi-agent system for supply chain K., McRae, G. J., Ind. Eng. Chem. Res.
management, in: Working Notes of 40(21) (2001), pp. 4513–4524.
the Agents ’99 Workshop on Agents for 18 Chen, H., Shonnard, D. R., Ind. Eng.
Chem. Res. 43(2) (2004), pp. 535–552.
References 85
19 Hugo, A., Pistikopoulos, E. N., En- 24 Subramanian, D., Pekny, J. F.,

vironmentally conscious planning Reklaitis, G. V., AIChE J. 47(10)
and design of supply chain networks, (2001), pp. 2226–2242.
in: Process Systems Engineering 2003, 25 Jung, J. Y., Blau, G., Pekny, J. F.,
Chen, B., Westerberg, A. W. (eds.), Reklaitis, G. V., Eversdyk, D., Com-
Elsevier, Amsterdam, 2004. put. Chem. Eng. 28 (2004), pp. 2087–
20 Puigjaner, L., Integration in sup- 2106.
ply chain management, in: Computer
26 Wan, X., Pekny, J. F., Reklaitis, G. V.,
Aided Process & Product Engineering
– CAPE, Puigjaner, L., Heyen, G.
1317–1328.
(eds.), Springer, Berlin, 2006.
21 ISO 14041, Gestión Ambiental, Análisis 27 Jensen, A. A., Hoffman, L., Iler,
del Ciclo de vida- Definición de la meta B. T. M., Schmidt, A., Christiansen,
y el alcance y análisis del inventario del K., Elkington, J., van Dijk, F., Life
ciclo de vida, 1998. Cycle Assessment (LCA). A Guide to
Approaches, Experiences and Informa-
22 ISO 14042, Gestión Ambiental, Análi-
sis del Ciclo de vida- Eva- luación del tion Sources, European Environmental
Agency, 1998.
impacto del ciclo de vida, 2000.
23 Subramanian, D., Pekny, J. F.,
Reklaitis, G. V., Comput. Chem. Eng.
24 (2000), pp. 1005–1011.
87
3
Environmentally Conscious Supply Chain Management
Metin Türkay
3.1
Introduction
A typical supply chain system consists of customers, retailers, distributors, ware-

houses, production facilities, and suppliers that can be represented as nodes in a
highly integrated network, as shown in Fig. 3.1. Traditionally, the main objective
of such supply chain systems has been to satisfy the demand by the customer for
maximizing the financial gain by the operation of the supply chain. The financial
concerns primarily include: the cost of raw material purchasing from the supplies,
the production cost at the production/manufacturing centers, the inventory and
material handling costs at the distribution centers, the cost of customer service at
the retailers, the revenue generated by the customers that covers all of the costs
and profits of the nodes in the supply chain network, and the cost of transportation
incurred by the movement of material/goods throughout the supply chain system
as shown in Fig. 3.1.
There has been a growing concern for the environment recently. The main rea-
sons for this are the effects that industrial and transportation activities are hav-
ing on the environment. The CO2 emission, which is the main contributor in the
greenhouse effect, can be categorized into four according to the end use by the sec-
tor: transportation, industrial, residential, and commercial. According to the An-
nual Energy Review 2005 report by the Energy Information Agency in the United
States, the emission by the transportation and industrial sectors are the sectors
with the largest emission of CO2 [1].
The contribution of the transportation and industrial sectors to CO2 emissions
is significant as seen in Table 3.1. The supply chain systems integrating different
nodes with transportation links play a major role in environmental effects includ-
ing the CO2 emissions. The environmental effects from the supply chain systems
include: the quality of raw materials from the supplies, the emissions during pro-
duction at the production/manufacturing centers, the inventory and material han-
dling effects at the distribution centers, the effects due to storage and material han-
dling systems at the distribution centers, the effects of packaging at the retailers,
ISBN: 978-3-527-31693-9
88 3 Environmentally Conscious Supply Chain Management
Fig. 3.1 Schematic representation of supply chain with

financial flows and environmental effects.
Table 3.1 CO2 emissions by different sectors in US in 2004 [1].
Sector CO2 emission (billion tons) % in total
Transportation 1.9 32.76

Industrial 1.7 29.31
Residential 1.2 20.69
Commercial 1.0 17.24
Total 5.8 100.00
and the waste generated after the product is consumed by the customers as shown
in Fig. 3.1.
The quality of raw materials has significant effect on the environmental perfor-
mance of the supply chain systems. Raw materials contain impurities that may be
harmful to the environment. For example, it is possible to generate electricity or
steam using different fuels such as natural gas, oils with different compositions,
and coal with different compositions in an energy supply chain operating with
conventional fossil fuels. Different fuels may contain different sulfur contents re-
sulting in varying affects on the environment. The production systems usually gen-
erate waste in gaseous, liquid or solid form. Although some of this waste is treated
in some cases, majority of the waste is released to the environment. The storage
and material handling systems at the distribution centers generate environmen-
tally harmful substances during their operation. For example, a food storage sys-
3.1 Introduction 89
tem needs to operate at low temperatures requiring energy and cooling agents. The
retailers usually sell the products in packages that are attractive to the customers.
Majority of the packaging material usually end up in nature without properly recy-
cling or utilizing waste disposal facilities. The customer generates waste after con-
suming the products. The waste from the product is usually harmful to the envi-
ronment. For example, when the life of electronic equipment is over, it contains sig-
nificant quantities of metals and other environmentally harmful substances. The
transportation system also generates environmentally harmful substances through
emissions from the vehicles in the system.
Traditionally, supply chain systems have been analyzed with the objective of fi-
nancial considerations such as the minimization of the cost or the maximization
of the profit. However, an efficient and responsive supply chain system needs to in-
clude environmental considerations for sustainability [2]. The environmental per-
formance of supply chain systems can be achieved by focusing on better manage-
ment of resources available in the supply chain [3].
We examine the environmental issues in supply chain systems in three cate-
gories:
1. product centric approaches (closed-loop supply chains);
2. production system centric approaches (environmentally
conscious production);
3. transportation system centric approaches (sustainable
transportation).
The product centric approaches focus on the design of the product for minimiz-
ing the use of environmentally harmful materials in the product and the recovery
and reuse of the product after it has been consumed by the end user [4]. The objec-
tive in the product centric approaches is to eliminate the product becoming a waste
after the product has completed its life time. These approaches include closed-loop
supply chains where the product is recovered for reuse/recycle and reverse logistics
that includes the planning of the logistics infrastructure of the products.
The production system centric approaches consider the selection of raw mate-
rials and the design of the production systems for minimizing the environmental
impact [5]. The objective of the production system centric approaches is to design
the production system so that the production system is flexible enough to elimi-
nate or reduce the generation of waste. One of the mechanisms is the use of dif-
ferent raw materials. Other mechanisms include changing the configuration of the
equipment or the operating conditions of the process system to reduce generation
of waste.
Transportation centric approaches consider the use of different transportation
systems that would reduce the environmental effects. For example, whenever pos-
sible using rail or sea transportation could reduce the emission of greenhouse
gases (GHG).
3.2
Closed-Loop Supply Chains
The closed-loop supply chains aim to eliminate the product or some parts of the
product becoming a waste after the product has completed its life time. The life of
the product ends after the product loses its functions. Due to accelerating speed
of innovation in certain industries, the life cycle of many of modern products are
shortened. When a product completes its life cycle, it is necessary to process the
product in different facilities and prevent the product becoming a complete waste.
Therefore, it is necessary to extend the classical supply chain to include product
recovery and processing for environmentally conscious supply chain management.
A typical closed-loop supply chain system is shown in Fig. 3.2.
The closed-loop supply chains include the waste collection, product remanufac-
ture, product disassembly, parts refurbishing and waste disposal nodes in addition
to the suppliers, manufacturers, distribution centers, and retailers [6, 7]. The prod-
ucts are collected in the collection sites and inspected for their conditions. If a
product can be reused by changing a small number of parts, the product is sent to
the remanufacturing node where new parts are replaced with the old ones. These
products are sent to distribution centers and sold as remanufactured products in
the retail outlets. The parts that are replaced by the new ones are sent to refurbish-
ing sites for further inspection. If a product requires changing/repairing of the
major parts, then they are disassembled to different parts. The parts that cannot be
repaired or reused are sent to the waste disposal facilities. The remaining parts are
then inspected at the refurbishing site for repair or disposal. The parts that can be
reused in the product (new or remanufactured) are stored as part inventory.
The closed-loop supply chains are instrumental in minimizing the effects of the
products that complete its life cycle. In addition to environmental gains, it is also
possible to generate some financial gains from closed-loop supply chains. The sales
of remanufactured or refurbished products are becoming a regular practice in elec-
tronic and computer industries.
Fig. 3.2 Schematic representation of a typical closed-loop supply chain.

3.3
Environmentally Conscious Production
The aim of the environmentally conscious production is to design and operate the
production system to eliminate or reduce generation of waste. This can be achieved
by using different raw materials when the production system is flexible, changing
the configuration of the equipment or the operating conditions of the process sys-
tem to reduce generation of waste. Since energy has strong effects on modern life,
the environmentally conscious production approaches can be illustrated on the en-
ergy supply chains.
Energy supply chain consists of raw materials, production facilities, and de-
mands for end products, i.e., energy. More than one-third of all primary energy
consumption goes into producing and delivering electricity and most of the world’s
primary energy consumption comes from fossil fuels, such as coal, natural gas, and
oil according to Energy Information Administration (EIA)’s Report [8]. Burning fos-
sil fuels release emissions that are harmful to the environment. These emissions
can be classified into two main groups: SOx and CO2 equivalent emissions. SOx
equivalent emissions are the particles that are quantified strictly and limited with
certain regulations. The case with CO2 equivalent emissions is different; these are
gases that have effect on global warming, i.e., greenhouse gases (GHG). The Kyoto
Protocol sets obligatory limits on the emission of GHGs by the industrialized coun-
tries during the period 2008–2012 [9]. The limits apply to an aggregate of emissions
of six gases covered in the protocol in CO2 equivalent units.
According to Energy Information Administration (EIA)’s Report, more than 80
percent of the human-originated greenhouse gas emissions are energy related [1].
Thus, electricity production and consumption are likely to be a major focus in
meeting Kyoto Targets. The electricity production systems and researchers are in
search for recipe solutions for reducing emissions without sacrificing from the
amount of energy production and low prices of production.
Since production and management of energy are very important, there has been
extensive research on planning and decision making in the energy supply chains.
Using the existing infrastructure more effectively can create benefit quickly in the
quest of low emissions and low prices. The primary performance criteria in energy
supply chains are the minimization of the costs and minimization of the emis-
sion of GHG to the environments that are conflicting with each other. The multi-
objective optimization problem can be represented using generalized disjunctive
programming framework [10, 11] as follows:

n
p
min zCOST = cj xj + c1k
j =1 k=1

n
p
min zGHG = dj xj + c2k
j =1 k=1
s.t.
Fig. 3.3 Flowsheet of an energy production system.
g(x) = 0
 
Yk

n 
 a x = bi 
 j =1 ij j 
 
k∈Dk  c1k = fk 
c2k = rk
(Y ) = True
xj , c1k , c2k 0, Yk ∈ {True, False} ∀j = 1, . . . , n, ∀k = 1, . . . , p. (3.1)
The multiobjective optimization problem given in Eq. (3.1) includes discrete and
continuous decision variables. The discrete decisions are modeled using Boolean
variables, Yk and the relationships between the Boolean variables and the con-
straints are modeled using disjunctions. The disjunctions model the operation of
the nodes in the energy production system as shown in Fig. 3.3. The detailed mod-
els for the nodes in the energy production systems using fossil fuels are given in
the following sections. The models include three courses of action simultaneously
to seek efficient solutions that minimize the total cost (zCOST ) and minimize the
emission of GHG gases (zGHG ):
1. using biofuel;
2. using Carbon Capture and Sequestering systems that
involves changes in the process topology;
3. determining operating conditions including turning on/off
the equipment.
3.3.1
Process Models for Energy Production Systems
The models for the energy production systems are presented with the objective
of assessing the synergy analysis that is conducted in the previous chapter with
systems that resemble real systems. A typical energy production system consists of
storage tanks to inventory raw materials, boilers that convert fuel into steam at high
pressures, turbines that expand higher pressure steam to lower pressure steam
and convert the mechanical energy released during this expansion into electricity
and mixing equipment for mixing compatible materials originating from different
sources in the system. Energy systems utilize fuel, air, and other materials to gen-
erate electricity and various grades of steam: high pressure (HP), medium pressure
(MP), and low pressure (LP) steam. The modeling of energy systems has been ad-
dressed in the literature [12, 13]. The mathematical models for common equipment
in energy production systems are summarized in the following sections.
3.3.1.1 Boiler Models and Use of Biofuels

The generation of HP steam is accomplished in the boilers by burning fuel, which
results in the emission of harmful substances such as GHG or SOx . The boilers
can be supplied with different fuels as raw material with minimal adjustments in
the operating conditions. This requires the selection of economically and/or envi-
ronmentally attractive fuel among the available alternatives. The alternatives may
be sulfurless oil, heavy oil, etc. which differ in calorie content, harmful emissions,
and cost. When environmental constraints appear, companies try to find new al-
ternatives for producing energy with minimum emissions. Biodiesel is a nontoxic
alternative fuel made from renewable fats and vegetable oils with a performance
a little lower than the petroleum-based diesel. Free of sulfur and aromatics, it can
be used in engines and boilers with few or no modifications. A biodiesel blend is
pure biodiesel blended with petrodiesel. Blends up to 20% biodiesel are compatible
with all known oil tanks and boilers. The compatibility of higher biodiesel blends
depends on the properties of the materials of the tanks, pumps and fuel lines. The
purchasing cost of biodiesel is a little higher than petrodiesel and holding cost is
higher because of its material properties [14]. The biodiesel can be mixed with only
one type of the fuel and the other fuels cannot be mixed to each other. The boiler
models consist of the following equations.
cck
Xij kfuel lcon t = XHFij kfuel t , (3.2)
ηij

Xij kHP lgen t = XHFij kt , (3.3)
k∈Fuel

Xij klcon t 0.2 Xij klcon t , (3.4)
k∈Biofuel k∈Fuel
XHFij kt M × YFUij kt , (3.5)

YFUij kt 1, (3.6)
k∈Fuel
k ∈Biofuel
/
Xij kMP lcon t = aij kMP Xij kHP lgen t , (3.7)
Xij kEL lcon t = aij kEL Xij kHP lgen t , (3.8)

Xij kSOx lgen t = sSOx kfuel Xij kfuel lcon t , (3.9)
k∈fuel
XjLkHP lgen i × Yij t Xij kHP lgen t Xij

U
kHP lgen × Yij t . (3.10)
The variable representing the HP steam production in a boiler (Xij kHP lgen t ) is dis-
aggregated into variables (XHFij kt ) for the fuel type it has been produced. Equation
(3.2) states that the HP steam production from a fuel is proportional to the calorific
value of fuel, cck , and the boiler efficiency, (1/ηij ). Equation (3.3) models that the
amount of HP steam produced in a boiler is equal to the sum of HP produced from
different fuels in that boiler. Equation (3.4) restricts the amount of biodiesel usage
to maximum 20% of the blend used in that period. If a particular type of fuel is
used in a boiler in that period, YFUij kt becomes 1 in Eq. (3.5), where M is a large
number. Equation (3.6) states that only one type of fuel can be used and mixed with
biodiesel in a period. Equations (3.7) and (3.8) model the electricity and MP steam
consumption in the boiler as a function of the HP steam generation. Equation (3.9)
models the SOx generation which is a function of the composition of the fuel and
the amount of fuel consumption in the boilers. Equation (3.10) determines the up-
per and lower bounds on the amount of HP steam generation in the boilers, if the
boiler is operating.
3.3.1.2 Turbine Models

Turbines generate electricity by expanding steam from higher pressures to lower
pressures. They receive HP steam and produce electricity as well as MP and LP
steams. Electricity generation in a turbine is a function of HP steam input and MP
and LP steam generation as shown in Eq. (3.11). The material balance around tur-
bines is expressed in Eq. (3.12). Equation (3.13) determines the upper and lower
bounds on the amount of MP, LP and electricity generation in turbines, if the tur-
bine is working. The parameters, eij k and gij k can be obtained from either design
specifications of the turbine or the operating data of existing turbines.

Xij kEL lgen t = eij kHP Xij kHP lin t − gij k Xij klgen t , (3.11)
k=MP,LP
Xij kHP lin t = Xij kMP lgen t + Xij kLP lgen t , (3.12)
L U
Xij klgen × Yij t Xij klgen t Xij klgen × Yij t . (3.13)
3.3.1.3 Fuel Tank Models

Different types of fuel are stored in fuel tanks with certain storage capacities and
initial inventory, Iij k0 . Equation (3.14) models the balance between a tank and the
boilers that use the fuel. Material balance around a fuel tank is modeled by Eq.
(3.15) such that the rate of flow out of tanks times the duration of period t plus
inventory at time t is equal to incoming fuel plus fuel remaining from the previous
period. Equation (3.16) is equivalent of Eq. (3.15) for the first time period. Equation
(3.17) enforces the inventory at any period to be between the total storage capacity
of the fuel tank and the safety stock level. Binary variable YPij kt is equal to 1 if fuel
k is purchased for tank j of company i in period t. There is an upper and a lower
limit for the fuel purchase amount as shown in Eq. (3.18). The cost of purchased
fuel is modeled in Eq. (3.19). Equation (3.20) models the fixed cost of purchase in
terms of the fixed cost of purchase νij k and the binary variable YPij kt . Finally, Eq.
(3.21) models the holding cost of fuel inventory, HCt , in terms of unit holding cost,
hij k and inventory level, Iij kt .

Xij kfuel lout t = Xij kfuel lin t , (3.14)
j ∈Boiler

Iij k(t−1) + Xij kfuel lin t − Iij kt − nt × Xij kfuel lin t = 0, t > 1, (3.15)
j ∈Boiler

Iij k0 + Xij kfuel lin t − Iij kt − nt × Xij kfuel lin t = 0, t = 1, (3.16)
j ∈Boiler
IijUk δ Iij kt IijUk , (3.17)

L U
pij kfuel × YPij kfuel t Xij kfuel lin t pij kfuel × YPij kfuel t , (3.18)

Ci = cfuel Xij klin t , (3.19)
j t k∈fuel

CPi = νij k YPij kt , (3.20)
j k∈fuel t

HCi = hij k Iij kt . (3.21)
j k∈fuel t
3.3.1.4 Mixer Models

Mixers receive and send one type of material from and to different units. There
is a mixer for each type of material in the system. Equation (3.22) represents the
material balances around mixers. In a steam mixer, the total amount of steam that
flows into the mixer from boilers, from other mixers and from other companies is
equal to the total amount of steam that flows from the mixer to the turbines, to the
boilers, to other mixers, to other companies and the demand.

Xij klout t + XEij i j t = Xij klin t + XEi j ij t + di j t . (3.22)
j j j j
3.3.1.5 Environmental Considerations

The boilers release GHG and SOx as waste products that results from burning
fuels. A model for energy production systems should include environmental limits.
Equations (3.23) and (3.24) state that the total releases of the companies should be
less than sum of their limits. The SOx emission limits are not included in the
Kyoto Protocol, but they are determined by local regulations. The total SOx and
GHG emissions are calculated over all periods. Here, the emission is calculated by
multiplying the emission rate by the length of period t, nt . Equation (3.25) models
the penalty cost of SOx release. Although the companies must decrease the GHG
emissions levels according to Kyoto Protocol, as long as they are below the limits,
they do not pay penalty for GHG emissions.

U
Xij kGHG lgen t × nt Xik GHG
, (3.23)
i j t i

U
Xij kSOx lgen t × nt Xik SO
, (3.24)
x
i j t i

CSi = Xij kSOx lgen t ςSOx × nt . (3.25)
t j
3.3.1.6 Material Balance

Equation (3.26) relates the states of materials to reflect the conservation of mass.
In order to maintain consistency in the material balances, Eq. (3.27) fixes some of
the states of materials to zero (for example, HP steam is not consumed and not an
input to boilers, so these variables are fixed to zero).
Xij klin t + Xij klgen t = Xij klout t + Xij klcon t , (3.26)
Xij kl t = 0. (3.27)
3.3.1.7 Electricity Purchase

The companies can buy electricity from the utility company and sell the excess
electricity to the utility company. This trade is modeled as an exchange activity
between the energy producing company and the utility company. The parameter
εij i j is positive for purchasing and negative for selling electricity. The electricity
cost for each company is determined with Eq. (3.28).

CELi = XEij i j t εij i j . (3.28)
j i j t
3.3.1.8 Operating and Startup Costs

If a boiler or a turbine of a company is operating in period t, the company spends
a fixed amount of money. The operating cost is modeled for boilers and turbines
with Eq. (3.29). While a process unit does not work in a period and works in the
next period, the company pays a fixed cost for the startup operation. The startup
cost for boilers and turbines is modeled with Eq. (3.30). Equation (3.31) models the
timing of startup such that if a unit does not work in a period and works in the next
period, the next period must be a startup period.

CWi = Yij t × oij , (3.29)
j t

CSUi = YSij t × σij , (3.30)
j t
Yij (t−1) − Yij t + YSij t 0, t > 1. (3.31)
3.3.1.9 Carbon Capture and Sequestering (CCS) Systems

The carbon capture and sequestering (CCS) involve capturing carbon emissions
from the boilers of fossil-fuel-based boilers and then injecting it underground.
There are three basic design systems: postcombustion, oxygen-combustion, and
precombustion [15]. Postcombustion capture has an important role in making
fossil-fuel-based energy production systems environmentally friendly in the transi-

tion period, since it can capture from the exhaust released from the plant. There-
fore, this technology is considered as an alternative to further reduce emissions in
the model. Benson [16] states that it has an “energy penalty” that it uses up to 30%
of the electricity produced. The separated carbon can be sequestrated in depleted
oil and gas reservoirs, coal-bed reservoirs and salt water filled formations. Burrus
[17] estimates that only depleted oil and gas reservoirs have a capacity for 40–50
years injection. CCS is a very complicated and costly investment, in addition to its
energy penalty. However, as technology matures its investment costs are expected
to decrease.
CCS system can be constructed by the companies as a low-carbon technology.
The CCS system is not modeled as a new unit in the system; its existence is mod-
eled by the interactions between disjunctions and Boolean variables. CCS system
can capture the CO2 equivalent materials emitted by the boilers at different per-
centages of fuel from different types of fuels. The capture ratio from natural gas
is about twice of the capture ratio from diesel. Since the type of fuel used and the
existence of CCS system have to be distinguished, a new variable representing the
type of fuel used in the boiler under the existence or absence of CCS system is
defined, Fij kct . Moreover, since the GHG emission to the atmosphere depend on
the existence of CCS, a new variable is introduced for the emitted GHG, disaggre-
gated on index c which indicates whether a CCS system is to be installed or not,
represented by Gij kct .

Xij kGHG lgen t = Gij kfuel ct , (3.32)
k c

Xij kfuel lcon t = Fij kfuel ct , (3.33)
c
Gij kct = sGHG,kc × Fij kct , (3.34)

Gij kcCCS t CCEit × M, (3.35)
Gij kcNoCCS t (1 − CCEit ) × M, (3.36)

t−1
CCEit CCit , (3.37)
t =1

CCit 1, (3.38)
t
Equation (3.32) states that the total of disaggregated GHG variables sum to total
GHG emissions and Eq. (3.33) states that the total of disaggregated fuel consump-
tion variables sum to total fuel consumption. Equation (3.34) sets the value of GHG
emissions with and without CCS system with the updated GHG emission parame-
ters. Equations (3.35) and (3.36) regulate the GHG emissions according to whether
CCS system exists for a company i, denoted by CCEit . Equation (3.37) ensures the
existence of CCS at period t if it has been constructed before t, and Eq. (3.38) limits
the CCS construction for a company i with 1.
The CCS in an energy production system uses some of the electricity produced
in the turbines. In order to incorporate this into the model, the production from
turbines was also disaggregated according to existence of CCS.
3.3.1.10 Objective Functions

The financial objective is the minimization of the total cost that consists of cost
of fuel purchased, fixed cost of purchase, holding cost of fuel, cost of installing
exchange equipment, penalty for SOx release and cost of electricity purchase.

min zCOST = (Ci + CPi + HCi + CEi + CSi + CELi + CWi + CSUi ).
i (3.39)
The environmental objective consists of the emission of GHG from the energy
production system.

min zGHG = sGHG,kfuel Xij kfuel lcon t . (3.40)
i j k∈fuel l t
3.3.1.11 Illustrative Example

In order to understand the model behavior accurately, the model is solved for an
energy producing companies whose schematic flowsheet is given in Fig. 3.3. The
energy production system has three fuel tanks, two boilers, two turbines, and one
mixer for each pressure level of steam.
The example energy production system is optimized using the mathematical
model given in Eqs. (3.2)–(3.40). Since the resulting optimization model is a mixed-
integer multiobjective optimization problem, the efficient frontier that includes the
collection of noninferior solutions exhibit gaps as shown in Fig. 3.4. The efficient
frontier is generated with the ε-Constraint method [18].
An important consideration is the installation of the CCS system. The CCS sys-
tem is able to reduce total GHG emissions while increasing the costs significantly
resulting in another gap in the efficient frontier. The first gap in the efficient fron-
tier shown in Fig. 3.4 corresponds to the discrete decision to install a CCS system.
Another important decision to examine is the usage of biofuel. Although biofuel
has lower calorific value compared to fossil fuels, the use of biofuel reduces the
GHG emissions while increasing the cost in the illustrative example. This corre-
sponds to the second gap in the efficient frontier shown in Fig. 3.4. The last gap in
the efficient frontier corresponds to the use of different fossil fuels.
When the decision maker does not want to degrade from any of the objectives,
the ideal compromising solution search method provides a guideline for the selec-
tion process. The goal of the ideal solution search method is to find the solution
which is closest to the utopia point [19]. The distance should be designed to equally
deal with all objective values, so they should be normalized between 0 and 1. The
following normalization can be applied to all points (for all i = 1, . . . , I ) on the
efficient frontier.
zi − zmin
z̃i = . (3.41)
zmax − zmin
Fig. 3.4 The efficient frontier for the problem in the illustrative example.
The distance (δp ) between the utopia point and the efficient points is defined with
Eq. (3.42) where p is the order of the norm:

p
p p1
δp = z̃GHG − zGHGmin + z̃COST − zCOSTmin 1 p ∞. (3.42)
The distance depends on the particular norm value, p. For example, for p = 2, the
distance is Euclidean distance that can be formulated as in Eq. (3.43).
p =2

min (z̃GHG )2 + (z̃COST )2 (3.43)
x
s.t. x ∈ Constraint Set.
Using p = 2 makes the model nonlinear, because of the square and root functions
of the normalized values. However, selecting the norms p = 1 and p → ∞ will
give the following deviations from the utopia point:
p=1
min {z̃GHG + z̃COST } (3.44)
x
p→∞
min max {z̃GHG + z̃COST } (3.45)
x
Equation (3.44) suggests using rectilinear distances and Eq. (3.45) suggests using
minimax distances. Equation (3.45) can be formulated as Eq. (3.46).
p→∞
min χ
x
s.t. χ z̃GHG (3.46)
χ z̃COST
x ∈ Constraint Set.
By selecting the norms p = 1 and p → ∞, the formulations do not change the
mixed-integer linear structure of the model, since Eqs. (3.44) and (3.45) are linear.
The optimum solutions for p = 1 and p = ∞, provide lower and upper bounds,
respectively for the sum of fractional deviations from the utopia point [19]. If there
is no other specific criterion for selection of an efficient point, the decision maker
can use p = 1 if he/she wants the minimum of the total of displacements from
the minimum values for environmental and economic objectives. And the decision
maker can use p → ∞ if he/she wants to minimize the maximum displacements
of the objective functions from the utopia point. In order to give the decision maker
the flexibility of selecting a desired solution on the efficient set, finding a sample
of efficient set and making the decisions from this set is preferred to the best com-
promise method.
3.4
Sustainable Transportation
Transportation systems are essential components in realizing the transfer of mate-

rials among the nodes of supply chains. In addition to incurring costs, transporta-
tion systems have a significant share in total CO2 emissions as shown in Table 3.1.
Since the transportation has strong effects on the financial and environmental per-
formance of supply chains, the financial and environmental improvements in the
transportation systems results in the performance improvements of supply chains.
The transportation activities are carried out using the following modes:
1. road transportation;
2. rail transportation;
3. sea transportation;
4. air transportation;
5. pipeline transportation.
A single or a combination of different transportation modes can be used in a

supply chain depending on the links among the nodes of the supply chain. If very
large quantities of liquid are to be transported between a supplier and a production
center, then it is possible to use pipeline transportation alone when these two nodes
are linked with a pipeline. When these nodes are not directly linked to each other
with a pipeline, then it is necessary to use a combination of different modes. For
example, petroleum supply chains use pipeline links to transfer oil between the
3.4 Sustainable Transportation 101
source and the sea ports, and then sea transportation is used to transport oil from
the seaport to the production center. Another example is the automotive supply
chains where a combination of sea transportation, rail transportation, and road
transportation are used to deliver finished automobiles from production centers
to customers. Modern supply chains continuously seek to improve the financial
efficiency of their transportation systems.
3.4.1
Intermodal Transportation
The use of different transportation modes in an integrated manner is an efficient

approach to improve the financial efficiency of the transportation systems. Inter-
modal transportation is the integration of more than one transportation mode
(road, rail, sea, air, and pipeline) in a single transportation chain [20]. The unit cost
of transportation in an intermodal setting offers financial advantages over single-
mode transportation. Different transportation modes have different cost structures
per unit distance traveled. In addition, fixed setup and changeover costs are in-
curred when the mode of transportation changes at terminals. Since changing
transportation modes at terminals do not involve displacement of material towards
destination, the setup and changeover costs are added to the total costs play a criti-
cal role in determining whether intermodal transportation is advantageous or not.
Beyond a certain breakeven point, intermodal transportation becomes economi-
cally more advantageous than single-mode transportation as shown in Fig. 3.5.
In addition to financial advantages, intermodal transportation has environmen-
tal advantages. An important environmental advantage is the emission of CO2 .
Similar to the cost of transportation, different transportation modes have different
emission levels per distance per weight of material being transported. Among the
transportation modes, the CO2 emissions per unit distance per unit weight can be
ranked from highest to the lowest as follows:
1. air transportation;
2. road transportation;
3. rail transportation;
4. sea transportation;
5. pipeline transportation.
Changing transportation modes at terminals involves the use of material han-

dling equipment such as cranes and forklifts, the emissions by these equipment
do not contribute towards moving the displacement of material towards desti-
nation, a discrete increase in the CO2 emissions are observed. Beyond a certain
breakeven point, intermodal transportation becomes environmentally more advan-
tageous than single-mode transportation as shown in Fig. 3.6.
The success of intermodal transportation highly depends on efficient solution
of two problems: the design and installation of terminal and ports that integrate
different modes of transportation and the scheduling of trips. The first problem in-
volves strategic level decisions while the second problem involves operational level
Fig. 3.5 Cost of transportation in intermodal setting.
Fig. 3.6 Emission of CO2 in intermodal transportation.
decision. Macharis and Bontekoning [21] discuss the strategic level decisions in in-
termodal transportation including the facility location and layout and fleet sizing.
Newman and Yano [22, 23] discuss the operational level decisions and compare
distributed and centralized scheduling approaches in intermodal transportation.
They recommend centralized scheduling although it is impractical to implement
in a realistic situation.
3.5 Conclusions 103
Intermodal transportation is a very interesting approach in transportation since

it involves economical and environmental advantages simultaneously. It is possi-
ble to reduce the total cost of transportation while decreasing CO2 emissions with
intermodal transportation.
3.5
Conclusions
The success in supply chain systems not only depends on the ability to deliver the
material to customers on time with minimum cost but also with minimum ef-
fect on environment. This fact forces many contemporary supply chain systems to
pay closer attention to the effects that the supply chains are having on the environ-
ment. The major environmental effects of a supply chain system can be categorized
into three:
1. products;
2. production systems;
3. transportation systems.
Three different approaches are studied in this chapter to improve environmental

performance of supply chain systems in addition to the economical performance.
The product centric approaches involve avoiding the use of environmentally harm-
ful materials in the product at the product design stage and the recovery and reuse
of the product after it has been consumed by the end user. Closed-loop supply
chains are examined in detail for product centric approaches. The production sys-
tem centric approaches consider the selection of raw materials and the design of
the production systems for minimizing the environmental impact. The energy pro-
duction systems are considered to illustrate the economical and environmental ef-
fects of using different raw materials including biofuels and the addition of new
equipment to the production process. Transportation centric approaches consider
the use of different transportation systems that would reduce the environmental
effects. Intermodal transportation is an effective approach to reduce costs as well
as the emission of environmentally harmful substances.
Acknowledgment
The financial support for this work by TUBITAK Career Project 104M322 is grate-
fully acknowledged.
References
1 EIA, Annual Energy Review 2005, En- optimal synthesis of process net-
ergy Information Administration, works. Comput. Chem. Eng. 20 (1996),
Report No. DOE/EIA-0384, 2005. pp. 959–978.
2 Handfield, R., Sroufe, R., Walton, 12 Türkay, M., Oruç, C., Fujita, K.,
S., Integrating environmental man- Asakura, T., Multi-company collabo-
agement and supply chain strategies. rative supply chain management with
Bus. Strat. Env. 14 (2005), pp. 1–19. economical and environmental con-
3 Anderson, L., Bateman, T., Individual siderations. Comput. Chem. Eng. 28
environmental initiative: champi- (2004), pp. 985–992.
oning environmental issues in U.S. 13 Soylu, A., Oruç, C., Türkay, M., Fu-
business organizations. Aca. Manag. J. jita, K., Asakura, T., Synergy analysis
43(4) (2000), pp. 548–570. of collaborative supply chain man-
4 Fleischmann, M., Bloemhof- agement in energy systems using
Ruwaard, J. M., Dekker, R., van der multi-period MILP. Eur. J. Oper. Res.
Laan, E., van Nune, Jo A. E. E., Van 174(1) (2006), pp. 387–403.
Wassenhove, L. N., Quantitative mod- 14 Belyaev, L. S., Marchenko, O.
els for reverse logistics: a review. Eur. V., Filippov, S. P., Solomin, S. V.,
J. Oper. Res. 103 (1997), pp. 1–17. Stepanova, T. B., Kokorin, A. L.,
5 Luo, Y., Zhou, M., Caudill, R. J., An World Energy and Transition to Sustain-
integrated E-supply chain model for able Development, Kluwer, Dordrecht,
agile and environmentally conscious 2002.
manufacturing. IEEE/ASME Trans. 15 Morgan, G., Apt, J., Lave, L., The
Mech. 6(4) (2001), pp. 377–386. U.S. Electric Power Sector and Climate
6 Kim, K., Song, I., Kim, J., Jeong, B., Change Mitigation, Carnegie Mellon
Supply planning model for reman- University, 2005.
ufacturing system in reverse logis- 16 Benson, S. M., Carbon Dioxide Cap-
tics environment. Comp. Ind. Eng. 51 ture and Storage in Underground Geo-
(2006), pp. 279–287. logic Formations, Pew Center / NCEP,
7 Beamon, B. M., Designing the green the 10–50 Solution, Technologies and
supply chain. Logistics Inf. Mgmt. 12(4) Policies for a Low-Carbon Future
(1999), pp. 332–342. Workshop, Washington DC, USA,
2004.
8 EIA, What Does the Kyoto Protocol
Mean to U.S. Energy Markets and 17 Burruss, R. C., Contributing Paper:
the U.S Economy, Energy Informa- Geologic Sequestration of Carbon Diox-
tion Administration, retrieved from ide in the Next 10 to 50 Years: An En-
www.eia.doe.gov on 02.19.2007, 2007. ergy Resource Perspective, Pew Center /
NCEP, the 10–50 Solution, Technolo-
9 UNFCC, Climate Change Information gies and Policies for a Low-Carbon
Kit, retrieved from http://www.unfccc. Future, Workshop, Washington DC,
int/resource/iuckit/cckit2001en.pdf, USA, 2004.
2001.
18 Haimes, Y., Ladson, L., Wismer, D.,
10 Raman, R., Grossmann, I. E., Model- On a bicriterion formulation of the
ing and computational techniques problems of integrated system iden-
for logic based integer program- tification and system optimization.
ming. Comput. Chem. Eng. 18 (1994), IEEE Trans. Syst., Man Cybernetics 1
pp. 563–578. (1971), pp. 269–297.
11 Türkay, M., Grossmann, I. E., Logic- 19 Grossmann, I. E., Drabbant, R.,
based MINLP algorithms for the Jain, R. K., Incorporating toxicol-
ogy in the synthesis of industrial
References 105
complexes. Chem. Eng. Commun. 17 Eur. J. Oper. Res. 153 (2004), pp. 400–
(1982), pp. 151–170. 416.
20 Jones, W. B., Cassady, C. R., Bowden, 22 Newman, A. M., Yano, C. A., Central-
R. O. Jr., Developing a standard def- ized and decentralized train schedul-
inition of intermodal transportation. ing for inter-modal operations. IIE
Transport. Law J. 27(3) (2000), pp. 345– Transact. 32 (2000a), pp. 743–754.
352. 23 Newman, A. M., Yano, C. A., Schedul-
21 Macharis, C., Bontekoning, Y. M., ing direct and indirect trains and
Opportunities for OR in inter-modal containers in an inter-modal setting.
freight transport research: a review, Transport. Sci. 34(3) (2000b), pp. 256–
270.
107
4
Optimal Design and Operational Planning of Responsive
Process Supply Chains
Fengqi You, Ignacio E. Grossmann
This chapter addresses the problem of optimal design and operational planning
of multiechelon, multisite process supply chain networks (PSCNs) with consider-
ation of responsiveness and profitability. A quantitative characterization of respon-
siveness for PSCNs is presented, which measures the response time or lead time
to changes in demands assuming zero inventories. This measure is incorporated
in a multiperiod mixed-integer nonlinear programming (MINLP) model, which
considers the selections of suppliers and manufacturing sites, process technology,
production levels, and scheduling. The problem is formulated as a bicriterion op-
timization model in which the objectives are to maximize the net present value
(NPV) and to minimize the lead time. This allows establishing trade-offs between
the economics and responsiveness of the supply chain network. The model pro-
duces a Pareto-optimal curve, which reveals how the optimal NPV, and therefore
the network structure of the PSCN, changes with different specifications of lead
time. The application of this model is illustrated through an example. The mod-
eling approach developed in this chapter and the results obtained suggest that a
new conceptual strategy has been developed that yields useful insights regarding
the responsiveness of PSCNs.
4.1
Introduction
The importance of response time as a competitive advantage for customer satis-

faction and market understanding has been recognized for some time [1]. There is
a growing recognition that individual business no longer competes as stand-alone
entities, but rather as supply chains [2], whose success or failure is ultimately de-
termined in the marketplace by the end consumer. The need to meet the demands
of customers for ever-shorter lead times, and to synchronize the supply to meet the
peaks and troughs of demand, requires the capability to manage the supply chain
in a way that enables quick response [3], which is of critical importance in this
area of time-based competition. Responsiveness is defined as the ability of a supply
ISBN: 978-3-527-31693-9
108 4 Design and Planning of Responsive Process Supply Chain Networks
chain to respond rapidly to changes in demand, both in terms of volume and mix
of products [4]. In today’s rapid changing industrial environment, responsiveness
has become not only the linchpin of companies to competitive success but also the
key to survival [4–6].
In the process industries, enterprise-wide optimization (EWO) has become a ma-
jor goal due to the increasing pressure for remaining competitive in the global
marketplace [7]. EWO involves optimizing the operations of supply, manufactur-
ing, and distribution activities of a company to reduce costs. While there has been
considerable progress in the development of the planning and scheduling models
that can be used to address specific problems, a major challenge that remains is
to integrate the consideration of responsiveness into the design and operations of
supply chain networks.
However, most of the research on responsive supply chains has been approached
qualitatively and no mathematical models are available for its modeling and opti-
mization. The major goal of this chapter is to propose a novel design and oper-
ational planning optimization model of process supply chain networks (PSCNs)
that takes into account the responsiveness. In this chapter, we propose a quan-
titative characterization of responsiveness for PSCNs, which determines the lead
time under the assumption of zero inventories. We incorporate this measure into
a multiperiod mixed-integer nonlinear programming (MINLP) model, which takes
into account the selections of suppliers, manufacturing sites, process technology,
production levels, and cyclic scheduling. Multiperiod modeling allows economic
parameters, such as demand and price, to vary with time. To establish the trade-
offs between the economics and responsiveness of the supply chain network, we
formulate a bicriterion optimization model in which the objectives are to maximize
the net present value (NPV) and to minimize the lead time. The multiobjective opti-
mization model yields a Pareto-optimal curve, which reveals how the optimal NPV,
and therefore the network structure of the PSCN, changes with different specifica-
tions of responsiveness.
The chapter is organized as follow. We first review the relevant literature. This
is followed by a motivating example and a formal definition of the problem ad-
dressed in this chapter. Next, the MINLP formulation for the multiobjective opti-
mization model is described. Results for an example are presented and conclusions
are drawn about this work.
4.2
Literature Review
The “responsiveness” literature for supply chains is highly qualitative and concep-
tual, and has not been subjected to the kind of quantitative analysis that is proposed
in this chapter. There are, however, related works that offer relevant insights. For-
rester [8] illustrated in a series of case studies the effect of dynamics in industrial
systems, which gives rise to the “bullwhip effect”. Lee et al. [9] further demonstrated
that “bullwhip” effect is a consequence of the information delay due to the structure
of supply chains, and the severity of this effect is positively related to lead times.
Responsiveness in the wider supply chain context has been discussed by Fisher
[5], who argues that the product characteristics (innovative or functional) and life
cycles need to be linked to the layout and functions (conversion and market medi-
ation) of the supply chain. He also pointed out that reducing the lead time enables
quick response to unpredictable demand to minimize stockouts, markdowns, and
obsolete inventory. Matson et al. [10] discussed the concepts and issues associated
with responsiveness in production and illustrated the audit tools they proposed
from a case study in the steel industry. Recently, several conceptual models on sup-
ply chain responsiveness have been proposed. Christopher and Towill [2] integrate
lead time and agility to highlight the differences in their approach, and combined
them to propose an integrated hybrid strategy for designing cost-effective respon-
sive supply chain with seamless connection between manufacturing and logistics.
In a later work, Yusuf et al. [11] have reviewed emerging patterns for creating re-
sponsive supply chain based on survey research driven by a conceptual model.
Holweg [12] proposed in his paper that product, process, and volume are three
key factors that determine the responsiveness of a supply chain system, and pro-
vided guidelines on how to align the supply chain strategy to these three factors in
order to balance responsiveness to customer demand and supply chain efficiency.
An examination on supply chain system in process industries from a responsive-
ness view point was carried out by Shaw et al. [13]. These authors also proposed a
conceptual management strategy to improve the responsiveness of process supply
chain system.
Another group of relevant papers to be considered are on supply chain design
and operation. A general review of this area is given in Kok and Graves [14], and a
specific review for supply chains in process industries is presented by Shah [15].
Some recent works include the following. Tsiakis et al. [16] presented a supply
chain design model for the steady-state continuous processes. Their supply chain
model was developed based on determining the connection between multiple mar-
kets and multiple plants with fixed locations. Jackson and Grossmann [17] pre-
sented a temporal decomposition scheme based on Lagrangean decomposition for
a nonlinear programming problem that models multisite production planning and
distribution models, where nonlinear terms arise from the relationship between
production and physical properties or blending ratios. Schulz et al. [18] described
two multiperiod MINLP models for short-term planning of petrochemical com-
plexes. Linearization techniques are applied to reformulate the nonconvex bilinear
constraints as mixed-integer linear programming (MILP) models. Sousa et al. [19]
presented a two-stage procedure for supply chain design with responsiveness test-
ing. In the first stage, they design the supply chain network and optimize the pro-
duction and distribution planning over a long time horizon. In the second stage,
responsiveness of the first stage decisions are assessed using the service level to the
customers (i.e., delay in the order fulfillment). However, all these models consider
supply chain networks with only dedicated processes. Multiproduct batch plants or
flexible processes were not taken into account, and hence no scheduling models
were included.
There are works on supply chain optimization with consideration of flexible

processes in the production network, but most of them are restricted to planning
and scheduling for a given facility in a fixed location, without extension to the mul-
tisite supply chain network design problems. Bok et al. [20] proposed a multiperiod
supply chain optimization model for operational planning of continuous flexible
process networks, where sales, intermittent deliveries, production shortfalls, de-
livery delays, inventory profiles, and job changeovers are taken into account. A
bilevel decomposition algorithm was proposed, which proved to reduce the compu-
tational time significantly. Kallrath [21] described a tool for simultaneous strategic
and operational planning in a multisite production network, where key decisions
include operating modes of equipment in each time period, production and sup-
ply of products, minor changes to the infrastructure and raw material purchases,
and contracts. A multiperiod model is formulated where equipment may undergo
one mode change per period. The standard material balance equations are adjusted
to account for the fact that transportation times are much shorter than the period
durations. Chen et al. [22] presented a multiproduct, multistage, and multiperiod
production and distribution planning model. They also proposed a two-phase fuzzy
decision-making method to obtain a compromise solution among all participants
of the multienterprise supply chain.
Thus, these papers either focus only on the long-term strategic design models or
are restricted to short-term planning and scheduling models. Hence, no quantita-
tive analyses are available for responsive supply chains. It is the goal of this chapter
to integrate supply chain decisions in different time scales (design, planning, and
scheduling), taking into account responsiveness with a quantitative approach. A
multiobjective optimization procedure is used in this work for trading off the two
different types of objectives – economics and responsiveness.
4.3
Motivating Example
To illustrate the issues associated with the problems addressed in this chapter,
consider the example in Fig. 4.1, which is taken from a real-world application.
A company wants to design its supply chain network in the USA for producing
polystyrene resins. Three types of candidate plants are included in the process net-
work (Fig. 4.1). Plant I is used to produce styrene monomers from ethylene and
benzene; Plant II is a multiproduct plant for the production of three different types
of solid polystyrene (SPS) resins; Plant III is also a multiproduct plant for the pro-
duction of two different types of expandable polystyrene (EPS) resins. As shown in
Fig. 4.2, two potential ethylene suppliers are located in OH and TX, and two po-
tential benzene suppliers are located in VA and TX. Three potential manufacturing
sites can be located in PA, TX, and AL. The PA manufacturing site can set up all
the three types of plants, the TX manufacturing site can only install Plant I, and
the AL manufacturing site can only set up Plants II and III. The customers for SPS
resins are located in CA and NY, while customers for EPS resins are located in GA
SPS - 1
II SPS - 2
Ethylene Styrene Multi Product SPS - 3
I
Benzene Single Product EPS - 1
III
EPS - 2
Multi Product
Fig. 4.1 Production network for motivating example.
Fig. 4.2 Location map for motivating example.
and MN. The corresponding superstructure of the supply chain network is given
in Fig. 4.3. Based on the given information, we want to design an economic and
responsive supply chain by making decisions on design (e.g., selection of suppli-
ers, plant sites, and transportation link) and operations (e.g., amount of purchase,
sale, production, and transportation). This problem is stated in general form in
Section 4.4.
4.4
Problem Statement
In general terms, the problem addressed in this work can be stated as follows.
Given is a potential PSCN that includes possible manufacturing sites, suppliers,
and customers (as shown in Fig. 4.4). A set of processes and a time horizon con-
sisting of a number of time periods are given. The processes may be dedicated or
Plant Site PA
Ethylene SPS
OH II
Styrene CA
I
III
VA Benzene EPS
NY
SPS
TX II
GA
Ethylene
Styrene
I III
Styrene
Benzene EPS MN
TX
Plant Site TX Plant Site AL
Fig. 4.3 Supply chain network superstructure.
flexible. Flexible processes are multiproduct processes that operate under differ-
ent production schemes, using different raw materials and/or producing different
products. The PSCN also involves a set of chemicals, which includes raw materi-
als, intermediate products, and final products. Demands are assumed to be known
in each time period. Raw materials are subject to availability constraints and final
products are subject to demand constraints (i.e., within lower and upper bounds).
Prices for raw materials and final products are assumed to be known over the en-
tire time horizon. For all production schemes, mass balances are expressed linearly
in terms of the main product’s production. The investment costs for installing the
plants are expressed by a cost function with fixed charges (Fig. 4.5). There could be
different transportation types, continuous (e.g., pipelines) or discrete (e.g., barges,
rail cars, tanker), for each route that connects the suppliers, plant sites, and cus-
tomers. For simplicity, we will assume that all the transportations of materials in
this problem are continuous. Thus, fixed charge cost functions provide good esti-
mations of transportation costs, and inventories for single product plants can be
neglected. The transportation time of each route and the residence time of each
product are assumed to be known.
As discussed before, responsiveness is the ability of supply chains to respond to
the change of customer demands and preferences [12]. Lead time is the time of a
supply chain network to respond to customer demand and preference changes in
the worst case. Therefore, by assuming zero inventories, lead time will be used as
a measure of responsiveness [9, 11]. As shown in Fig. 4.6, a supply chain network
with long lead time implies that its responsiveness is low and vice versa. Thus, to
design a responsive supply chain, one objective function of this problem is to min-
imize the lead time of the entire supply chain network. From the economic aspect,
another objective function is to maximize the NPV over the specified long-range
time horizon. The income from sales, along with investment, operating, trans-
portation, and purchase costs are taken into account in the NPV objective function.
Fig. 4.4 Typical process supply chain network.
Cost
0 , y 0
cost ®
¯D E x , y 1
x (capacity)
Fig. 4.5 Cost function with fixed charge.
Responsiveness
Lead Time
Fig. 4.6 Conceptual relationship between lead time and responsiveness.
Since the two conflicting objectives need to be optimized simultaneously, the

corresponding problem yields an infinite set of alternative solutions denoted by
the Pareto-optimal curve. These solutions are Pareto-optimal, in the sense that it is
impossible to improve both objective functions simultaneously [23]. This implies
that any designs, for which the NPV and the lead time can be improved simultane-
ously, are “inferior” solutions that do not belong to the Pareto-optimal curve. The
aim of this problem is to determine the supply chain network configurations that
define the Pareto-optimal solution by maximizing NPV and minimizing the lead
time.
4.5
Model
The model will be formulated as a multiperiod MINLP problem, which predicts the
detailed design decisions, production profiles, and schedules for the PSCN with
different specifications of the lead time. A list of indexes, sets, parameters, and
variables are given in the Appendix. Three types of constraints are included in this
model. They are network structure constraints, operational planning constraints,
and cyclic scheduling constraints.
Constraints (4.1) to (4.8) determine the network structure, constraints (4.9) to
(4.14) refer to the operational planning constraints, constraints (4.15) to (4.30) are
used for the cyclic scheduling of multiproduct plants. Finally, inequalities (4.33)
to (4.35) define the lead time and Eq. (4.37) defines the NPV, both of which are
objective functions to be optimized.
4.5.1
Network Structure Constraints
To determine the topology of network structure and model the selection of sup-
pliers, plant sites, together with the transportation links between them, the binary
P , Y O , Y I ) for plants, and transportation links are introduced for
variables (Yk,i k,l k,k
design decisions. Three types of network structure constraints are applied to rep-
resent the relationships between each node in the supply chain network.
4.5.1.1 Supplier: Plant Site

The first type of relationship is between suppliers and plant sites, as shown in
Fig. 4.7. A transportation link for raw material j from supplier ls to plant site k
exists, only if at least one plant that consumes raw material j exists in site k. The
relationships discussed above can be expressed by the following logic proposition:
O P
Yk,ls ⇒ ∨i∈Ij Yk,i . (4.1a)
These logic propositions can be further transformed into inequalities as de-

scribed by Raman and Grossmann [24].

O P
Yk,ls Yk,i ∀k ∈ Ki , ls ∈ LSj . (4.1b)
i∈Ij
4.5 Model 115
ls1
ls2
. Plant
. Site k
.
lsn
Fig. 4.7 Relationship between suppliers and manufacturing sites.
On the plant side, if a plant that consumes raw material j is set up, at least
one transportation link from the supplier ls to site k must be selected. The logic
propositions are:
P O
Yk,i ⇒ ∨ls∈LSj Yk,ls . (4.2a)
Then they can be transformed to inequalities:

P O
Yk,i Yk,ls ∀k ∈ Ki , i ∈ Ij . (4.2b)
ls∈LSj
4.5.1.2 Plant Site: Customer

The second type of relationship is between plant sites and customers as shown
in Fig. 4.8. A transportation link for product j from plant site k to customer ld
exists, only if at least one plant that consumes raw material j exists in site k. On
the plant side, if a plant that consumes raw material j is set up, there should be at
least one link from the customer ld to site k exist. Similarly, transforming from the
corresponding logic propositions leads to the following inequalities:

O P
Yk,ld Yk,i ∀k ∈ Ki , ld ∈ LDj , j ∈ JPk,i , (4.3)
i∈Oj

P O
Yk,i Yk,ld ∀k ∈ Ki , i ∈ Oj , j ∈ JPk,i . (4.4)
ld∈LDj
4.5.1.3 Input and Output of a Plant

The third type of network structure relationship is the input and output relation-
ship of a plant as in Fig. 4.9. This type of relationship is somewhat more compli-
cated than the previous two because the intersite transportation must be taken into
account. If an intersite transportation link from site k to site k is installed for chem-
ical j , it implies that at least one plant i in site k is installed that produces chemical
j , and also at least one plant i in site k is installed that consumes chemical j ,

YkI ,k P
Yk,i ∀k, k ∈ K, j ∈ (JMk,i ∩ JPk ,i ), (4.5)
i∈Ij

I
Yk,k
P
Yk,i ∀k, k ∈ K, j ∈ (JPk,i ∩ JMk ,i ). (4.6)
i∈Oj
ld1
ld2
.
Plant
Site k .
.
ldn
Fig. 4.8 Relationship between manufacturing sites and customers.
ls I2
…
…
Plant i
I1 ld
Fig. 4.9 Input and output relationship of a plant.
If a plant i in site k is installed, which consumes chemical j , then site k is con-

nected to one of the suppliers of chemical j denoted as ls, or connected to another
site k that produces chemical j , or there is another plant i in site k that produces
chemical j . The logic propositions can be written as follow:
P O
Yk,i ⇒ ∨ls∈LSj Yk,ls ∨k ∈Ki YkI ,k ∨i ∈Oj Yk,i
P
, (4.7a)
which can be transformed into inequalities as:

P O
Yk,i Yk,ls + YkI ,k + P
Yk,i
ls∈LSj k ∈Ki i ∈Oj
∀k ∈ Ki , i ∈ Ij , j ∈ JMk,i . (4.7b)
Similarly, if the chemical j is produced by plant i in site k, then at least one other
plant i in the same site is installed that consumes chemical j , or there is at least
one transportation link to a customer or the other site that consumes chemical j :

P O I P
Yk,i Yk,ld + Yk,k + Yk,i
ld∈LDj k ∈Ki i ∈Ij
∀k ∈ Ki , i ∈ Oj , j ∈ JPk,i . (4.8)
Constraints (4.7b) and (4.8) are defined for all the chemicals (raw materials, inter-
mediate products, final products). When the chemical j is raw material, constraint
(4.7b) will reduce to constraint (4.2b); when the chemical j is final product, con-
straint (4.8) will reduce to constraint (4.4).
4.5 Model 117
4.5.2
Operational Planning Constraints
In the operational planning model, investment in plant capacity and purchase,

sale, production, transportation, and mass balance relationships are considered,
together with the restriction of these activities due to the supply chain structure.
4.5.2.1 Production Constraints

All the chemical flows Wk,i,j,s,t associated with production scheme s in plant i of
site k other than the main product j are given by the mass balance coefficient. The
following equation relates the input (Wk,i,j,s,t ) to the output (Wk,i,j ,s,t ) of each
process by the material balance coefficient µi,j,s :
Wk,i,j,s,t = µi,j,s Wk,i,j ,s,t ∀k ∈ Ki , i ∈ Ij , j ∈ Ji ,s ,

j ∈ JPi,s , s ∈ Si , t ∈ T . (4.9)
The production amount should not exceed the design capacity Qk,i for each
process:
Wk,i,j,s,t ρi,s Qk,i Lenpt ∀k ∈ Ki , i ∈ Oj , j ∈ JPk,i , s ∈ Si , t ∈ T . (4.10)
For flexible processes, the maximum production rate rk,i,s of the each main prod-
uct j of production scheme s is proportional to the capacity of the plant [25]:
rk,i,s = ρi,s Qk,i ∀k ∈ Ki , i ∈ Oj , j ∈ JPk,i , s ∈ Si . (4.11)
4.5.2.2 Sale Constraints

L
The amount of sale Sk,j,ld,t has to lie between given the lower bounds dj,ld,t and
U
upper bounds dj,ld,t of the demand:

L U
dj,ld,t Sk,j,ld,t dj,ld,t ∀j ∈ J, ld ∈ LD, t ∈ T . (4.12)
k∈K
4.5.2.3 Upper Bound Constraints

Purchases Pk,j,ls,t from supplier ls to plant site k take place only if the transporta-
tion link between them is set up:
U I
Pk,j,ls,t Pk,j,ls,t Yk,ls ∀k, j, ls, t. (4.13a)
Intersite transportation TRk,k ,j,t from site k to site k take place only if the trans-
portation link between them is set up:
TRk,k ,j,t TRU YI

k,k ,j,t k,k
∀k, k , j, t. (4.13b)
Sales Sk,j,ld,t from plant site k to customer ld take place only if the transportation
link between them is selected:
U O
Sk,j,ld,t Sk,j,ld,t Yk,ld ∀k, j, ld, t. (4.13c)
Nonzero production flows Wk,i,j,s,t are allowed in plant i of site k only if the plant
is installed:
U P
Wk,i,j,s,t Wk,i,j,s,t Yk,i ∀k, i, j, s, t, (4.13d)
Qk,i QU YP
k,i k,i
∀k, i. (4.13e)

Mass balance for chemical j in manufacturing site k at time period t is given as
follows:

Pk,j,ls,t + TRk,k ,j,t + Wk,i,j,s,t = Sk,j,ld,t
ls∈LS k∈K i∈Oj s∈Si ld∈LD

+ TRk,k ,j,t + Wk,i ,j,s ,t ∀k ∈ Ki , j ∈ J, t ∈ T . (4.14)
k ∈K i ∈Ij s ∈Si
4.5.3
Cyclic Scheduling Constraints
To address detailed operations of the multiproduct plants, we have considered a

cyclic scheduling policy [26]. Under this policy, sequences to produce each product
are decided, together with the cycle time (Figs. 4.10 and 4.11), and then identical
schedule is repeated over each time period (Fig. 4.12). The trade-offs between in-
ventories and transitions are established by optimizing the cycle times (as shown
in Fig. 4.13).
Important decisions in cyclic scheduling, including the sequence of production
S
(Yk,i,s,sl,t ) and precedence relationship for changeovers between pairs of products
(Zk,i,s,s ,sl,t ), are restricted by assignment and sequence constraints (constraints
(4.15) to (4.19)).
Scheme A Transition
Scheme B
Scheme C
Fig. 4.10 Changeovers of flexible processes.
Transition Processing jobs
Tc
(cycle time)
Time
Fig. 4.11 Cyclic scheduling in a cycle.

4.5 Model 119
The Whole Time Period
Cycle Time Cycle Time Cycle Time
Fig. 4.12 Cyclic scheduling of each time period.
Cost
Inventory Cost
Total Cost
Transition Cost
Cycle Time
Fig. 4.13 Trade-off between inventory and transition in cyclic scheduling.
4.5.3.1 Assignment Constraints

The assignment constraints state that exactly one time slot must be assigned to one
production scheme and vice versa. The total number of time slots will be exactly
equal to the total number of products.

S
Yk,i,s,sl,t = 1 ∀k ∈ Ki , i ∈ Ij , s ∈ Si , t ∈ T , (4.15)
sl∈SLi

S
Yk,i,s,sl,t =1 ∀k ∈ Ki , i ∈ Ij , sl ∈ SLi , t ∈ T . (4.16)
s∈Si
4.5.3.2 Sequence Constraints

The sequence constraints state that exactly one transition from product s occurs
in the beginning of any time slot if and only if s was being processed during the
previous time slot. On the other hand, exactly one transition to product s occurs
in the time slot if and only if product s is being processed during that time slot.
As suggested by Wolsey [27], the transition variables Zk,i,s,s ,sl,t can be replaced
by continuous variables between 0 and 1, instead of binary variables. This signif-
icantly reduces the number of discrete variables and improves the computational
efficiency.

S
Zk,i,s,s ,sl,t = Yk,i,s ,sl−1,t
s∈Si
∀k ∈ Ki , i ∈ Ij , s ∈ Si , sl ∈ SLi , t ∈ T , (4.17)

S
Zk,i,s,s ,sl,t = Yk,i,s,sl,t ∀k ∈ Ki , i ∈ Ij , s ∈ Si , sl ∈ SLi , t ∈ T , (4.18)
s ∈Si
0 Zk,i,s,s ,sl,t 1 ∀k, i, s, s , sl, t. (4.19)


The production amount of product s in a cycle (Wk,i,s,t S ) is equal to the processing
rate rk,i,s times the processing time δk,i,s,sl,t :

S
Wk,i,s,t = rk,i,s δk,i,s,sl,t ∀k ∈ Ki , i ∈ Ij , s ∈ Si , t ∈ T . (4.20)
sl∈SLi
The amount to produce for each product in time period t (Nk,i,t cycles in the time
period) should be not less than the total production predicted from operational
planning in this time period:
S
Wk,i,s,t Nk,i,t Wk,i,j,s,t ∀k ∈ Ki , i ∈ Oj , j ∈ JPk,i , s ∈ Si , t ∈ T . (4.21)
4.5.3.4 Timing Constraints

Constraints (4.22) to (4.28) are used to restrict the timing issues in the cyclic
scheduling.
The processing time δk,i,s,sl,t in a certain time slot is equal to the summation of
the processing times assigned to all the products in this time slot

δk,i,sl,t = δk,i,s,sl,t ∀k ∈ Ki , i ∈ Ij , sl ∈ SLi , t ∈ T . (4.22)
s∈Si
The cycle time TCk,i,t is equal to the summation of all the processing times in
each time slot plus the summation of transition times in this cycle

TCk,i,t = δk,i,sl,t + Zk,i,s,s ,sl+1,t τi,s,s
sl∈SLi s∈Si s ∈Si sl∈SLi
∀k ∈ Ki , i ∈ Ij , t ∈ T . (4.23)
The start time Tsk,i,sl,t and end time Tek,i,sl,t of each time slot are given by:
δk,i,sl,t = Tek,i,sl,t − Tsk,i,sl,t ∀k ∈ Ki , i ∈ Ij , sl ∈ SLi , t ∈ T , (4.24)

Tsk,i,1,t = Zk,i,s,s ,1,t τi,s,s ∀k ∈ Ki , i ∈ Ij , t ∈ T , (4.25)
s∈Si s ∈Si

Tsk,i,sl,t = Tek,i,sl−1,t + Zk,i,s,s ,sl,t τi,s,s
s∈Si s ∈Si
∀k ∈ Ki , i ∈ Ij , sl ∈ SLi , t ∈ T . (4.26)
The total production time should not exceed the duration of each time period Hk,i,t :
TCk,i,t Nk,i,t Hk,i,t ∀k ∈ Ki , i ∈ Ij , t ∈ T . (4.27)
The production for scheme s in time slot sl can take place only if the time slot is
assigned to the production scheme:
U S
δk,i,s,sl,t δk,i,s,sl,t Yk,i,s,sl,t ∀k ∈ Ki , i ∈ Ij , s ∈ Si , sl ∈ SLi , t ∈ T . (4.28)
4.5 Model 121
Inventory
Level
a time period
Time
cycle time processing time
Fig. 4.14 Inventory level change in cyclic scheduling.
4.5.3.5 Cost Constraints

To integrate the cyclic scheduling with the strategic planning, the inventory and
transition costs from cyclic scheduling are considered as part of the operating cost.
Constraint (4.29) represents that cost from scheduling in a time period for a certain
plant. The first term on the right-hand side of the equation stands for the total
transition cost in a time period. The second term on the right-hand side of the
equation is the inventory cost for all the chemicals involved in the production. The
change of inventory level in a time period is given in Fig. 4.14. In the work by
Pinto and Grossmann [26], they consider inventory only for final products, as their
model is for single plant. In our case, each manufacturing site may have more
than one production plant, and inventory for materials of multiproduct plants must
be also taken into account. Since we assume that material balances are expressed
linearly in terms of the main product’s production, the cumulative inventory levels
for raw materials are also related linearly to the cumulative inventory level of main
product in each production scheme, and the coefficients of the linear relationships
are exactly the absolute value of mass balance coefficients. This leads to the second
term on the right-hand side of the following constraint. Thus, operating cost for
the inventory and transition of flexible processes is given by:

COSTSk,i,t = CTRi,s,s Zk,i,s,s ,sl,t Nk,i,t
s∈Si s ∈Si sl∈SLi

S
+ (|µi,j,s |εj )(rk,i,s Hk,i,t − Wk,i,s,t Nk,i,t )δk,i,sl,t /2
s∈Si j ∈Ji sl∈SLi
∀k ∈ Ki , i ∈ Ij , t ∈ T . (4.29)
This constraint is nonlinear and nonconvex, with bilinear and triple-linear term.
If all the processes in the production network are dedicated, cyclic scheduling need
not be taken into account and thus this constraint can be discarded.
4.5.3.6 Upper Bound Constraints

As a multisite problem, we need to make sure that if a plant i in site k is not
installed, the associated scheduling cost should be 0. To model this, we introduce
the upper bound constraint (4.30) for the number of cycles Nk,i,t in each time
period for each multiproduct plant in each manufacturing site:
U P
Nk,i,t Nk,i,t Yk,i ∀k ∈ Ki , i ∈ Ij , t ∈ T . (4.30)
Also assignment constraints are revised, so that all the scheduling activities can
take place only if the plant is installed:

S P
Yk,i,s,sl,t = Yk,i ∀k ∈ Ki , i ∈ Ij , s ∈ Si , t ∈ T , (4.31)
sl∈SLi

S P
Yk,i,s,sl,t = Yk,i ∀k ∈ Ki , i ∈ Ij , sl ∈ SLi , t ∈ T . (4.32)
s∈Si
4.5.4
Lead Time Definition
One of the challenges in this work is to quantitatively define the lead time of
PSCNs, which is a quantitative measure of responsiveness. The definition of lead
time in PSCNs should integrate with the supply chain network structure and op-
erational details. Lead time corresponds to the response time in the worst case,
assuming zero inventories. We define in this work the lead time of the PSCN as
the length of the longest time path of chemical flows from a supplier to a customer
by way of several manufacturing sites as shown in Figs. 4.15 and 4.16. The length
of each time path is equal to the summation of all the time delays incurred by
transportation and production in the path. If a plant or a transportation link is not
selected, the associated delay is 0. Binary variables are used to model the lead time
TP with the following inequalities:

n
O O
TP Yk,ls λk,ls + YkPx ,ix θkx ,ix +
x=1

n−1
+ YkIx ,kx+1 λIkx ,kx+1 + YkOn ,ld λO
kn ,ld
x=1
∀(ls, k1 , k2 . . . kn , ld) ∈ Pathls,k,l . (4.33a)
According to constraint (4.33), the length of each time path of chemical flow is
equal to the summation of products of all the corresponding binary variables (for
installation of transportation links and manufacturing plants) multiplied by the
associated time delay (by transportation or production). Furthermore, the lead time
must be greater than the lengths of time for all the paths. The set Pathls,k,l includes
all the possible paths of chemical flow in a given potential PSCN network. All the
elements in the set Pathls,k,l are in the form of ls, k1 , k2 . . . kn , ld, where supplier ls
is the start of the path, k1 , k2 . . . kn are the manufacturing sites that the associated
chemical flow goes through, and customer ld is the end of the path of the chemical
flow.
It is obvious that the time delay in each transportation link is equal to the trans-
portation time, and for dedicated process, the time delay by production is equal to
the residence time.
The production time delay for a multiproduct plant is not so obvious. Before in-
troducing our definition, consider a motivating example as shown in Figs. 4.10 and
4.17. A multiproduct plant produces three chemicals A, B, and C. Assume there
4.5 Model 123
Fig. 4.15 A path of chemical flow in a PSCN.
Time delayed by Path 1
Time delayed by Path n

Time
Lead Time
Fig. 4.16 Lead time of PSCNs.
is a demand change of chemical A. The worst case is when we just finished pro-
ducing A, and there is no extra inventory of A besides those committed to former
demand, then we must wait until the plant produces A again, so that we can adjust
the production to meet the demand change. This takes some time which is given
by the processing time of chemical B and C, plus residence time of A. In this way
we define for multiproduct plant, the time delay for each product as cycle time plus
Processing
Time
Time delayed by Production
Input
Output
Residence Cycle Time

Time
Fig. 4.17 Time delay by production.
residence time minus its processing time. Therefore, the time delay by production
for a multiproduct plant is equal to the maximum time delay for each product:

R
θk,i TCk,i,t + θi,s − δk,i,s,sl,t ∀k ∈ Ki , i ∈ Ij . (4.34)
sl∈SLi
In this definition, cycle times of each plant are taken into account as part of the
delay due to production, so that we have integrated the production details into the
quantitative definition of responsiveness.
The bilinear terms (binary variable times the time delay of flexible process) in
the lead time definition can be linearized. We use a continuous variable TYk,i to
P θ
replace the Yk,i k,i term in the lead time constraint:
P
Yk,i θk,i = TYk,i ∀k, i, (4.35a)

n
n−1
O O
TP Yk,ls λk,ls + TYkx ,ix + YkIx ,kx+1 λIkx ,kx+1 + YkOn ,ld λO
kn ,ld
x=1 x=1
∀(ls, k1 , k2 . . . kn , ld) ∈ Pathls,k,l . (4.33b)
The Eq. (4.35a) is equivalent to the following disjunction:

P
Yk,i ¬Yk,i
P
∨ . (4.35b)
TYk,i = θk,i TYk,i = 0
Applying the convex hull reformulation [28] to the above disjunctive constraint
leads to:
TYk,i + TY2k,i = θk,i ∀k, i, (4.35c)

U P
TYk,i θk,i Yk,i ∀k, i, (4.35d)
U P
TY2k,i θk,i (1 − Yk,i ) ∀k, i, (4.35e)
where TY2k,i is a new continuous variable introduced as a slack variable.

It is interesting to note that this model reduces to a MILP problem if all the
processes are dedicated, since the nonlinear terms in the lead time constraints are
linearized and the time delays of the dedicated plants are known.
4.5 Model 125
4.5.5
Nonnegative Constraints
All continuous variables must be nonnegative and the binary variables should be
integer:
Qk,i , Wk,i,j,s,t , Pk,j,ls,t , TRk,k ,j,t , Sk,j,ld,t , TP 0, (4.36a)
S
Zk,i,s,s ,sl,t , Wk,i,s,t , rk,i,s , δk,i,s,sl,t , δk,i,sl,t , TCk,i,t , Tek,i,sl,t ,
Tsk,i,sl,t , Nk,i,t , θk,i , COSTSk,i,t 0, (4.36b)
O P I S S
Yk,l , Yk,i , Yk,k , Ym,ld , Yk,i,s,sl,t ∈ {0, 1}. (4.36c)
4.5.6
Net Present Value
The NPV of the supply chain network is given by the following equations:
NPV = Income − Cpurchase − Coperate − Ctranport − Cinvest , (4.37)
where

Income = ϕj,ld,t Sk,j,ld,t ,
k j ld t

Cpurchase = ϕj,ls,t Pk,j,ls,t ,
k j ls t

Coperate = σi,s,t Wk,i,j,s,t + COSTSk,i,t ,
k i s j ∈JPi,s t k i t

O I
Ctranport = ωk,j,ls,t Pk,j,ls,t + ωk,k ,j,t TRk,k ,j,t
k j ls t k k j t

O
+ ωk,j,ld,t Sk,j,ld,t ,
k j ld t

P P P O O
Cinvest = ωk,i Qk,i + γk,i Yk,i + γk,l Yk,l
k i k i k l

I I
+ γk,k Yk,k .
k k
All the parameters in the above formulation are discounted at a specified interest
rate and include the effect of taxes on the NPV.
4.5.7
Solution Procedure
In order to obtain the Pareto-optimal curve for a bicriterion optimization problem,

one of the objectives is specified as an inequality with a fixed value for the bound,
which is treated as a parameter. There are two major approaches to solve the prob-
lem in terms of this parameter. One is to simply solve it for a specified number
of points to obtain an approximation of the Pareto-optimal curve. The other is to
solve the problem as a parametric programming problem [29], which yields the
exact solution for the Pareto-optimal curve. While the latter provides a rigorous so-
lution approach, the former one is simpler to implement. For this reason we have
selected this approach. The procedure includes the following three steps: The first
one is to minimize the lead time TP to obtain the shortest lead time TPS , which in
turn yields the lowest Pareto-optimal NP V . The second step is to maximize NP V
that in turn yields the longest Pareto-optimal lead time TPL . In this case, the objec-
tive function is set as NP V − ε · TP(ε is a very small value, e.g. it is on the order
of 0.001). The last step is to fix the lead time TP to discrete values between TPS
and TPL , and to optimize the model by maximizing NP V at each selected point.
In this way, we can obtain an approximation to the Pareto-optimal curve, together
with the optimal configurations of PSCN for different values of lead time.
4.6
Numerical Example
In this section, we consider the PSCN for producing polystyrene resins in the mo-
tivating example (as shown in Fig. 4.1). There are four suppliers, four customers,
three manufacturing sites, and three types of processes. Process I is dedicated;
Processes II and III are flexible multiproduct plants. The lower bounds of the de-
mands are set to zero, which implies that the model can select which customer
markets to include in the optimal supply chain network. The time horizon is 10
years, and 3 time periods are considered, with length of 2 years, 3 years, and 5
years, respectively. An annual interest rate of 5% has been considered for the cal-
culation of the NPV.
The model consists of 133 binary variables, 2249 continuous variables, and 3041
equations. It was modeled with GAMS [30]. Due to the nonconvexity of the MINLP,
it was solved using the BARON solver [31] on an Intel 3.2 GHz machine with 512-
MB memory. Nine points in the Pareto-optimal curve take 133 h computational
time for global optimum.
The Pareto curve is shown in Fig. 4.18, which required 133 h of computation.
From this figure, it can be seen that the Pareto curve ranges from 8.85 to 14.4 days
in the lead time, and from $159 million to $1261 million for the NPV. It is interest-
ing to see that in the range between 11.6 and 13.7 days, the NPV remains constant
at $794 million, which means that the lead time of 11.6 days can be obtained over
a range with no penalty to the NPV.
The optimal network structures under different lead times are shown in Figs.
4.19–4.22. It is interesting to see that all the three sites are selected, but different
types of plants are installed in all the network structures. With the shortest lead
time, 8.85 days (Fig. 4.19), only the customer markets in NY and GA are selected to
sell the products, four plants in the three sites are installed, and all the four suppli-
4.6 Numerical Example 127
d)
c)
b)
a)
Fig. 4.18 Pareto-optimal curve for the numerical example.
Plant Site PA
Ethylene 5.12
OH 0.5 SPS
1 Styrene II CA
I 1
1 Benzene
VA
NY
TX 0.5
GA
Ethylene 1
1 1 5.35
Styrene
I III
Styrene
0.5 EPS
TX Benzene MN
Fig. 4.19 Optimal network structure at lead time = 8.85 days.
ers are selected and connected to the associated nearest plant sites. As the lead time
increases to 9.55 days, the customer market in MN can be linked to the AL manu-
facturing site, which leads to larger sales, in turn increase the NPV (Fig. 4.20). As
shown in Fig. 4.21, Plant I in PA site and the links between suppliers to PA site are
removed, while a new intersite link from TX site to PA site for the transportation of
styrene monomer is added. The change of network structure reduces the total in-
vestment cost, but also increases the lead time. With the longest lead time of 14.42
days, all the four customer markets are connected to the plant sites, and this struc-
Plant Site PA
Ethylene 5.12
OH 0.5 SPS
1 Styrene II CA
I
1
1 Benzene
VA
NY
TX 0.5
Ethylene 1 GA
1 1
5.05
Styrene
I III 2
Styrene
0.5 EPS
TX Benzene MN
Plant Site PA
6.64
OH SPS
Styrene II CA
VA
NY
2.5
TX 0.5
Ethylene 1 GA
1 1 5.65
Styrene
I III 2
0.5 Styrene
Benzene EPS MN
TX
ture leads to the highest NPV up to $1.2 billon (Figs. 4.22 and 4.23). This example
then shows the importance of establishing trade-offs between responsiveness and
economics in the design and planning of a PSCN for the improvement of overall
earning and performance of a company.
4.7
Conclusions
This chapter has presented a quantitative approach for designing responsive sup-
ply chain. Lead time under the assumption of zero inventories was proposed as
4.7 Conclusions 129
Plant Site PA
6.42
OH SPS 4
Styrene II CA
1
VA
NY
2.5
TX 0.5
Ethylene GA
1 1 1
5.14
Styrene
I III 2
0.5 Styrene
Benzene EPS MN
TX
Fig. 4.23 Material flows in the location map for longest lead time (14.42 days) case.
a measure of responsiveness, and was defined quantitatively with integration of

supply chain network structure and operational details. A MINLP mathematical
model was developed to support the decision making in design and planning of
PSCNs. The model integrates the long-term strategic decisions (e.g., installation of
plants, selection of suppliers, manufacturing sites, and transportation links) with
the short-term operational decisions (e.g., product changeovers and inventory pro-
files for flexible processes) for the multisite multiechelon PSCN.
A bicriterion optimization framework was implemented to obtain the trade-offs
between responsiveness and economics in this model. The multiobjective opti-
mization model was solved with the ε-constraint method for the sake of compu-
tational simplicity. An example reflected to styrene production was solved to illus-
trate the industrial application of this model. The result shows that small changes
in lead time can lead to significant changes in the NPV and the network struc-
ture, which in turn suggests the importance of integrating responsiveness into the
design and operations of PSCN.
In terms of the limitations of this model, we have assumed that all the trans-
portation links are continuous and that there are no capacity expansions over the
whole time horizon. Our next aim is to further develop this model so that it can
accurately capture all the aspects of the PSCN, such as capacity expansion and the
inventory effect of single product plants due to intermittent transportation. Uncer-
tainties in the demands and prices of chemicals should also be taken into account.
Finally, a more efficient algorithm to solve this type of nonconvex MINLP problem
also needs to be developed.
Acknowledgments
The authors gratefully acknowledge financial support from the National Science
Foundation under Grant No. DMI-0556090.
References
1 Stalk, G., Time – the next source of Indust. Market Manage. 29 (2000), pp.
competitive advantage. Harv. Bus. Rev. 37–44.
66(4) (1988), pp. 41–51. 7 Grossmann, I. E., Enterprise-wide op-
2 Christopher, M., Logistics and Sup- timization: A new frontier in process
ply Chain Management, Prentice-Hall, systems engineering. AICHE J. 51
Englewood Cliffs, NJ, 2005. (2005), pp. 1846–1857.
3 Sabath, R., Volatile demand calls for 8 Forrester, J., Industrial Dynamics.
quick response. Int. J. Phys. Distrib. MIT Press, and John Wiley & Sons,
Logist. Manage. 28(9/10) (1998), pp. New York, 1961.
698–704. 9 Lee, H. L., Padmanabhan, V.,
Whang, S., Information distortion
4 Christopher, M., Towill, D., An in-
in a supply chain: The bullwhip ef-
tegrated model for the design of agile fect. Management Science 43(4) (1997),
supply chains. Int. J. Phys. Distrib. Lo-
pp. 546–558.
gist. Manage. 31 (2001), pp. 235–246.
10 Matson, J. B., McFarlane, D. C., As-
5 Fisher, M. L., What is the right sup- sessing the responsiveness of existing
ply chain for your product? Harv. Bus. production operations. Int. J. Oper.
Rev. 75(2) (1997), pp. 105–116. Prod. Manage. 19(8) (1999), pp. 765–
6 Christopher, M., The agile supply 784.
chain competing in volatile markets. 11 Yusuf, Y. Y., Gunasekaran, A., Adel-
eye, E. O., Sivayoganathan, K., Agile
References 131
supply chain capabilities: Determi- 21 Kallrath, J., Combined strategic and

nants of competitive objectives. Eu- operational planning: An MILP suc-
ropean J. Oper. Res. 159 (2004), pp. cess story in chemical industry. OR
379–392. Spectrum 24 (2002), p. 315.
12 Holweg, M., The three dimensions 22 Chen, C. L., Wang, B. W., Lee, W. C.,
of responsiveness. Int. J. Oper. Prod. Multiobjective optimization for a mul-
Manage. 25 (2005), pp. 603–622. tienterprise supply chain network.
Indust. Eng. Chem. Res. 42 (2003), pp.
13 Shaw, N. E., Burgess, T. F., Mattos,
1879–1889.
C., Stec, L. Z., Supply chain agility:
The influence of industry culture on 23 Halshall-Whitney, H., Thibault,
asset capabilities within capital inten- J., Multi-objective optimization for
sive industries. Int. J. Prod. Res. 43(16) chemical processes and controller
(2005), pp. 3497–3516. design: Approximating and classify-
ing the pareto domain. Computer and
14 Kok, A. G., Graves, S. C., Supply
Chemical Engineering 30 (2006), pp.
Chain Management: Design, Coor-
1155–1168.
dination and Operation, in: Handbooks
24 Raman, R., Grossmann, I. E., Sym-
in Operations Research and Manage-
bolic integration of logic in mixed in-
ment Science, Elsevier, Amsterdam,
teger linear programming techniques
2003.
for process synthesis. Computer and
15 Shah, N., Process industry supply Chemical Engineering 27 (1993), p. 909.
chains: Advances and challenges.
25 Norton, L. C., Grossmann, I. E.,
Strategic planning model for complete
1225–1235.
process flexibility. Ind. Eng. Chem. Res.
16 Tsiakis, P., Shah, N., Pantelides, 33 (1994), pp. 69–76.
C. C., Design of multi-echelon supply 26 Pinto, J. M., Grossmann, I. E., Op-
chain networks under demand uncer- timal cyclic scheduling of multistage
tainty. Ind. Eng. Chem. Res. 40 (2001), continuous multiproduct plants. Com-
pp. 3585–3604. put. Chem. Eng. 18 (1994), pp. 797–
17 Jackson, J. R., Grossmann, I. E., 816.
Temporal decomposition scheme 27 Wolsey, L. A., MIP modeling of
for nonlinear multisite production changeovers in production planning
planning and distribution models. and scheduling problems. Eur. J. Oper.
Indus. Eng. Chem. Res. 42 (2003), pp. Res. 99 (1997), pp. 154–165.
3045–3055. 28 Balas, E., Disjunctive programming
18 Schulz, E. P., Diaz, M. S., Bandoni, and a hierarchy of relaxations for dis-
J. A., Supply chain optimization of crete continuous optimization prob-
large-scale continuous processes. lems. SIAM J. Algebr. Discrete Methods
Comput. Chem. Eng. 29 (2005), pp. 6(3) (1985), pp. 466–486.
1305–1316. 29 Dua, V., Pistikopoulos, E. N., Para-
19 Sousa, R. T., Shah, N., Papageorgiou, metric optimization in process sys-
L. G., Supply chain design and plan- tems engineering: Theory and algo-
ning with responsiveness testing – a rithms. Proc. Indian Nat. Sci. Acad.
two-level holistic approach to an in- 69A (2004), pp. 429–444.
dustrial case. AICHE Annual Meeting, 30 Brooke, A., Kendrick, D., Meeraus,
San Francisco, CA, 2006. A., Raman, R., GAMS – A User’s Man-
20 Bok, J. K., Grossmann, I. E., Park, ual, GAMS Development, Washington
S., Supply chain optimization in con- DC, 1998.
tinuous flexible process networks. 31 Sahinidis, N. V., BARON: A general
Indust. Eng. Chem. Res. 39 (2000), pp. purpose global optimization software
1279–1290. package. J. Glob. Opt. 8(2) (1996),
pp. 201–205.
Appendix
Nomenclature
Indexes/Sets
k Manufacturing sites
i Processes
j Chemicals
ls Suppliers
ld Customers
l Markets (suppliers and customers)
t Time periods
s, s Production schemes
Ki Set of sites that can set up process i
Ji,s Set of chemicals involved in scheme s of process i
JPi,s Set of main products for production scheme s of process i
JPk,i Set of main products of process i in site k
JMk,i Set of materials of process i in site k
Si Set of production schemes for process i
SLi Set of time slot for process i in the production scheduling
LSj Set of suppliers that supply chemical j
LDj Set of customers that need chemical j
Ij Set of processes that consume chemical j
Oj Set of processes that produce chemical j
Pathls,k,l Set of possible path of chemical flow from a supplier to some
sites and then ends at a customer. Elements are in the form of
(ls, k1 , k2 . . . kn , ld)
Parameters
Lenpt Length of each time period t
L
dj,ld,t Lower bound of demand of chemical j in market ld during time pe-
riod t
U
dj,ld,t Upper bound of demand of chemical j in market ld during time pe-
riod t
ϕj,ld,t Price of chemical j in market l during time period t
εj Inventory cost of chemical j
R
θi,s Residence time of the main product for production scheme s of
process i
O
γk,l Setup cost of pipelines from suppliers to plant sites or from plant sites
to customers
I
γk,k Fixed cost of interplant site transportation

P
γk,i Fixed cost of installation of each plant
O
ωk,j,l,t Variable transportation cost of chemicals between markets l and sites k
I
ωk,k Unit transportation cost of chemicals for interplant site transportation
,j,t
P
ωk,i Variable cost of installation of each plant
Appendix 133
λOk,l Transportation time from between markets and plant sites

λIk,k Transportation time from one plant site to the other
ρi,s Relative maximum production rate of main product in scheme s of
process i
µi,j,s Material balance coefficients of chemical j in scheme s of process i
σi,s,t Unit operating cost of scheme s of process i during period t
τi,s,s Transition time from production scheme s to s in process i
CTRi,s,s Transition cost from production scheme s to s in process i
Hk,i,t Total production time in plant i of site k in period t
U
Pk,j,ls,t Upper bound of purchase of chemical j from supplier ls to site k dur-
ing period t
TRUk,k ,j,t Upper bound of transportation of product j from site k to k in period t
U
Sk,j,ld,t Upper bound of sales of product j to market ld from site k during
period t
U
Wk,i,j,s,t Upper bound of production of chemical j in plant i of site k in period t
QU
k,i Upper bound of capacity of each plant
Capacity of each plant
Continuous
Variables
Qk,i
rk,i,s Processing rate of production scheme s in plant i of site k
Wk,i,j,s,t Amount of chemical j produced in plant i of site k in period t
Pk,j,ls,t Purchase of chemical j from supplier ls to site k during period t
Sk,j,ld,t Sales of product j to market ld from site k during period t
TRk,k ,j,t Transportation of product j from site k to k in period t
TP Total lead time of the whole supply chain network
NP V Net present value of the supply chain network
θk,i Time delay by production of process i in site k
Tsk,i,sl,t Starting time of slot sl in process i of site k in period t
Tek,i,sl,t End time of slot sl in process i of site k in period t
δk,i,s,sl,t Processing time of scheme s in slot sl of process i in site k
δk,i,sl,t Processing time of the time slot sl of process i in site k
TCk,i,t Cycle time of process i in site k in period t
S
Wk,i,s,t Amount produced of main product in scheme s of process i of site k
in period t
Nk,i,t Number of cycle in plant i of site k in period t
COSTSk,i,t Total cost for inventories and transitions of process i in site k in pe-
riod t
Binary Variables
P
Yk,i 1 if plant i in site k is installed
O
Yk,l 1 if transportation link from site k to market l is set up
I
Yk,k 1 if transportation link from site k to k is set up
S
Yk,i,s,sl,t 1 if the slot sl is assigned to the production scheme s in process i of
site k in period t
Zk,i,s,s ,sl,t 1 if production scheme s is preceded by production scheme s in time
slot sl of process i in site k in period t
135
5
Reverse Production Systems – Optimization Modeling to
Support Supply Chains for Product Recovery
Tiravat Assavapokee, Matthew J. Realff, Jane C. Ammons
5.1
Introduction
Industrial systems have evolved over the last 150 years to deliver a vast array of
goods from global sources of production to globally distributed consumers. For-
ward production systems have taken advantage of the historical development of
mass transportation from canals through steam ships and railways, motorway and
interstate systems, and mass air freight. These transportation developments in turn
lead to increased concentration and flexible location of mass production systems,
Chandler [1]. In many cases, highly optimized, sophisticated, production systems
for delivering products to consumers are paired with remarkably primitive, unso-
phisticated, means for recovering products from them. Some effective reverse pro-
duction systems do exist, for basic materials, such as steel and paper, and for some
products that contain these materials, but these are the exceptions rather than the
rule.
In the EU, legislation has been necessary to try to raise the effectiveness of re-
cycling systems for cars, electronics, packaging, paper, and more generally high
thermal content materials. In the US, such legislation has been difficult to develop
within the Federal system for many reasons. The motivation of limited landfill
space is not as compelling in a relatively lightly populated country, and the general
resistance to legislation that forces consumers and businesses to adopt possibly
costly changes is high. Therefore, in the US, there has been a higher motivation
to develop systems that require little or no government intervention, either in the
form of subsidies or taxes, or explicit landfill bans, or mandated take-back targets.
In this chapter, we will describe supply chain modeling efforts that have been un-
dertaken to support the development of these systems based on the minimization
of cost.
We will describe a supply-chain model that has been developed for reverse pro-
duction system design for carpet recycling. The model has some important fea-
tures, and questions asked of it that are different from forward systems. In particu-
lar, there are key uncertainties that must be addressed in the schemes for collecting
ISBN: 978-3-527-31693-9
136 5 Reverse Production Systems – Optimization Modeling to Support Supply Chains
carpet at the early stages of the system planning, in order to achieve the necessary
scale for economic operation. In Section 5.2, we will briefly review the literature
in the recycling system planning area. In Section 5.3, we will introduce key facts
about carpet and carpet recycling that drives the modeling effort. In Section 5.4, we
will describe the basic model and the assumptions that drive it. In Section 5.5, the
case study will be developed, and in Section 5.6, we will present case study results
and discussion and finally conclude.
5.2
Literature Review of Reverse Product System Modeling
Product recovery systems have been an area of significant activity in the last decade
and a half. Fleischmann et al. [2], Fleischmann et al. [3], and Bostel et al. [4] give
systematic overviews of the logistics network of reuse and recycling. Other research
papers have focused on recycling and resource recovery for specific materials,
such as printer toner cartridge [5–7] and mobile phones [8]. Several demonstra-
tion projects in the US have been conducted to study the management of end-of-
life electronic products [9–11]. A number of frameworks for reverse supply chain
management are also provided in [12–15].
The use of mathematical models in strategic level to allocate product recovery fa-
cilities in reverse logistics system is described in a number of studies for different
types of discarded products. Ammons, Realff, Newton, and Cerav [16] present com-
mon features of the reverse production system and develop a generic MILP model
to determine the infrastructure of the recovery system. Pochampally and Gupta [17]
develop a three-phase mathematical programming approach to design a reverse lo-
gistics network. Phase one is to identify the most economical product from a set
of used products. Then, potential facilities are identified in phase two from a set
of candidate recovery facilities by using analytic hierarchy process. Phase three is
to solve a MILP model to obtain location of facilities and product flows across the
established recovery network.
For electronic products, Ammons et al. [18] and Hong et al. [19] present mathe-
matical programming tools to support robust strategic decision making in product
recovery system under uncertainty. The uncertainty is considered in terms of costs,
prices, and volumes in the reverse production system. Other researches focusing
on electronic products include Jayaraman, Guide, and Srivastava [20], Jayaraman
et al. [21], and Krikke et al. [22]. Kusumastuti et al. [23] provide an approach to
design reverse logistics networks, incorporating multiobjective functions, multi-
period planning horizon, and uncertainties. A case study of computer refurbish-
ment is presented. The authors introduce a genetic algorithm to find solutions
and employ a simulation model to study the impact of uncertainty. Shih [24] stud-
ies a reverse logistics system planning for electrical appliances and computers in
northern Taiwan through an MILP model to determine the optimal solution for
infrastructure design and reverse flow.
5.3 Carpet and Carpet Recycling 137
For nonelectronic products, there are a number of studies extending back to the
early nineties. These studies compliment earlier work in the planning of waste
hauling and location problems, as an example Caruso et al. [25] develop a multi-
objective mathematical model of location allocation for planning urban solid waste
management systems. The studies in this area reflect products that are of public
concern either due to their widespread use and rapid disposal or because they are
potentially toxic in landfills. Wang et al. [26] formulate an MILP model for reverse
logistics of recovered paper in the State of Iowa. Kroon and Vrijens [27] study a
case of a large reverse logistics organization service for returnable containers in the
Netherlands. Louwers et al. [28] present a linear programming model encounter-
ing facility location allocation problem for carpet recycling. Realff et al. [29] develop
an MILP model to aid in decision making at the strategic level for carpet recy-
cling infrastructure design. Realff et al. [30] further develop a framework of robust
optimization for reverse production system, based on their earlier MILP model.
Schultmann et al. [31] propose a MILP model to solve a facility location problem.
The authors apply the model to a case study of recycling networks for spent batter-
ies in Germany. Srivastava et al. [32] consider the reuse of traditionally single use
devices in the healthcare industry.
In addition to studies focused on consumer products, there has also been in-
terest in industrial wastes and materials. Spengler et al. [33] develop MILP model
to deal with location and allocation planning of recycling installations of the steel
industry in Germany. Barros, Dekker, and Scholten [34] formulated an MILP prob-
lem as a mathematical model for the sand-recycling problem in the Netherlands.
Listes and Dekker [35] extend this formulation, presenting a stochastic program-
ming approach to account for uncertainties.
Recently, the structure of recovery processes in closed-loop supply chains has
been studied to understand their impact on profits, Savaskan et al. [36]. In this
work, different structures of coordination between manufacturer and retailer are
examined, where the manufacturer can collect directly, subcontract to a third party,
or provide incentives to a retailer.
There has also been significant activity on planning and optimizing the disas-
sembly strategy for used products to balance the amount of effort put into the
recovery process and the amount of value that is recovered from the operations.
Research on disassembly strategy using mathematical model or MILP model for
specific materials with recyclables focusing on cost benefit, environment, selling
price, recyclers include Krikke et al. [22], Chen et al. [37], Das and Yedlarajiah [38],
and Nagurney and Toyasaki [39].
5.3
Carpet and Carpet Recycling
The focus of the modeling case study is carpet recycling in the United States. This
is a product for which there has been extensive activity in the last few years with
regards to recycling; with the invention of several processes for recovering useful
Table 5.1 Carpet recycling face fiber type assumption.
Face fiber type % of Stream
Nylon 6 30
Nylon 66 35
Polypropylene 20
PET and other 15
materials. A key issue in carpet recycling is the recovery and classification of the
carpet from the consumer. Typically, this can be facilitated through the installer
who would return the carpet to the retail site after it is removed from the building,
or directly to a recycling facility, thus avoiding an expensive curb-side recycling
system.
Carpet is composed of two main subsystems, the face and the backing. Carpet has
many grades and uses several different fiber types for the visible face fiber, and has
several different backing constructions for the underlying physical support of the
face fiber. A typical carpet has approximately 50% of its weight in the face fiber and
the rest in the backing. The backing is typically one layer of a polypropylene fabric
through which the face fiber is either tufted or woven, a layer of “glue” that is latex
filled with calcium carbonate, and a final fabric layer. A reasonable characterization
of the composition of the recycling stream in terms of the different face fibers is
given in Table 5.1.
A key point is that the latex and calcium carbonate filler can be as much as 40–
50% of the weight of the carpet and yet is not a valuable material from a recycling
perspective. A second issue is that the different face fiber polymers have different
recycling uses and values. This means that the carpet must be sorted into different
fiber types at some point in its journey from the consumer to its recycling destina-
tion. Finally, carpet is a heavy and bulky material that is not easy to handle when
loose and it is not compressible. Shipping loose carpet is significantly more costly
because of the lack of density than baling it. Shipping can be a considerable ex-
pense and it is important to decide what transportation mode to use and at what
point to bale the carpet for easier handling, if at all. This interacts with the des-
tinations of the material, because some face fiber types may not be required by
the recycler and so the decision must be made whether to ship and sort centrally,
and hence incur extra shipping cost, or to sort and ship only the material that is
needed by the recycler. These questions are the focus on the mathematical model
presented in the next section.
5.4
Model Description
To answer the questions of the location of processes and which collection sites
to open to meet volume targets, we formulate this as a mixed integer linear pro-
gram (MILP). The complete MILP model formulation is given at the end of this
section and the objective function and constraints of the formulation will be de-
scribed briefly. The model can represent the operation of the system on two dif-
ferent timescales. The larger, in our case yearly, time period controls the strategic
growth of the network by requiring that certain minimum demands for the re-
quired material are met. In order to meet these demands, the network is allowed
to grow on a finer, monthly timescale, and to vary its operation monthly to reflect
constraints with respect to shipments and labor.
The decisions that are made on a monthly basis are which sites to open and close,
although in the case study the system is in an expansion mode and so shutting
sites down is not considered, and materials management at each site. The materials
management takes two forms. There are discrete types of material that are collected
from the sources, and each type of material can be shipped or stored or processed.
The shipping can be done either directly to the customer, or to another site that
acts as a hub for the material so that it can be aggregated for further shipment. The
hubs themselves can ship material to other hubs for yet further aggregation. The
collection and hub locations have certain transportation modes available to them.
Each material type can be further processed into types that are demanded by
the eventual customer. The main process capacity is discrete and involves different
capital and operating cost depending on the process. Each process can perform a
number of subprocesses that must fit within the installed capacity. This reflects the
idea that the different material types may consume a fraction capacity of a certain
process.
5.4.1
Objective Function
The objective function of this model is to minimize the overall cost of the supply
chains infrastructure, given the constraint of reaching certain collection volumes.
For example, collecting, processing, and transporting old carpets in order to obtain
a certain quantity of nylon face carpet at a specific set of locations for each specific
time period.
Constraint (1): (Material Flow Balance Constraints)

The material flow balance constraints are used to balance the physical flow of
each commodity at each physical location in the supply chain during each time
period. It ensures the conservation of commodities within a network by balanc-
ing the flow dynamics of those commodities. These constraints state that the total
amount of each commodity transported out, used in production, and stored at each
location during each time period must have been transported in from other loca-
tions, produced at the location, or stored at the location since previous time period.
These constraints were articulated in the classical state-task-network formulations
in chemical batch process scheduling.
Constraint (2)–(14): (Capacity Constraints)

The capacity constraints are introduced in the model in order to specify the capacity
limitation on collection, supply, production, storage, and transportation at each site
and each connection in the supply-chains network.
Constraints (2) and (9) enforce the limitation of material collection based on the
number of available vehicles. Constraints (3) and (10) enforce the limitation of ma-
terial collection based on the number of available shifts. Constraint (4) enforces the
limitation of material collection based on the available supply. Constraints (5) and
(11) enforce the limitation of material processing based on the number of produc-
tion resources available. Constraints (6) and (12) enforce the limitation of material
storage based on the available storage space. Constraints (7), (8), (13), and (14) en-
force the limitation of material transportation based on the number of available
vehicles.
Constraint (15): (Minimum Quantity of Material Constraints)

This constraint enforces the decision to obtain a certain minimum quantity of a
certain type of material at a certain set of locations for each certain time period.
Constraint (16)–(22): (Binary Logical Constraints)

These constraints enforce the decision to satisfy a set of logical requirements. For
example, in order to perform some operations (collection, production, storage,
transportation) at a site, this site must be opened and operated.
Constraint (23)–(24): (Binary Logic for Site Opening and Closing)

These constraints enforce the decision to correctly account for opening and closing
sites. The closing of sites is considered only in longer term problems where the net-
work may be initially expanded based on the willingness of collectors to participate,
and then reconfigured as additional collectors become available for incorporation
into the network.
Constraint (25)–(30): (Must and Allow Constraints)

These constraints enforce the decision to satisfy some conditions required by the
decision maker. For example, the decision maker may want to enforce the model
to always use some existing sites.
Constraint (31)–(33): (Nonnegativity and Integer Restriction Constraints)

These constraints are common in all MILP models enforcing some variables to be
non-negative, binary, or integer values.
5.4.2
Mathematical Formulation
For model indices, model superscripts, parameters and variables please see appen-
dix to this chapter.
Mathematical model
Minimize (Objective) Minimize Total Cost

(St) (St) (Co) (Co) (Co) (Co)
Fit yit + Fit yit + Fimt yimt
t i t i t m i
(Si) (Si) (Si) (Si) (Si) (Si)
+ (Fit yit + Fit yit + Fit yit )
t i
(Pr) (Pr)
+ Fipt yipt
t p i
(Tr) (Tr)
+ Fii mt yii mt
t m i i =i
(Tr) (Tr)
+ Ficmt yicmt Fixed Costs
t m c
i
(St) (St)
+ Vij t xij t Storage Costs
t j i
(Co) (Co)
+ Vij mt xij mt Collection Costs and Fees
t j m i
(Pr) (Pr)
+ Vi(sp)t xi(sp)t Processing Costs
t sp
i (Pr) (Tr)
+ Vjj ∗ct xij cmt Additional Processing Costs
i j =j ∗ c m t
(Tr) (Tr)
+ Vii mt xij i mt dii m
t m i j i =i
(Tr) (Tr)
+ Vicmt xij cmt dicm Shipping Costs
t m i j c
Subject to:
(St) (St)
(Co)
(Tr)
(Tr)
xij t = xij (t−1) + xij mt + xi j imt − xij cmt
m i =i m c m

ρj (sp) xi(sp)t
(Tr) (Pr) (Pr)
− xij i mt + − ρj (sp) xi(sp)t ∀i, j, t (5.1)
i =i m sp sp

1 (Co) (Co)
(Co)
x ij mt yimt ∀i, m, t (5.2)
(Co) Cij mt
j |Cij mt =0

1 (Co) (Co)
(Co)
xij mt yit ∀i, t (5.3)
(Co) Cij t m
j |Cij t =0
(Co) (Si)
xij mt Sij t ∀i, j, t (5.4)
m

1 (Pr) (Pr)
(Pr)
xi(sp)t yipt ∀i, p, t (5.5)
(Pr) Ci(sp)t
sp∈SPp |Ci(sp)t =0

1 (St) (St)
(St)
xij t yit ∀i, t (5.6)
(St) Cij t
j |Cij t =0

1
∀i, i = i, m, t
(Tr) (Tr)
(Tr)
xij i mt yii mt (5.7)
(Tr) Cii j mt
j |Cii j mt =0

1
(Tr)
xij(Tr) (Tr)
cmt yicmt ∀i, c, m, t (5.8)
(Tr) Cicj mt
j |Cicj mt =0
(Co) (Co)
xij mt = 0 ∀i, j, m, t|Cij mt = 0 (5.9)
(Co) (Co)
xij mt = 0 ∀i, j, t|Cij t =0 (5.10)
m
(Pr) (Pr)
xi(sp)t =0 ∀i, sp, t|Ci(sp)t = 0 (5.11)
(St) (St)
xij t =0 ∀i, j, t|Cij t =0 (5.12)
∀i, i = i, j, m, t|Cii j mt = 0
(Tr) (Tr)
xij i mt = 0 (5.13)
xij(Tr)
cmt =0
(Tr)
∀i, c, j, m, t|Cicj mt =0 (5.14)
LB(Cu)
(Tr)
Dcj ∗T Pjj ∗ xij cmt ∀c, T (5.15)
t in year T i m j
(Co) (Si)
yit yit ∀i, t (5.16)
(Co) (Co) (Co)
yimt Mimt yit ∀i, m, t (5.17)
(Pr) (Pr) (Si)
yipt Mipt yit ∀i, p, t (5.18)
(St) (Si)
yit yit ∀i, t (5.19)
(Tr) (Tr) (Si)
yii mt Mii mt yit ∀i, i , m, t (5.20)
(Tr) (Tr) (Si)
yi imt Mi imt yit ∀i, i , m, t (5.21)
(Tr) (Tr) (Si)
yicmt Micmt yit ∀i, c, m, t (5.22)
(Si) (Si) (Si)
yit − yi(t−1) yit ∀i, t (5.23)
(Si) (Si) (Si)
yi(t−1) − yit yit ∀i, t (5.24)
(Si) (Si) (Si)
mit yit ait ∀i, t (5.25)
(Co) (Co) (Co) (Co)
mit yit Mit ait ∀i, t (5.26)
(Pr) (Pr) (Pr) (Pr)
mipt yipt Mipt aipt ∀i, p, t (5.27)
(St) (St) (St)
mit yit ait ∀i, t (5.28)

m(Tr) (Tr) (Tr) (Tr)
ii mt yii mt Mii mt aii mt ∀i, i , m, t (5.29)
(Tr) (Tr) (Tr) (Tr)
micmt yicmt Micmt aicmt ∀i, c, m, t (5.30)
(Co) (St) (Tr) (Tr) (Pr)
xij mt , xij t , xij i mt , xij cmt , xi(sp)t 0 ∀sp, i, c, j, m, t, i = i (5.31)
(St) (Si) (Si) (Si)
yit , yit , yit , yit ∈ {0, 1} ∀i, t (5.32)
(Co) (Co) (Tr) (Tr) (Pr) +
yit , yimt , yii mt , yicmt , yipt ∈ I ∪ {0} ∀i, c, m, p, t, i = i (5.33)
Table 5.2 Material types, NFF is nylon face fiber carpet.
Material ID Material name
M1 Loose carpet
M2 Baled carpet
M3 NFF carpet
M4 Baled NFF carpet
M5 Landfillable material
5.5
Case Study Data and Scenarios
The model described in Section 5.4 has been applied to study the specific problem
of establishing a network of carpet collectors in the SE region of the United States to
provide the feedstock for a large recycling process that is being opened in Augusta,
Georgia. The facility needs approximately 60–100 million lbs of nylon 6 carpet on
an annual basis and we explore the collection of this from the SE region. The higher
number reflects what the facility needs as a whole and the lower number represents
the case where a significant proportion will come from outside the region. This
carpet will generate about 30–50 million lbs of nylon 6 material annually through
a large-scale chemical depolymerization process.
We use carpet sales data that we have for region in which the recycling system
will be located. The data is available by retailer in a given zip (postal) code in terms
of the number of square yards of carpet sold. A replacement rate is assumed: 2/3
of all carpet sold is for replacement. This, combined with the total sales for a given
year, in a given location, gives an estimate of the carpet volumes. For this case
study, we need to breakdown the carpet by the face fiber type using Table 5.1 since
the processor is only interested in nylon carpet. Both the nylon 6 and nylon 66
carpet will be shipped to the customer, but only the nylon 6 carpet generates nylon
6. The nylon 66 is considered to be neutral in cost once it reaches the customer.
The material types in the model are given in Table 5.2.
These materials allow the model to change the material from its loose form to
baled and to sort the nylon face fiber (NFF) carpet from the other material that
can be landfilled. The nylon face carpet can be then be baled, or not, for further
shipment.
Five processes are considered in the model, these are sorting, two different types
of baling that have different capital and throughput specifications, and unbaling
and landfilling. The emphasis around baling and unbaling enables the examination
of decisions about whether it is worth baling the carpet before shipping it, and
whether at an intermediate point the carpet should be unbaled and then sorted and
possibly rebaled. This will add handling costs but lower shipping costs.
The subprocesses are given in Table 5.3 along with the material type inputs and
outputs, the default ρ values are 1 and any others are given in brackets. This is
Table 5.3 Subprocesses in the carpet model and throughputs.
Subprocess Subprocess name Input Output materials Throughput (1000

ID materials lbs/shift/month)
MP1-P1 Sorting Loose carpet Landfillable (0.35) 1498

NFF carpet (0.65)
MP2-P2 Horizontal baling Loose carpet Baled carpet 1920
loose carpet
MP2-P3 Horizontal baling NFF carpet Baled NFF carpet 1,920
NFF carpet
MP3-P4 Vertical baling loose Loose carpet Baled carpet 256
carpet
MP3-P5 Vertical baling NFF carpet Baled NFF carpet 256
NFF carpet
MP4-P6 Unbaling loose carpet Baled carpet Loose carpet 2304
MP4-P7 Unbaling NFF carpet Baled NFF NFF carpet 2304
carpet
MP5-P8 Landfilling Landfillable Landfilled material 2304
Table 5.4 Dimensions for scenarios.
Model scenario dimension High setting Low setting
Volume required (VR) 50 million lbs/year 30 million lbs/year

Transportation cost (TC) US$ 4/load/mile US$ 2/load/mile
Participation rate (PR) 60% 40%
a very simple process system but could lead to a web of processes based on the
transshipment of material from collectors to hubs.
A total of 200 collection sites were considered as sources for the material and
100 of these were considered as possible hubs for locating the process equipment
and for transportation consolidation. Eight possible scenarios were considered for
the model from the factor of three dimensions each at low and high settings; these
are defined in Table 5.4. First the volume required (VR) is the amount the customer
needs to satisfy their process from this region. The second is the transportation cost
(TC) for trucking the material, which has been increasing sharply with fuel prices,
and is expressed as how much it costs to ship a full load (44,000 lbs in the US) for
1 mile. The third is the participation rate (PR) of the retailers in the network. Not
all the retailers will want to participate in the recycling scheme and so the average
distance that material will have to be shipped will increase as the participation rate
decreases.
The eight scenarios that are formed by the combinations were examined with
and without the ability to have rail links to the final customer from the hubs. Cost
data used in the case study is given in Table 5.5 and other data in Table 5.6.
Table 5.5 Cost data for case studies.
Fixed cost Variable cost Basis unit

US$/basis unit
Hub site (labor) 16,800 (US$/month/shift) 0.002 (L), 0.004 (H) lbs
Hub site (rental) 5000 (US$/month)
Truck 710 2 (L) load/mile

(US$/truck/month) 4 (H)
Rail 3570 0.09 m+1.67 m = miles
(US$/rail/month) 0.18 m+1.67 basis per ton
Sorting 2970 0.005 Lbs
(US$/equipment/month)
Horizontal baling 4900 0.005 Lbs
Vertical baling 4867 0.005 Lbs
Unbaling 2400 0.005 Lbs
(US$/resource/month)
Landfilling 2400 33 (L) Per ton
(US$/resource/month) 46 (H)
Table 5.6 Other main data for case study.
Data category Value
Truck capacity unbaled 30,000 lbs/load

Truck capacity baled 44,000 lbs/load
Rail capacity unbaled 140,000 lbs/load
Rail capacity baled 200,000 lbs/load
The landfill cost is quite low, but the SE region does not have high tipping fees
in general, and we have assumed a long travel distance to allow for flexibility. Since
the material is of one type, it can be segregated and landfilled separately which can
also reduce the cost.
The model had 32,500 binary, 212,861 continuous variables, and 125,862 con-
straints. It was solved on a Windows XP-based Pentium® 4 CPU 3.60 GHz per-
sonal computer with 2.00 GB RAM using a C++ program and CPLEX 9 for the
optimization process. MS-Access is used for the case study input and output data-
base. Visual Basic 2005 is used to query required data for the case study. Each model
run took approximately 3.5 h and the tolerance on the gap between the lower and
upper bound was set to 5%.
Table 5.7 Results for scenarios when rail is allowed from the hubs.
Scenario VR TC PR Objective function Hubs Processes N6 Face

(US$ cost/month) Sorting Landfilling 1000 lbs/year
1 L L L US$ 350,000 2 2,2 1,1 30200

2 L L H US$ 295,000 2 2,2 1,1 30200
3 L H L US$ 420,000 3 2,2 1,1 30200
4 L H H US$ 350,000 3 0,2,1 0,1,1 30150
5 H L L US$ 685,000 4 0,3,0,0 0,1,0,0 50130
6 H L H US$ 540,000 4 2,3,3 1,1,1 50380
7 H H L US$ 820,000 5 0,3,2,1 0,1,1,1 50300
8 H H H US$ 625,000 5 0,2,2,3,1 0,1,1,1,1 50390
Table 5.8 Transportation results when rail is allowed from the hubs.
Scenario Collection Collection T2 T3

sites trucks Trucks/rails Trucks/rails
1 24 26 12 12/2
2 16 18 6 10/2
3 27 30 14 13/3
4 15 18 6 9/3
5 80 84 49/1 34/4
6 31 35 16/1 15/4
7 82 87 61/1 21/4
8 30 35 18/1 12/4
5.6
Results and Discussion
The numerical results for the eight scenarios are given in Tables 5.7–5.10 for the
cases with and without the use of rail from the hubs. The results show that it is
desirable to use hubs and rail links in all cases, but naturally the higher truck cost
and volume required encourages greater use of hubs. The use of baling was not
preferred in the rail model and as a result the processes associated with them are
zero, therefore, the only ones that are selected at hubs are sorting and landfilling
the remaining material. The list of numbers for each process type is the number
of them installed at a given hub. A zero indicates that this hub must transship the
material only, doing no sorting.
Table 5.8 shows the various transportation needs that are generated in the model.
The collection trucks are assumed to be required on the basis on one truck per
collector and per hub, which also acts as a collection point. This truck will perform
the pick-ups from the retailers themselves. The T2 information shows the number
of collection points that ship to the hubs and then the transshipment from one
hub to another via rail. For example, in scenario 7, 61 of the collection sites ship
Table 5.9 Results for scenarios when rail is not allowed from the hubs.
Scenario VR TC PR Objective function Hubs Processes N6 Face

(US$ cost/month) Sorting Horizontal Landfill 1000 lbs/year
baling
1 L L L US$ 440,000 1 3 0 1 30110

2 L L H US$ 355,000 0 0 0 0 30000
3 L H L US$ 590,000 1 2 0 1 30110
4 L H H US$ 470,000 0 0 0 0 30000
5 H L L US$ 830,000 2 3,2 0 1,1 50250
6 H L H US$ 680,000 1 3 0 1 50160
7 H H L US$ 1,125,000 3 2,2,2 1,1,0 1,1,1 50290
8 H H H US$ 910,000 2 2,2 1,1 1,1 50190
Table 5.10 Transportation results when rail is not allowed from the hubs.
Scenario Collection Collection T2 T3

sites trucks Trucks/rails Trucks/rails
1 35 36 6 33
2 31 31 0 32
3 46 47 3 46
4 43 43 0 43
5 84 84 18 66
6 47 48 4 50
7 85 87 26 67
8 56 58 8 54
to a hub, and one hub ships to another hub via rail. The T3 information shows the
number of direct connections to the customer. For example, in scenario 7, 21 of the
collection points, the balance of those that do not ship to the hubs, and four of the
hubs ship directly to the customer.
The results for the problem without the use of rail are given in Tables 5.9 and
5.10.
In these solutions, the major change is for the high transportation cost and high
volume requirement scenarios where the use of a high capacity baler is recom-
mended. The low volume scenarios all make very limited use of hubs because there
is no real advantage to the consolidation of the material, if there is no lower cost
transportation mode available from the hub. The monthly costs are considerably
higher for the no-rail solution than the rail solution, which suggests that the deci-
sion on whether or not to pursue rail hubs should be more fully explored by the
customer.
Two further case studies were considered to answer questions around expand-
ing the capacity of the network. The infrastructure for the 30 million lb of nylon 6
was assumed to have been established according to the model. This infrastructure
Table 5.11 Objective function values for models where the
network is expanded from a low volume to a high volume.
VR TC PR Objective function Objective function

(US$ cost/month)rail allowed (US$ cost/month)no rail
L→H L L US$ 700,000 US$ 830,000

L→H L H US$ 540,000 US$ 685,000
L→H H L US$ 825,000 US$ 1,125,000
L→H H H US$ 625,000 US$ 910,000
is both the collection points and the hubs. Then the necessary additional collec-
tion system for 50 million lbs of nylon 6 was solved for. The results are given in
Table 5.11 where it can be seen that the difference in cost between the optimal
network from the start at this capacity and the expanded network is small when
comparing the objective function values to the earlier tables. This means that the
customer can proceed with the lower collection target and not be overly constrained
if the amount required increases in the future.
The abstract results in the tables can also be translated into maps showing the
locations of the collection centers and hubs. Figure 5.1 shows the results for the
most optimistic and low collection volume scenario which is then expanded to the
higher volume, and Fig. 5.2 shows the results for the most pessimistic scenario
in terms of participation and transport cost. Therefore, the final networks in both
cases are comparable in the volume they are collecting.
In these figures, the blue squares are the original collection sites. The black out-
lined squares are the new capacity. The green circles are the original hubs and the
black outlined green circles are the new hubs. The two maps are quite a contrast
in both the number and dispersion of the collection points. This demonstrates the
impact of the participation rate, in particular, in the structure of the network. This
indicates that this is a very important parameter to establish with reasonable fi-
delity in the early planning phases. An interesting further optimization that is not
represented in the model would be the coordination of transportation routes be-
tween collectors back towards the hubs rather than point-to-point logistics. There
are several routes from Florida back towards Atlanta and Augusta that would sup-
port this. However, it is important to first analyze whether the results indicate that
transportation is a significant effort and cost factor.
Table 5.11 shows the transportation results for two scenarios in terms of the total
tonmiles per month and network miles that are present in the solutions. These
factors are a reasonable measure of the complexity and effort involved in the trans-
portation. These two scenarios hold the transport cost and the volume constant, so
these results are a function of the participation rate difference. This shows that the
participation rate changes the total truck ton miles and distances involved in the
network significantly, by a factor of three for the site to site transportation. This is
because the sites are more spread out at the lower participation and more effort
Fig. 5.1 Thirty million lbs growing to 50 million lbs with rail
high participation and low transport cost.
Fig. 5.2 Fifty million lbs with rail low participation and high transport cost with rail.
Table 5.12 Transport analysis for two comparable scenarios.
VR TC PR Short ton/month (ton miles) Total miles/month (miles)

Site to site Site to customer Site to site Site to customer
Truck Rail Truck Rail Truck Rail Truck Rail
H H L 370,450 316,600 101,850 146,6550 53,200 4760 15,510 21,760

H H H 113,200 383,850 96,350 1482,050 15,610 5650 13,080 21,530
Table 5.13 Cost breakdown for scenario 7.
Fixed opening cost US$ 112,500 13.75%

Fixed collection cost US$ 84,000 10.27%
Variable collection cost US$ 67,100 8.20%
Fixed processing cost US$ 25,000 3.06%
Variable processing cost US$ 95,640 11.69%
Fixed transportation cost US$ 137,700 16.83%
Variable transportation cost US$ 173,500 21.21%
Additional cost US$ 122,800 15.01%
must be made to aggregate the material before it can be shipped by high volume
rail routes.
To determine whether transportation should be a continued focus, a more de-
tailed cost analysis can be performed for any scenario. In Table 5.13, the monthly
costs are shown for scenario 7. Overall the transportation is 38% of the total, this is
for the high transportation cost, situation, and shows that this cost can be the sig-
nificant driver of recycling systems. The rest of the costs are evenly split between
the collection and the processing of the material and between the fixed costs of
establishing the network and the variable costs of operating it.
5.7
Conclusions
The design of large-scale recovery and recycling systems for electronics, carpet,
and other complex products with relatively low value at their end-of-life is still in
its early stages. There is a significant opportunity for strategic logistics planning
tools to be used to help companies and governments understand the tradeoffs
involved in establishing the infrastructure. In particular, there are uncertainties
around many aspects of the system. For example, how sources of material will re-
spond to requests to recycle, and transportation costs. Mathematical programming
models can help establish priorities on which uncertainties will have the biggest
impact, and how sensitive the infrastructure is to them. It can inform decision
makers as to where to focus their efforts in recruitment of sources of material and
the overall target costs that are achievable.
References 151
References
1 Chandler, A. D., The Visible Hand, M. M., Hetzel, C., Minnesota’s multi-
Belknap, Harvard, 1977. stakeholder approach to managing
2 Fleischmann, M., Bloemhof- electronic products at end-of-life. IEEE
Ruwaard, J. M., Dekker, R., v.d. International Symposium on Electronics
Laan, E. A., v. Nunen, J. A. E. E., van and the Environment, 2000, pp. 310–
Wassenhove, L. N., Quantitative mod- 317.
els for reverse logistics: A review. Eur. 11 Diggelman, C., Rau, G., Brachman,
J. Oper. Res. 103 (1997), pp. 1–17. S., Ziemke, B., Wisconsin end-of-life
electronics study. IEEE International
3 Fleischmann, M., Krikke, H. R.,
Symposium on Electronics and the Envi-
Dekker, R., Flapper, S. D. P., A char-
ronment, 2003, pp. 352–357.
acterization of logistics networks for
product recovery. Omega 28 (2000), 12 Bettac, E., Maas, K., Beullens, P.,
pp. 653–666. Bopp, R., RELOOP: Reverse logistics
chain optimization in a multi-user
4 Bostel, N., Dejax, P., Lu, Z., The
trading environment. IEEE Interna-
design, planning and optimization
tional Symposium on Electronics and the
of reverse logistics networks, in: Lo-
Environment, 1999, pp. 42–47.
gistics Systems: Design and Optimiza-
tion, Langevin, A., Riopel, D. (eds.), 13 Dowlatshahi, S., Developing a the-
Springer, Berlin, 2005. ory of reverse logistics. Interfaces 30(3)
(2000), pp. 143–155.
5 Bartel, T., Recycling program for
14 Knoth, R., Hoffmann, M., Kopacek,
printer toner cartridges and optical
B., Kopacek, P., A logistic concept to
photoconductors. IEEE International
improve the re-usability of electric and
Symposium on Electronics & the Envi-
electronic equipment. IEEE Interna-
ronment, 1995, pp. 225–228.
6 Williams, J., Shu, L. H., Analysis of Environment, 2001, pp. 115–118.
toner-cartridge remanufacturer waste 15 Dowlatshahi, S., A framework for
stream. IEEE International Symposium strategic factors in reverse logistics.
on Electronics and the Environment, Proceedings – Annual Meeting of the
2000, pp. 260–265. Decision Sciences Institute, 2002, pp.
7 Degher, A., HP’s worldwide take 425–430.
back and recycling programs: Lessons 16 Ammons, J. C., Realff, M. J., New-
on improving program implementa- ton, D., Cerav, S., Determining re-
tion. IEEE International Symposium on verse production systems infrastruc-
Electronics and the Environment, 2002, ture for electronic assemblies. IEEE
pp. 224–227. International Symposium on Electronics
8 Wright, L., McLaren, J., Jackson, T., and the Environment, 1999, pp. 257–
Parkinson, S., Mobile phone takeback 262.
and recycling: Analysis of the ECTEL 17 Pochampally, K. K., Gupta, S. M., A
project. IEEE International Symposium multi-phase mathematical program-
on Electronics & the Environment, 1998, ming approach to strategic planning
pp. 54–59. of an efficient reverse supply chain
9 Jung, L. B., Bartel, T., An industry network. IEEE International Sympo-
approach to consumer computer re- sium on Electronics and the Environ-
cycling: The San Jose Project. IEEE ment, 2003, pp. 72–78.
International Symposium on Electronics 18 Ammons, J. C., Assavapokee, T., New-
& the Environment, 1998, pp. 36–41. ton, D., Realff, M. J., Reverse pro-
10 Hainault, T., Smith, D. S., Cauchi, duction system design for recycling
D. J., Thompson, D. A., Fisher, under uncertainty. Proc. SPIE – Int.
Soc. Opt. Eng. 4569 (2002), pp. 1–12.
19 Hong, I., Assavapokee, T., Am- 28 Louwers, D., Kip, B. J., Peters, E.,
mons, J., Realff, M., Boelkins, C., Souren, F., Flapper, S. D. P., A fa-
Gilliam, K., Oudit, D., Vannicola, cility location allocation model for
J., Wongthatsanekorn, W., Planning reusing carpet materials. Comput. Ind.
the E-scrap reverse production sys- Eng. 36 (1999), pp. 855–869.
tem under uncertainty in the state 29 Realff, M. J., Ammons, J. C., New-
of georgia: a case study. IEEE Trans. ton, D. J., Carpet recycling: Deter-
Electron. Pack. Manuf. 29(3) (2006), pp. mining the reverse production system
150–162. design. J. Polym. Plastics Technol. Eng.
20 Jayaraman, V., Guide, V. D. R. Jr., 38(3) (1999), pp. 547–567.
Srivastava, R. A closed-loop logistics
30 Realff, M. J., Ammons, J. C., New-
model for remanufacturing. J. Oper.
ton, D. J., Strategic design of reverse
Res. Soc. 50(5) (1999), pp. 497–508.
production systems. Comput. Ind. Eng.
21 Jayaraman, V., Patterson, R. A., Rol- 24(2) (2000), pp. 991–996.
land, E., The design of reverse dis-
tribution networks: models and solu- 31 Schultmann, F., Engels, B., Rentz,
tion procedures. Eur. J. Oper. Res. 150 O., Closed-loop supply chains for
(2003), pp. 128–149. spent batteries. Interfaces 33(6) (2003),
22 Krikke, H. R., van Harten, A., pp. 57–71.
Schuur, P. C., Business case Oce: 32 Srivastava, R., Jayaraman, V., Kriche,
reverse logistic network re-design E., Reverse logistics in healthcare.
for copiers. OR Spektr. 21 (1999), pp. Proceedings – Annual Meeting of the
381–409. Decision Sciences Institute, 2003, pp.
23 Kusumastuti, R. D., Piplani, R., 2079–2084.
Lim, G. H., An approach to design 33 Spengler, T., Puchert, H.,
reverse logistics networks for prod- Penkuhn, T., Rentz, O., Environ-
uct recovery. IEEE International En- mental integrated production and
gineering Management Conference, recycling management. Eur. J. Oper.
vol. 3, Proceedings – 2004 IEEE In- Res. 97(2) (1997), pp. 308–326.
ternational Engineering Management 34 Barros, A. I., Dekker, R., Scholten,
Conference: Innovation and Entrepre-
V., A two-level network for recycling
neurship for Sustainable Development sand: A case study. Eur. J. Oper. Res.
(IEMC 2004), 18–21 October 2004,
110 (1998), pp. 199–214.
Singapore. pp. 1239–1243.
35 Listes, O., Dekker, R., A stochastic
24 Shih, L. H., Reverse logistics system
planning for recycling electrical ap- approach to a case study for product
recovery network design. Eur. J. Oper.
plicances and computers in Taiwan.
Resour. Conserv. Recycl. 32 (2001), pp. Res. 160 (2005), pp. 268–287.
55–72. 36 Savaskan, R. C., Bhattacharya, S.,
25 Caruso, C. C. A., Paruccini, M., The Van Wassenhove, L. N., Closed-loop
regional urban solid waste manage- supply chain models with product
ment system: a modelling approach. remanufacturing. Manage. Sci. 50(2)
Eur. J. Oper. Res. 70(1) (1993), pp. 16– (2004), pp. 239–252.
30. 37 Chen, R. W., Navin-Chandra Prinz,
26 Wang, C. H., Evan, J. C., Adams, F., A cost–benefit analysis model of
S. K., A mixed-integer linear model product design for recyclability and its
for optimal processing and transport application. IEEE Transactions on Com-
of secondary materials. Resour. Con- ponents, Packaging, and Manufacturing
serv. Recycl. 15 (1995), pp. 65–78. Technology Part A, 1994, pp. 502–507.
27 Kroon, L., Vrijens, G., Returnable 38 Das, S., Yedlarajiah, D., An integer
containers: an example of reverse lo- programming model for prescribing
gistics. Int. J. Phys. Distrib. Logistics
Manage. 25(2) (1995), pp. 56–68.
Appendix 153
material recovery strategies. IEEE In- Comput. Ind. Eng. 36(4) (1999), pp.
ternational Symposium on Electronics 811–853.
and the Environment, 2002, pp. 118– 45 Istvan, Z., Garamavolgyi, E., Re-
122. verse logistics and management of
39 Nagurney, A., Toyasaki, F., Reverse end-of-life electric products. IEEE In-
supply chain management and elec- ternational Symposium on Electronics
tronic waste recycling: A multitiered and the Environment, 2000, pp. 15–19.
network equilibrium framework for
46 Lu, Q., Christina, V., Stuart, J. A.,
re-cycling. Transport. Res. Part E: Lo-
Rich, T., A practical framework for
gistics Transport. Rev. 41(1) (2005), pp.
reverse supply chain. IEEE Interna-
1–28.
40 de Ron, A., Penev, K., Disassembly Environment, 2000, pp. 266–271.
and recycling of electronic consumer
products: An overview. Technovation 47 Realff, M. J., Ammons, J. C., New-
15(6) (1995), pp. 363–374. ton, D. J., Robust reverse production
system design for carpet recycling. IIE
41 Del Castillo, E., Cochran, J. K.,
Optimal short horizon distribution Trans. (Inst. Ind. Eng.) 36(8) (2004),
pp. 767–776.
operations in reusable container sys-
tems. J. Oper. Res. Soc. 47(1) (1996), 48 Sodhi, M. S., Reimer, B., Models
pp. 48–60. for recycling electronics end-of-life
42 Farkash, R., Mueller, K., The evolu- products. OR Spektrum 23 (2001), pp.
tion of environmental initiatives for 97–115.
toner containers at Xerox corporation. 49 Thierry, M. S. M., Van Nunen, J.,
IEEE International Symposium on Elec- Van Wassenhove, L., Strategic issues
tronics and the Environment, 1995, pp. in product recovery management.
267–272. Calif. Manage. Rev. 37(2) (1995), pp.
43 Guide, V. D. R., Jayaraman, V., Sri- 114–135.
vastava, R., Benton, W. C., Supply- 50 Winka, M., Carpenter, J., New Jer-
chain management for recoverable sey department of environmental
manufacturing systems. Interfaces protection/union county demanufac-
30(3) (2000), pp. 125–142. turing program. IEEE International
44 Gungor, A., Gupta, S. M., Issues in Symposium on Electronics and the Envi-
environmentally conscious manufac- ronment, 1998, pp. 328–330.
turing and product recovery: A survey.
Appendix
Nomenclature
Model Indices
i sites
j material type
j∗ required material type
m transportation mode
sp subprocess type
p main-process type
c customer
t month time period
T year time period
Model Superscripts
Co collection
St storage
Tr transportation
Pr process
Cu customer
Si site
Model Parameters
(Si)
Sij t amount of material j that is supplied at site i (if open) in month t
Pjj ∗ proportion of material j ∗ in one unit of material j
LB(Cu)
Dcj T lower bound on amount of material j that is demanded at customer c
in year T
(St)
Vij t storage cost per standard unit of material j per time period at site i in
month t
(Co)
Vij mt collection cost per standard unit of material j at site i in month t by
mode m
(Pr)
Vi(sp)t processing cost per standard unit for subprocess sp at site i in month t
transportation cost per standard unit per distance from site i to site i
(Tr)
Vii mt
using transportation mode m in month t
(Tr)
Vicmt transportation cost per standard unit per distance from site i to cus-
tomer c using transportation mode m in month t
dii m distance from site i to i by transportation mode m
dicm distance from site i to customer c by transportation mode m
(Si)
Fit fixed site operating cost if site i is opened in month t
Fit(Si) fixed site opening cost of site i in month t
(Si)
Fit fixed site closing cost of site i in month t
(St)
Fit fixed storage cost at site i in month t
(Co)
Fit fixed collecting cost site i in month t (Labor)
(Co)
Fimt fixed collecting cost at site i in month t by transportation mode m
(Vehicles)
(Pr)
Fipt fixed processing cost for using main process p at site i in month t
fixed cost for transportation from site i to site i using one transporta-
(Tr)
Fii mt
tion mode m in month t
(Tr)
Ficmt fixed cost for transportation from site i to customer c using one trans-
portation mode m in month t
(Co)
Cij t maximum collection capacity to collect material type j at site i in
month t
(Co)
Cij mt maximum collection capacity to collect material type j at site i in
month t using one transportation mode m
(St)
Cij t maximum amount of material type j that can be stored at site i in
month t
Appendix 155
maximum amount of material j that can be shipped for site i to site i

(Tr)
Cii j mt
using one transportation mode m in month t
(Tr)
Cicj mt maximum amount of material j that can be shipped for site i to cus-
tomer c using one transportation mode m in month t
(Pr)
Ci(sp)t maximum amount of material that subprocess sp can produce at site
i in month t
(Si)
ait 1 if site i is allowed to be opened in month t, 0 otherwise
ait(St) 1 if storage is allowed at site i in month t, 0 otherwise
(Tr)
aii mt 1 if shipment by transportation mode m is allowed between sites i and
i in month t, 0 otherwise
(Tr)
aicmt 1 if shipment by transportation mode m is allowed between sites i and
customer c in month t, 0 otherwise
(Pr)
aipt 1 if main process p is allowed at site i in month t, 0 otherwise
(Co)
ait 1 if collection is allowed at site i in month t, 0 otherwise
m(Si)
it 1 if site i must be opened in month t
(St)
mit 1 if storage at site i must be used in month t, 0 otherwise
(Tr)
mii mt 1 if shipment by transportation mode m must be used between sites i
and site i in month t, 0 otherwise
(Tr)
micmt 1 if shipment by transportation mode m must be used between sites i
and customer c in month t, 0 otherwise
(Pr)
mipt 1 if main process p must be used at site i in month t, 0 otherwise
(Co)
mit 1 if collection must be done at site i in month t, 0 otherwise
ρj (sp) proportion of material type j consumed by subprocess sp
ρj (sp) proportion of material type j produced by subprocess sp
SPp set of all subprocesses which can be performed by the main process p
(Co)
Mimt maximum number of transportation mode m used for collection at site
i in month t
(Pr)
Mipt maximum number of main process p at site i in month t
(Tr)
Mii mt maximum number of transportation mode m used between site i and
site i in month t
(Tr)
Micmt maximum number of transportation mode m used from site i to cus-
tomer c in month t
Mit(Co) maximum number of shift at site i in month t
cost per unit of material j for obtaining material j ∗ at customer c in
(Pr)
Vjj ∗ ct
month t
Model Variables
(Co)
xij mt amount of material collected of type j by mode m at site i in month t
(St)
xij t amount of material stored of type j at site i in month t
amount of material shipped from site i to site i of type j using trans-
(Tr)
xij i mt
portation mode m in month t
(Tr)
xij cmt amount of material shipped from site i to customer c of type j using
transportation mode m in month t
(Pr)
xi(sp)t amount of material processed by subprocess sp at site i in month t
(Co)
yit number of shifts for collection at site i in month t
(Co)
yimt number of transportation mode m used for collection at site i in
month t
(Tr)
yii mt number of transportation mode m used for the shipment between sites
i and site i in month t
(Tr)
yicmt number of transportation mode m used for the shipment between sites
i and customer c in month t
(Pr)
yipt number of main process p used at site i in month t
(St)
yit 1 if storage is to be used at site i in month t, 0 otherwise
(Si)
yit 1 if site i is decided to be opened in month t, 0 otherwise
(Si)
yit 1 if site i is decided to be closed down in month t, 0 otherwise
yit(Si) 1 if site i is operated in month t, 0 otherwise
157
6
Optimal Design of Supply Chain Networks using
Mathematical Programming
Panagiotis Tsiakis, Lazaros G. Papageorgiou, Michael C. Georgiadis
Enterprise optimization can rapidly strip significant “bottom line” costs out of oper-
ations, giving companies a real competitive edge. The benefits of managing supply
chain networks by integrating operational design and financial decisions have been
acknowledged by the industrial and academic community.
This chapter presents the application of a mathematical programming model
to two real examples where the aim is to determine the optimal configuration of
a production and distribution network subject to operational and financial con-
straints. Operational constraints include quality, production, and supply restric-
tions, and are related to the allocation of the production and the work-load balance
between production sites. Financial constraints include production costs, trans-
portation costs, and duties for the material flowing within the network subject to
exchange rates. As a business decision the outsourcing of production is considered
whenever the organization cannot satisfy the demand. A mixed integer linear pro-
gramming (MILP) model is proposed to describe the optimization problem. The
model addresses both infrastructure and operational decisions.
Two case studies are presented for two different types of industries. The first
case study focuses on a European household appliances company that wishes to
establish a new distribution center to satisfy its expanding plans in new markets.
A second case study, based on a fine chemicals business unit of a global specialty
chemicals manufacturer, is used to demonstrate the applicability of the approach
in a number of operating scenarios. Through those two distinct types of examples
the generality of the formulation is explained.
6.1
Introduction
The management of supply chains has been studied extensively in two well known
papers. First, Lee and Billington [8] present pitfalls and opportunities for managing
the supply chain inventory since, in their view, inventory levels are what the sup-
ply chain tries to control. The problems appearing in a supply chain environment
ISBN: 978-3-527-31693-9
158 6 Optimal Design of Supply Chain Networks using Mathematical Programming
are usually the outcome of the complex interactions among its constituent parts
since the management of a supply chain is very different from the management of
a single site. They identified three areas of potential difficulties: (1) pitfalls related
to information definition and supply chain management, (2) operational related
problems, and (3) strategic and design related problems. In adressing the above
pitfalls, a number of opportunities arise, especially in the areas of product design,
integration, and improvement of database systems, more robust decisions about
planning and control, and finally better metrics to characterize supply chain per-
formance. The design problems and their complexity for a distribution network as
part of the supply chain were identified by Geoffrion and Powers [9] who focused
on warehouse related issues. The issue of number and location of warehouses is
constrained by four decisions: (1) warehouse mission, (2) warehouse design, (3)
warehouse location, and (4) operational utilization. These issues are also the con-
straints in an optimization problem of this type. They discuss a comprehensive dis-
tribution planning system which is composed of the “sources” (plants, copackers,
vendors), existing and candidate “distribution centers (DCs)” (master distribution
centers, regional distribution centers, branch warehouses), and “customers” (ag-
gregations of individual customers). The work of Geoffrion and Powers [9] was a
pioneering work that introduced mathematical programming approaches to study
the design of supply chain networks.
Mathematical programming techniques began to emerge in the mid-1970s in or-
der to address the design and operation problem of supply chains and distribution
networks. In their pioneering work, Geoffrion and Graves [10] present a model to
solve the problem of designing a distribution system with optimal location of the
intermediate distribution facilities between plants and customers. Williams [25]
developed a dynamic programming algorithm for simultaneously determining the
production and distribution batch sizes at each node within a supply chain net-
work. Brown et al. [5] present an optimization-based decision algorithm for a sup-
port system used to manage complex problems involving facility selection, equip-
ment location and utilization, and manufacture and distribution of products. A
mixed integer nonlinear programming (MINLP) formulation is presented by Co-
hen and Lee [7] seeking to maximize the total after-tax profit for the manufacturing
facilities and distribution centers. Managerial (resource and production) as well
as logical consistency (feasibility, availability, demands limits) constraints were ap-
plied. Camm et al. [6], develop an integer programming model for Procter and
Gamble to determine the location of distribution centers and to assign those se-
lected to customer zones. Pirkul and Jayarama [22] studied a tri-echelon multicom-
modity system concerning production, transportation, and distribution planning.
In today’s rapidly changing economic and political conditions, global corpora-
tions face a continuous challenge to constantly evaluate and optimally configure
their supply chain operations to achieve key performance indices (KPIs), either it
is profitability, cost reduction or customer service. In this environment, operations
managers and planners need to address accurately questions such as, which plants
to operate, what product mix per plant, which distribution center supplies which
customer, what inventory levels are necessary to maintain service levels, which sup-
pliers are to be used. A global production and supply network can de defined as a set
of existing or potential manufacturing facilities, warehouses and distribution cen-
ters with multiple supply configurations, and customers with demands. All these
facilities can be located in different countries where different tax breaks and cur-
rencies apply.
The necessity to develop strategic and tactical level supply chain planning mod-
els in order to address issues in a quantitative manner rather than the qualitative
approaches used till now is acknowledged by the industry [23] and still remains
an active research area. This has created many challenges for both researchers and
practitioners who wish to implement supply chain support systems successfully. In
a recent literature survey, Bilgen and Ozkarahan analyze previous research and re-
view models for the production and distribution problem. A classification scheme
was employed to categorize and compare strengths and weaknesses as they are re-
ported in a table at the end of the review. Klose and Drexl [19] review the status
of research in facility location models for the design of distribution systems. Their
work reviews state-of-the-art model formulations on strategic planning, distribu-
tion system design, facility location, and mixed integer programming.
Previous attempts to improve the performance of supply chain networks have
mainly focused on the logistic aspects rather than the business decisions associ-
ated with the plant operation and production for the design of such networks. In
an early attempt, Arntzen et al. [1] developed an MILP “global supply chain model”
(GSCM) aiming to determine: (1) the number and location of distribution cen-
ters, (2) customer distribution center assignment, (3) number of echelons, and
(4) the product–plant assignment. The objective of the model is to minimize a
weighted combination of total cost (including production, inventory, transporta-
tion, and fixed costs) and activity days as a bill of material problem. To the same
direction and as supply chains become increasingly global, additional aspects such
as differences in tax regimes, duty drawback and avoidance, and fluctuations in ex-
change rates also become important [26]. Tsiakis et al. [24] have addressed a large
number of the production and logistics issues reported by Bilgen and Ozkarahan in
a comprehensive work that models supply chain networks under demand uncer-
tainty focusing on the production and transportation issues. The resulting MILP
problem is solved to optimality using decomposition methods to reduce the com-
putational effort required. The model reported does not include financial aspects
such as duties and exchange rates. Guinet [15] examines the economics of multi-
site production systems using a two-level approach to allocate production to sites
and address the workshop scheduling problem. A primal-dual heuristic approach
was used to solve cases of this problem aiming to minimize variable and fixed
costs. Goetschalckx et al. [12], review the area of modeling and design of global
logistics and propose two models, one for a global case where the aim is to calcu-
late the transfer prices and a second one for a single country where the focus is
on tackling efficiently seasonality in the demand. The focus is on the logistic as-
pects with the plants assumed multiproduct without limitations. Kaihara [17] uses
an agent based approach to manage supply chains in terms of product allocation
and resource distribution. The proposed model has the form of a discrete resource
allocation problem under dynamic environment. Guillen et al. [14] considered the
design and retrofit of a supply chain consisting of several production plants, ware-
houses, distribution systems, and markets. A key feature of their approach is the
consideration of uncertainty and risk into the overall supply chain design problem.
Kallrath [18] presented an MILP model for the simultaneous strategic and opera-
tional planning and applied it to a real-world problem from the chemical industry.
Salient features of the model include a detailed nonlinear price structure for the
purchase of raw materials and a detailed discussion of transport times.
The design of a supply chain network is characterized as a complex issue and
the methods used to perform such an analysis vary depending on the problem
type and the scope Ballou [2]. The author supports the idea that location analysis is
most likely sufficient for most supply chain problems. Arguably, in a continuous
changing environment and dynamic economic conditions such as costs, exchange
rates, transportation structures, and duties detail modeling is required to manage
the data available and achieve expected performance. Harrison [16] identifies the
issues related to global supply chain design and operation while he points out the
benefits of optimization. Supply chain design needs to address issues such as man-
ufacturing strategy, supplier base selection, outsourcing policies, and new product
and process design.
The key business issue of supply chain partnership was investigated by Gjer-
drum et al. [11]. In their work, a mathematical programming formulation is pre-
sented for fair, optimized profit distribution between members of multienterprise
supply chains based on a novel approach applying game theoretical Nash-type
models to find the optimal profit level for each enterprise subject to given mini-
mum profit requirements. Model decision variables include intercompany transfer
prices, production and inventory levels, resource utilization, and flows of products
between echelons, subject to a deterministic sales profile, minimum profit require-
ments for each enterprise.
Based on the work of Papageorgiou et al. [21] on the strategic design of pharma-
ceutical supply chains, Levis and Papageorgiou [20] extended the model to study the
long-term capacity planning problem under uncertainty. The problem was formed
as a mathematical programming model (MILP) to determine both product port-
folio and plant capacity taking into account the uncertainty of clinical trials to de-
velop a new product. The resulting large scale MILP problem was solved using two-
stage multiscenario hierarchical algorithm, and the applicability was demonstrated
through illustrative examples. Grossmann [13] presented an excellent overview of
the emerging area of enterprise-wide optimization for the process industries. He
highlighted some of the major research challenges in this area and he presented a
number of examples to illustrate the nature of the applications and the problems
that are faced.
The above summary of the literature review provides an insight into the cur-
rent mathematical programming techniques addressing the design and operation
problem of entire manufacturing networks and supply chains. However, to our
knowledge, none of the current literature models reviewed has considered the
combined interactions between plants, warehouses, and distribution centers in a
general framework. Most of the developed approaches consider specific problems

rather than generic formulations.
The current paper proposes a strategic planning model for multiechelon supply
chain networks, integrating components associated with production, facility loca-
tion, and distribution along side with financial and business issues such as im-
port duties, plant utilization, exchange rates, and plant maintenance. The network
comprises a number of manufacturing sites, each using a set of flexible, shared
resources for the production of a number of products. The manufacturing sites al-
ready exist at given locations, so do the customer zones. Although the location of
distribution facilities is given, the design in terms of capacity of each distribution
center and its connectivity is considered as part of the optimization problem. The
model developed and presented in the next sessions is more generic and can be
used to address design issues of the supply chain such as production plant location
and distribution center location.
Section 6.2 presents the problem statement and description and discusses the
formulation proposed for this type of problems. The problem is presented as an
MILP formulation and aims for the optimal design and operation of global supply
chain networks. Section 6.3 presents an industrial case study for a European house-
hold appliances manufacturer for the establishment of a new distribution center.
A reduced version of the model is employed for this purpose. Section 6.4 presents
another industrial case study for a fine chemical manufacturer who wishes to re-
design their global operations in an optimal way. Finally, Section 6.5 concludes with
a summary of the main points of this work and future directions.
6.2
Optimal Design of Production and Distribution Network
This work considers the optimal design and operation of multiproduct, multiech-
elon production, and distribution networks. These networks can be global or re-
gional. The principles of designing such systems are the same irrelevant of the
scope. As shown in Fig. 6.1, the network may consist of a number of existing mul-
tiproduct manufacturing sites at fixed or potential locations, a number of existing
or potential distribution centers, and finally a number of customer zones at fixed
locations.
In general, each product can be produced at several plants at different locations
and this is the reason we talk about multipurpose pants. The production capacity
of each manufacturing site is modeled in terms of a set of linear constraints relat-
ing the mean production rate per product to the availability of the plant and the
number of changeovers between campaigns of products. The operating principles
of production sites are the same for discrete and process manufacturing.
Distribution centers are secondary warehousing locations and are described by
upper and lower bounds on their material handling capacity, and they can be sup-
plied from more than one manufacturing plants and can supply more than one
Production plant
Distribution center
Customer zone
Possible connection
Fig. 6.1 A typical supply chain representation.
customer zone. However, “single sourcing” constraints requiring that a distribu-

tion center be supplied by a single plant can also be accommodated.
Each customer zone places demands for one or more products. These demands
may be assumed to be known a priori. A customer may be served by more than one
distribution center. Alternatively, single sourcing constraints, according to which a
customer zone must be served by a single distribution center, may be imposed by
modifying certain allocation constraints.
Operational costs include those associated with production, handling of mater-
ial at distribution centers, transportation, and duties. Transportation costs are as-
sumed to be linear functions of the actual flow of the product from the source stage
to the destination stage. Additionally to transportation costs we may have taxes and
duties for transporting material between countries. For regional operations or sin-
gle country operations duty and tax costs are reduntant.
The decisions to be determined include the product portfolio per production
plant, production amounts, utilization level, and transportation links to establish in
the network along side with material flows. Infrastructural costs include the setup
and shutdown of production plants and establishment of distribution centers.
The objective is the minimization of the total annualized cost of the network,
taking into account both infrastructure and operating costs. This work considers
a steady-state form of the above problem according to which demands are time-
invariant and all production and transportation flows determined by the optimiza-
tion are considered to be time-averaged quantities.
The above problem is formulated mathematically as an MILP optimization prob-

lem. We aim to obtain an optimal design of such networks. The formulation de-
scribed is more generic and can be easily modified to accommodate variances of
networks and data availability.
6.2.1
Notation
The notation to be used in this section is described in the appendix to this chapter.
6.2.2
Mathematical Formulation of Deterministic Problem
The mathematical model proposed for this problem is an MILP problem as de-
scribed below.
6.2.2.1 Network Structure Constraints

A link between a production plant j and a distribution center k may exist only if
production plant j is established:
Xj k Yj , ∀j, k. (6.1)
If a distribution center k is established, it can then be served by more than one

production plant j:

Xj k Yk , ∀k. (6.2)
j
This constraint can be transformed into a single source constraint for case where
this restriction applies. In this case the constraint has to be written as:

Xj k = Yk , ∀k. (6.3)
j
The link between a distribution center k and a customer zone l may exist only if
the distribution center is established:
Xkl Yk , ∀k, l. (6.4)
Each customer can be supplied by more than one distribution centers to satisfy
demand:

Xkl 1, ∀l. (6.5)
k
In accordance with the “single sourcing,” each customer zone can be served by
exactly one distribution center and the constraint has to be modified as follows:

Xkl = 1, ∀l. (6.6)
k
6.2.2.2 Logical Constraints for Transportation Flows

Flow of material from production plant j to distribution center k can take place
only if the corresponding connection exists:

j k Xj k
Qmin Qij k Qmax
j k Xj k , ∀i, j, k. (6.7)
i
Flow of material from distribution center k to customer zone l can take place only
if the corresponding connection exists:

kl Xkl
Qmin Qikl Qmax
kl Xkl , ∀i, k, l. (6.8)
i
Values for the upper bounds Qmax max

j k , Qkl appearing on the right hand sides of con-
straints (6.7) and (6.8) are determined by contracts in place for the transportation
of materials. The bounds apply on the total amount sent to each destination.
6.2.2.3 Material Balances

The actual rate of production of product i by plant j must equal the total flow of
this product from plant j to all distribution centers k:

Pij = Qij k , ∀i, j. (6.9)
k
The proposed model assumes that production plants operate independently each
other, therefore, no interaction exists. This does not limit the generality of the
model.
Assuming that there is no stock accumulation or depletion (i.e., steady-state op-
eration), the total rate of flow of each product leaving a distribution center must
equal the total rate of flow entering this node of the supply chain network:

Qij k + Oik = Qikl , ∀i, k. (6.10)
j l
Ideally, the total rate of flow of each product i received by each customer zone l
from the distribution centers must be equal to the corresponding market demand.
In the case where this amount is not sufficient to cover the demand this is partially
filled by out-sourced material which is supplied to the distribution centers:

Qikl = Dil , ∀i, l. (6.11)
k
In case where outsourcing is not available or not allowed the corresponding vari-
ables Oik are set to zero.

An important issue in designing the distribution network is the ability of the man-
ufacturing plants to cover the demands of the customers as expressed through the
orders received from the warehouses.
The rate of production of each product at any plant cannot exceed certain limits.
Thus, there is always a maximum production capacity for any product; moreover,
there is often a minimum production rate that must be maintained while the plant
is operating:
Yj Pijmin Pij Pijmax Yj , ∀i, j. (6.12)
If a decision to shut-down a plant is taken the above constraint enforces the pro-
duction to zero.
Production is restricted by the number of changeovers and number of campaigns
performed at a site. Also maintenance has to be taken into account:

Tij Hj − Mj Yj − τ Nijc Wij , ∀j. (6.13)
i
The number of days available for the production of product i in production site j
is limited by the total availability of the plant, subject to the plant producing this
product:
Tijmin Wij Tij Tijmax Wij , ∀i, j. (6.14)
The production per site is organized in terms of campaigns due to the seasonal
nature of the business, the demand patterns, and the portfolio of products which
require long cleaning periods of the production lines between products. Therefore,
the production capacity is reduced based on the number of changeovers. We as-
sume that on average the same length of cleaning or retooling is required between
products.
The total production of each product depends on the daily production capability
of the plant and the number of days allocated per year:
Pij = rjd Tij ∀i, j. (6.15)
The utilization of each plant is equal to the number of production days per year
and is given by:

Uj = Tij ∀j. (6.16)
i
Due to multisite production we want to ensure that production is distributed evenly

between sites, therefore, utilization per site should be similar with the difference
limited to a specified range. The following constraints ensure that the total utiliza-
tion per site remains close and balanced among the sites:
U Uj − Uj ∀j, j = j,
U Uj − Uj ∀j , j = j , (6.17)
U
ζ.
6.2.2.5 Capacity of Distribution Centers

The capacity of a distribution center k generally has to lie between given lower and
upper bounds, Dkmin and Dkmax , provided, of course, that the distribution center is
actually established (i.e., Yk = 1):
Dkmin Yk Dk Dkmax Yk , ∀k. (6.18)
We generally assume that the capacities of the distribution centers are related lin-
early to the flows of materials that they handle. This is expressed via the constraints:

Dk δik Qikl , ∀k. (6.19)
i,l
where δik is a given coefficient that relates product and distribution center capacity.
6.2.2.6 Objective Function

In general, a distribution network involves both capital and operating costs. The
former are one-off costs associated with the establishment of the infrastructure
of the network, and in particular its warehouses and distribution centers. On the
other hand, operating costs are incurred on a daily basis and are associated with
the cost of production of material at plants, the handling of material at warehouses
and distribution centers, and the transportation of material through the network.
Fixed Infrastructure Costs The infrastructure costs considered by our formulation

are related to the establishment or shut-down of a production plant or a distribution
center at a candidate location. These costs are expressed in the following objective
function terms:
P ,e P ,s D,e D,s
Cj Yj + Cj (1 − Yj ) + Ck Yk + Ck (1 − Yk ).
j j k k
If we assume that the production plants are already established, we need to provide
the option of expansion. Therefore, we do consider the capital cost associated with
their redesign and construction.
We ignore any infrastructure cost associated with customer zones.
Production Cost The production cost is given by the product of the production rate
Pij of product i in plant j , with the unit production cost CijP . The corresponding
term in the objective function is of the form:

CijP Pij .
i,j
The production cost CijP is expressed in terms of local currency.

As part of the production cost we have to account the cost of changeovers as lost
production

τ CijP Nijc Wij .
ij
As production costs we account for the material out-sourced to third parties to

satisfy excess demand.

CiS Oik .
i,k
Material Handling Costs at Distribution Centers Handling costs can usually be ap-
proximated as linear functions of the total throughput. They can be expressed as
follows:

DH
Cik Qikl .
i,k l
Handling and temporary storage costs at production plants are included in the
production cost.
Transportation Cost The unit transportation cost is constant independent of the

product transported and depends only on the source and destination node.
Transportation costs that we include are those between production plant to dis-
tribution centers and distribution centers to customers. Therefore, the following
term needs to be added to the objective function:

CijT k Qij k + T
Cikl Qikl .
i,j,k i,k,l
Duties In the specific model duties apply to the material transferred from the pro-
duction sites to the distribution centers and are calculated in currency of the im-
porting country.
The unit duty cost is a function of the production and transportation costs be-
tween the source and destination. It is applied as a coefficient on the combined
production and transportation cost:
CijDk = aij k (βCijP + γ CijT k ).
The contribution to the objective function is given by

CijDk Qij k .
i,j,k
The unit duty cost is a function of the handling and transportation costs between
the source (distribution centre) and destination (customer). It is applied as a coef-
ficient on the combined material handling and transportation cost:
D DH T
Cikl = aikl (βCik + γ Cikl ).
The contribution to the objective function is given by:

D
Cikl Qikl .
i,k,l
Overall Objective Function By combining the cost terms derived in the above-
mentioned Sections, we obtain the total cost of the supply chain network, which
is to be determined by the optimization:
min equation destorted?

production plant infrastructure costs

P ,e P ,s
Cj Yj + Cj (1 − Yj ) +
j j
distribution center infrastructure costs

D,e D,s
Ck Yk + Ck (1 − Yk ) +
k k
production cost

CijP Pij + CiS Oik + τ CijP Nijc Wij + (6.20)
i,j i,k i,j
material handling cost at distribution centers

DH
Cik Qikl +
i,k l
transportation costs

CijT k Qij k + T
Cikl Qikl +
i,j,k i,k,l
duties cost

CijDk Qij k + D
Cikl Qikl .
i,j,k i,k,l
The above minimization is subject to all the constraints presented in Sections

6.2.2.1–6.2.2.5.
The objective function can be modified to define a maximization problem based
on the profit by satisfying the product demand at a specific product price.
6.3
Distribution Center Location Case Study
The purpose of this case study is to design the distribution network of an enter-
prise based on forecasted annual data and redesign their network to accommodate
the company’s strategy. Based on robust forecasting techniques a demand projec-
tion was extracted for their operating business environment. The redesign of the
distribution network is required to accommodate the new demand pattern.
The enterprise comprises two production sites with virtually unlimited ware-
housing capacity which supply directly to local customers and to secondary distrib-
ution centers on various geographical locations. The management of the company
wants to examine the options of establishing a distribution center in a European
country to serve its local customer basis.
The above problem is modeled based on a reduced version of the above described
mathematical formulation.
A number of variables and constraints are reduntant for this case and have to
be dropped. The variables used to support structural decisions are based on two
main types of binary (decision) variables. The decisions taken are the location and
establishment of distribution centers, and the connections between warehouses
and distribution centers as well as the connections between distribution centers
and customers.
On the operational aspects we need variables to describe the rate of production
of each product at each production plant, the rate of transportation flow of each
product from plants/warehouses to distribution centers and from there finally to
customers. At last a number of variables are introduced to model the capacity of
the warehouses and distribution centers.
6.3.1
Problem Description and Data Analysis
The problems that we want to address through this case study are:
A. The selection of the best location for a new distribution center in Germany to
serve the local market.
The criteria for the optimum decision and evaluation of the possible positions
will be:
• the “geographical position” of the customers;
• the total volume of products transported to every customer;
• the cost per distance of each transportation mean;
• the decision between the construction or the rental of the
new distribution centre;
• the possible storage of raw materials through the new
distribution center.
B. The mathematical model should optimize infrastructural and operational deci-

sions for the network.
The following sections provided analyze the available data and present the results
obtained when the model of Section 6.2 was applied in a reduced form.
6.3.2
Production
Production is happening at two different sites. Based on experience and capacity

analysis studies made prior to this that the production capacity is sufficient to cover
the demand placed. Therefore, the production contraints are reduntant and they ap-
ply only as maximum available capacity. Production capacity expresses the ability of
the plant to produce certain volume of products forecasted. There is an interaction
between the plants in terms of intermediate materials used for the assembling of
the final products but the information provided does not allow for further exploita-
tion of this relationship and is not of the interest of the current problem.
Table 6.1 Production cost per product (amounts are in euros/item).
Production plant Product Labor cost Raw material cost Overhead costs Total cost
PL1 P1 12.50 69.50 2.66 84.66

PL1 P2 12.52 77.41 2.67 92.60
PL1 P3 10.47 70.99 2.42 83.88
PL1 P4 10.47 81.29 2.42 94.18
PL1 P5 10.47 78.96 2.42 91.85
PL1 P6 10.47 97.58 2.42 110.47
PL1 P7 16.54 89.81 3.82 110.17
PL1 P8 17.25 98.55 3.99 119.79
PL1 P8 16.54 94.44 3.82 114.80
PL1 P9 16.54 108.35 3.82 128.71
PL2 P10 1.90 3.53 0.19 5.62
PL2 P11 3.29 7.96 0.30 11.55
Analytical information on the production costs is provided by breaking the costs

to labor costs, raw material, and industrial costs (see Table 6.1). There is no equip-
ment or general resource information that allows detailed modeling of the plants
or an indication of a bottleneck. Based on historical data, the plants can satisfy the
product demands allocated without facing capacity problems.
6.3.3
Warehousing
There are two types of storage for the finished products:

• Major warehouses which are located next to each plant and
they are already established. The warehousing ability of these
sites is assumed to be unlimited and no handling or
inventory costs are provided. The existence of these
warehouse is a virtual echelon between production and
distribution to the customers, but it does not affect the
problem decisions since they are already established.
• Distribution centers sites described as secondary warehouses
to be designed and allocated. These are sites with limited
warehousing capacity and depend on the market size they
are allocated to serve. Since we have a design case, only the
expected minimum and maximum storing capacity may be
known. In our case is only the maximum capacity and
furthermore we want to select one out of eight possible sites
to establish a new distribution center which will serve the
local market. The selection of the suitable distribution
centers depends on the transportation costs, establishment
costs and the required capacity. The infrastructure costs
consist of the personel costs and the cost to rent the space.
Table 6.2 Labor and rental costs to determine infrastructure costs per location.
Potential DC Labor cost Rental cost Total infrastructure cost

(Euro/h) (Euro/m2 ) (Euro)
DC1 8.69 2.56 116,810

DC2 9.72 3.58 131,730
DC3 7.67 2.56 103,550
DC4 9.20 3.07 124,205
DC5 9.72 3.58 131,730
DC6 9.20 3.07 124,205
DC7 9.20 3.07 124,205
DC8 10.23 4.09 139,725
The infrastructure costs are calculated on the assumption of

five employees per site and a warehouse of 1500 m2 surface
enough to accommodate approximately 70,000 items. We
assume 40 h per week for 52 weeks per year labor costs and
overheads of 25% (Table 6.2).
• By local market we mean the biggest selling points of the
products. Six such points have been identified. Regarding the
cost information, we have the cost details for material
handling and establishment costs. These are expressed as an
annualized amount.
6.3.4
Transportation
Transportation in our system occurs between the following stages:

• Production plants and distribution centers (Table 6.3).
• Distribution centers and customers (Table 6.4).
The information provided included the transportation costs between warehouses

and distribution centers, and between distribution centers and customers. The
costs depend on the product and distance of the two points. Transportation costs
occuring between plants and warehouses are negligible compared to the others
since the warehouses are located in the same site.
6.3.5
Demand
The potential distribution center will cover a market where the demand is given in
Table 6.5. Demand is given in terms on annual amounts as it has been forecasted
based on historic profiles and projection.
Table 6.3 Transportation costs between production plants and
potential distribution centers (euros/item).
Potential DC Plants
PL1 PL2
DC1 9.75 11.00

DC2 11.75 13.00
DC3 11.25 12.50
DC4 10.25 11.50
DC5 8.95 10.20
DC6 11.75 13.00
DC7 12.50 13.75
DC8 10.00 11.25
Table 6.4 Transportation costs from potential distribution centers to customers (euros/item).
Potential DC Customers
C1 C2 C3 C4 C5
DC1 5.00 5.10 3.90 1.10 2.70

DC2 1.50 2.65 3.20 3.30 7.50
DC3 2.85 1.45 4.20 4.50 6.20
DC4 4.85 2.90 3.80 2.30 4.00
DC5 7.50 4.95 5.90 3.00 0.80
DC6 2.60 2.40 2.20 3.10 7.50
DC7 4.60 5.60 0.00 1.90 5.50
DC8 8.25 5.90 5.50 3.90 0.00
Table 6.5 Annual customer demand data for base case scenario (in terms of items per product).
Customers Products
P1 P2 P3 P4 P5 P6 P7 P8 P9 P10 P11
C1 95 140 115 140 65 55 60 35 165 745 7,165

C2 365 290 65 275 90 70 135 90 340 3,276 5,860
C3 1,240 1,110 845 440 210 565 205 105 235 4,765 7,160
C4 370 255 305 170 115 200 140 85 470 3,425 10,420
C5 900 485 320 290 145 325 190 130 105 2,680 7,280
6.3.6
Results – Network Description
The mathematical formulation was implemented in GAMS [4] and CPLEX is the
solver used to solve the MILP problem.
8,780 it ems
43,900 Euro
C1
10,856 it ems
55,365 Euro
PL1
C2
12,545 it ems Capacit y 65,321 it ems
122,314 Euro 116,810 Euro 16,880 it ems
Production cost
1,224,230 Euro DC1 65,832 Euro
C3
PL2
C4
52,776 it ems 15,955 it ems
580,536 Euro 17,550 Euro
Production cost 12,850 it ems
521,259 Euro 34,695 Euro C5
Fig. 6.2 Optimal location for distribution center as part of the supply chain network.
One distribution center need to be established to serve these five major customer
points. The location selected purely based on financial objectives and the network
configuration is given in Fig. 6.2.
The figure shows the break down of costs as they have been determined by the
optimization.
Each arrow shows the flow of material and the corresponding cost associated.
Production costs are shown for each plant as well as the cost for establishing the
distribution center and its estimated capacity.
The selection of DC1 as the location of the new distribution center is explained
by both its geographical location and the associated costs for establishing and op-
erating the site. It can serve two of the biggest customers at a minimal cost and
also has the least cost for transporting material from the plants. The infrastructure
costs are competive and within the margins.
The solution provided in this case study and its quality is data driven, as it hap-
pens in most mothematical formulations which have been designed to be generic
and easy to adapt. The solution time is considerably small (a few seconds on an
Intel Pentium Centrino processor) and strongly depends on the data and the prop-
erties of the formulation.
6.3.7
Comments
The problem solved represents a steady-state case of a supply chain operation and
design. All costs and material flows are time-averaged and small disturbances do
not affect the whole operation. This case can be considered as a single scenario
network supply chain design problem. We cannot examine the robustness of the
design over time or on other occasions due to the lack of customer demand data
for more periods or an other relevant data.
6.4
Optimal Production Allocation and Distribution Supply Chain Networks
To further demonstrate the applicability of the proposed model we examine the per-
formance of a global organization with a supply chain extending in many countries
in different continents where multiple currencies and taxes apply.
The model was applied on a network consisting of six manufacturing plants lo-
cated at different countries on different continents. Each plant can produce six
types of products using a number of shared production resources, therefore, only
one product may be produced at a time. A number of possible distribution cen-
ters are considered to handle the material based on the market demand placed by
the customer zones. Customer zones comprise eight large geographical areas. The
network superstructure is given in Fig. 6.3. The superstructure assumes complete
connectivity between all sites which is subject to constraints regarding material
transfers.
PL1
PL2
C8 DC1
DC2
PL3 C2
DC3
C1
C6
C3
PL5
DC5
PL4
C5
PL6
C7
DC4 C4
DC6
Fig. 6.3 Network superstructure.
We examine different scenarios of operation to demonstrate the benefits of opti-

mization over decisions imposed by practice or heuristics. These scenarios are:
6.4 Optimal Production Allocation and Distribution Supply Chain Networks 175
• free optimization where all decisions are determined by the

model subject to the constraints presented in the following
data sections;
• optimize network flows when product allocation decisions
and network structure has been fixed a priori by
management.
The model described above is implemented in GAMS [4] and CPLEX is the solver
used to solve the MILP problem. Computationally the problem does not pose sig-
nificant challenges since the size is limited to only a few dozens of binary variables.
6.4.1
Problem Description – Data Analysis
6.4.1.1 Production Plants

The theoretical maximum production capacity of each plant with respect to each
product (i.e., the parameter Pijmax of the formulation) is given as a result of Ta-
ble 6.6 and the maximum availability of the plant. The corresponding minimum
production rates Pijmin are taken to be zero unless the product is selected.
The production per plant in term of tonnes of product per day is given in Ta-
ble 6.6. This is the bulk product capacity.
The operating horizon is 365 days per plant and maintenance duration per pro-
duction site is 10 days except plants 5 and 6 where we have 7 days.
Maximum production for each product per site is Tijmax = 365 days, while the
minimum campaign length is Tijmin = 7 days.
The changeover duration is assumed to be τ = 1 day and we assume that the
utilization coefficient is ζ = 10 days.
The unit production costs per product for each plant are given in Table 6.7.
Note that all costs displayed have been converted to relative money units based
on certain exchange rates for confidentiality reasons. Sensitivity analysis was per-
formed on the exchange rate fluctuation and the average price was used.
6.4.1.2 Distribution Centers

The distribution centers to be designed are assumed to have sufficient capacity to
handle the material received and distributed, and the logistics management is not
an issue at this level. In order for a connection between a production plant and
distribution center to exist that must be at least for a total flow of 1000 tonnes (i.e.,
j k = 1,000 ∀j, k).
Qmin
Table 6.6 Daily production rate per production site (tonnes/day).
Plants
PL1 PL2 PL3 PL4 PL5 PL6
73.4 12.7 35.1 34.4 11.0 10.0

Table 6.7 Unit production cost per product per plant (rmua /kg).
Product Production site

PL1 PL2 PL3 PL4 PL5 PL6
P1 4.68 4.53 5.17 5.04 4.27 4.26

P2 3.23 3.17 3.21 3.87 3.12 3.97
P3 4.18 4.19 5.17 4.35 3.71 3.70
P4 2.48 2.71 2.54 2.74 2.43 2.42
P5 1.43 1.49 1.55 1.41 1.28 1.27
P6 2.15 2.15 2.22 2.40 2.07 2.06
a rmu: relative money units
Table 6.8 Transportation cost between production sites and distribution centers (rmu/kg).
From/to Distribution center

Plant DC1 DC2 DC3 DC4 DC5 DC6
PL1 0.00 0.062 0.081 0.053 0.051 0.056

PL2 0.065 0.089 0.127 0.077 0.071 0.077
PL3 0.083 0.126 0.00 0.141 0.125 0.119
PL4 0.107 0.138 0.182 0.00 0.143 0.144
PL5 0.183 0.236 0.117 0.038 0.00 0.038
PL6 0.168 0.178 0.236 0.117 0.038 0.00
The transportation costs from the production plants to the distribution centers
are given in Table 6.8. This is the unit transportation cost and is independent of the
product type and the amount transported.
The amount transported to the distribution centers is subject to duties as they
have been defined in the section “Duties”. Table 6.9 describes the coefficients used
between production sites and distribution centers.
For this case parameter β = 1.1 and γ = 1.0. The relationship can then be
written as:

αij k 1.1 · CijP + CijT k
The establishment of a distribution center occurs fixed costs for the organization.
These amounts have been amortized to annual amounts. Depending on the mate-
rials stored there is a handling cost per product. Both costs are given in Table 6.10.
6.4.1.3 Customer Zones

The customer zones are located close to the distribution centers and the allocation
for the specific problem is considered fixed, therefore, distribution costs can be
ignored.
The demand per product and customer zone is given in Table 6.11.
Table 6.9 Values of coefficient αij k to apply on duties structure (%).

Plant Product group DC1 DC2 DC3 DC4 DC5 DC6
PL1 P1–P6 0.0 0.0 7.1 53.1 10.0 5.0

P5 0.0 0.0 3.8 53.1 10.0 12.5
PL2 P1–P6 0.0 0.0 7.1 53.1 10.0 5.0
P5 0.0 0.0 3.8 53.1 10.0 12.5
PL3 P1–P6 6.5 9.0 0.0 53.1 10.0 5.0
P5 6.5 9.0 0.0 53.1 10.0 12.5
PL4 P1–P6 0.0 9.0 4.1 0.0 10.0 5.0
P5 0.0 9.0 0.3 0.0 10.0 12.5
PL5 P1–P6 0.0 9.0 7.1 53.1 0.0 5.0
P5 0.0 9.0 3.8 53.1 0.0 12.5
PL6 P1–P6 6.5 9.0 4.1 53.1 10.0 0.0
P5 6.5 9.0 0.3 53.1 10.0 0.0
Table 6.10 Costs associated with the distribution centers upon establishment.
Distribution center costs

DC1 DC2 DC3 DC4 DC5 DC6
Fixed infrastructure cost (,000 rmu)

4,300 2,900 3,100 2,200 1,300 1,500
Material handling cost per product (rmu/kg)
P1 0.07 0.068 0.077 0.076 0.064 0.064
P2 0.048 0.048 0.048 0.058 0.047 0.060
P3 0.063 0.063 0.077 0.065 0.056 0.056
P4 0.037 0.041 0.038 0.041 0.036 0.036
P5 0.021 0.022 0.024 0.021 0.019 0.019
P6 0.032 0.032 0.033 0.036 0.031 0.031
Table 6.11 Customer zone demand per product (tonnes).
Product Customer zones

C1 C2 C3 C4 C5 C6 C7 C8
P1 5,701 3,665 3,398 2,686 1,342 2,310 2,436 1,819

P2 2,116 365 1,468 1,355 910 630 1,049 748
P3 3,246 887 1,369 1,353 609 410 1,219 679
P4 2,582 295 3,016 466 556 700 1,436 906
P5 3,683 437 2,284 922 794 690 1,578 844
P6 294 64 311 1,173 130 110 172 323
Table 6.12 Transportation cost between distribution center and customers (in rmu/kg).
Distribution center Customer zones

C1 C2 C3 C4 C5 C6 C7 C8
DC1 0.00 0.62 0.81 0.53 0.51 0.56 1.31 1.66

DC2 0.65 0.00 1.27 0.77 0.73 0.77 1.38 1.77
DC3 0.90 1.40 0.0 1.41 1.25 1.19 1.35 1.29
DC4 1.13 1.52 1.82 0.0 1.43 1.44 1.93 1.54
DC5 1.89 2.16 2.61 1.66 0.00 0.57 1.57 1.34
DC6 1.74 1.78 2.36 1.17 0.00 0.59 1.36 1.23
Table 6.13 Duty coefficient for material transferred between

distribution center and customer αikl (%).
Distribution center Customer zones

C1 C2 C3 C4 C5 C6 C7 C8
DC1 0.0 0.0 7.1 53.1 12.5 11.0 17.0 15.0

DC2 0.0 0.0 7.1 53.1 12.5 11.0 17.0 15.0
DC3 6.5 9.0 0.0 53.1 12.5 11.0 17.0 15.0
DC4 6.5 9.0 4.1 0.0 12.5 11.0 17.0 15.0
DC5 6.5 9.0 7.1 53.1 0.0 11.0 17.0 15.0
DC6 6.5 9.0 4.1 53.1 12.5 11.0 17.0 15.0
The transportation costs from distribution centers to customer demand zones

are given in Table 6.12.
The amount transported to the customer demand zones is subject to duties as
they have been defined in the section “Duties”. Table 6.13 describes the coefficients
used between distribution centers and customers.
For this case parameter β = 1.0 and γ = 1.0. The relationship can then be
written as:

DH T
aikl Cik + Cikl .
6.4.2
Network Optimization – Base Scenario
The first scenario is a free optimization exercise where the model is allowed to de-
termine all the decisions required such as, product allocation per plant and produc-
tion level, establishment of distribution centers and customer allocation, material
flows, and distribution center capacity.
The optimal solution for the network is given in Fig. 6.4.
Table 6.14 shows the product allocation per plant and the number of days of
production per product. Out-sourced material is purchased to cover demand that
cannot be satisfied by internal production and is presented as the last column of
PL1
PL2
DC1
DC2
C8
PL3 C2
DC3 C6
C1
PL5
C3
DC5
PL4
C5
C7
PL6
DC4
C4
Fig. 6.4 Optimal network configuration for base scenario.
Table 6.14 Product allocation per plant and production days allocated.
Product Plants
PL1 PL2 PL3 PL4 PL5 PL6 Out-sourced (tonnes)
P1 148 321 0 173 131 133 0

P2 54 0 61 42 108 0 0
P3 69 0 50 66 0 58 0
P4 0 0 124 0 0 48 5,097
P5 58 0 95 0 0 94 0
P6 0 14 0 50 96 0 2,289
Utilization
Total 329 335 330 331 335 333
the same table. This material is directly supplied to the distribution centers and
from there to the customer zones assigned.
The total material supplied from the production plants to distribution centers is
shown in Table 6.15. This is subject to the material flow constraint that applies for
an established link. At the bottom of the table, the total material handling capacity
of each distribution center is given. This includes the amount received by third
parties as out-sourced production.
Material flows from a distribution center, where it has been established to the
assigned customers, is equal to the total demand placed by the customer. The flow
Table 6.15 Material flow from production plants to distribution centers (in tonnes).

Production site DC1 DC2 DC3 DC4 DC5 DC6
PL1 20,765 2,096 1,000 0 0 0

PL2 1,000 3,258 0 0 0 0
PL3 1,000 0 10,538 0 0 0
PL4 1,418 0 3,656 6,316 0 0
PL5 1,000 0 0 0 2,340 0
PL6 1,000 0 1,337 0 1,315 0
Total handling capacity 30,252 5,713 17,165 7,955 4,341 0
per arc equals to the total product demand per customer since we enforce single
source constraints.
The total annual cost of operation for this case is 311,681,000 rmu. Individual
contributions to the objective function are given in Table 6.18.
6.4.3
Fixed Product Allocation and Customer Assignment Scenario
This scenario assumes that production allocation and network configuration in

terms of customer allocation to distribution centers are decisions already taken
as part of a pre-planning exercise. Product allocation is made according to plant
capabilities and bidding process from the plants. Secondary, customer demands
are assigned to the nearest possible geographically site to minimize transportation
costs.
As a result, the model size is reduced in terms of number of decisions variables
need to be determined since Xkl and Wij are fixed.
The model is allowed to optimize the production levels (as days per year and
number of campaigns) and flows within the network as well as connections be-
tween plants and distribution centers. Table 6.16 shows the activity levels for each
plant based on the pre-allocation of products. Out-sourced material purchased for
this case is given as the last column of Table 6.16. More products are out-sourced
and in total is higher than the base case.
The flow of material between plants and distribution centers is given in Ta-
ble 6.17.
The total cost of operation in this case is 327,049,000 rmu. The break-down of
individual costs is given in Table 6.18.
The solution of the problem is given in Fig. 6.5 which also shows the customer
assignment to distribution centers. These connections were fixed.
Table 6.16 Production per plant as days per product.
Product Plants
PL1 PL2 PL3 PL4 PL5 PL6 Out-sourced (tonnes)
P1 91 327 84 165 69 133 1,711

P2 50 0 60 42 91 42 0
P3 62 0 42 70 59 58 134
P4 42 0 56 0 28 28 4,323
P5 74 0 75 49 80 66 0
P6 0 0 0 0 0 0 2,467
Total 319 327 317 326 327 327
Table 6.17 Material flow between production plants and distribution centers (in tonnes).
From/to Distribution centers

Production site DC1 DC2 DC3 DC4 DC5 DC6
PL1 19,869 2,197 0 0 0 1,353

PL2 1,000 3,156 0 0 0 0
PL3 0 0 9,482 0 0 1,634
PL4 1,000 0 1,000 6,316 0 2,936
PL5 0 0 0 0 3,260 0
PL6 0 0 0 0 0 3,583
Total handling capacity 22,362 5,713 11,846 7,955 4,341 13,209
PL1
PL2
DC1
DC2
PL3 C2
DC3 C6
C1
PL5
C3
DC5
PL4
C5
PL6
DC6
C7
DC4
C4
C8
Fig. 6.5 Network configuration for second case.

Table 6.18 Cost analysis and comparison of cases (,000 rmu).
Cost analysis Case 1 Case 2
Global production costs 225,641 223,558

Global distribution costs 22,060 21,397
Global duties paid 4,503 5,942
Global out-sourcing cost 42,230 57,458
Infrastructure costs 13,800 15,300
Material handling cost 3,447 3,394
Total costs 311,681 327,049
Difference 0.00% +4.93%
6.4.4
Results Analysis
The results obtained demonstrate clearly the savings of optimization against

heuristic decisions that dictate part of the network structure. The savings are signif-
icant to justify the redesign of the supply chain to comply with the results obtained
design.
The benefits are a result of a better utilization of the plants by selecting suitable
product portfolios, minimization of the out-sourced material and savings on total
transportation and duties. Savings also result from infrastructure decisions since
only 5 distribution centers are needed to serve the customer basis.
Although the plants are multipurpose, it is beneficial for the organization to op-
erate specific plants in a dedicated mode. This reduces switches between products
and production costs.
Other scenarios to be considered are:
• optimize the network when each plant can manufacture up
to a certain number of allocated products;
• examine the network sensitivity by changing some of the
parameters such as utilization constraint, changeover
duration, and others.
6.4.5
Comments
This case study has demonstrated the application of an integrated model based on a
detailed mathematical programming formulation that addresses some of the com-
plex issues related to the design and operation of a global supply chain network.
The focus is on financial and tactical operational aspects within the organization
taking into account production balancing among sites.
Between other business benefits is the operational and distribution efficiency of
the network, visibility and control of the supply chain and capability to perform
toward KPIs such as operational cost, customer satisfaction, and product quality.
6.5 Conclusions 183
Moreover, the impact of decisions on the design and tactical operation can be quan-
tified and evaluated. The data used is extracted directly from ERP systems as busi-
ness objects making such approaches easy to use and easy to update with fresh
data.
The availability of models and subsequent tools with this capability have given
the opportunity to the company to examine their network’s performance every
quarter. In case where a decision is needed it can be justified quantitatively.
6.5
Conclusions
The management of supply chains has been studied extensively over the last two
decades with numerous approaches. Mathematical programming approaches are
still into their infancy due to the extensive use of heuristic techniques and qual-
itative rather than quantitative solution methods. Optimization techniques have
gained ground during the last years due to cheap computational power and the
implementation of ERP systems that allow data to be exported and made available
easily. All these models are highly data sensitive and their success depend strongly
on good quality data. The ability of the models to examine multiple scenarios si-
multaneously minimizes the effort required to analyze operating strategies and
tactical operational decisions. Moreover, optimization can provide the best avail-
able answer.
Due to their nature, supply chains are subject to uncertainty in many aspects
such as costs, product demands, raw material availability, exchange rates, du-
ties, etc. Research is trying to address these considerations by proposing new ap-
proaches to incorporate uncertainty either as stochastic variables or as discrete sce-
narios. There is no unique answer to what approach is preferable but strongly de-
pends on the problem nature and the use of the results obtained.
References
1 Arntzen, B. C., Brown, G. G., Har- GAMS Development Corporation,

rison, T. P., Trafton, L. L., Global 1998.
supply chain management at digital 5 Brown, G. G., Graves, G. W., Hon-
equipment corporation. Interfaces 25 czarenko, M. D., Design and oper-
(1995), pp. 69–93. ation of a multicommodity produc-
2 Ballou, R. H., Unresolved issues in tion/distribution system using primal
supply chain network design. Inf. Syst. goal decomposition. Manage. Sci. 33
Frontiers 3(4) (2001), pp. 417–426. (1987), pp. 1469–1480.
3 Bilgen, B., Ozkarahan, I., Strategic 6 Camm, J. D., Chorman, T. E., Dill,
tactical and operational production- F. A., Evans, J. R., Sweeney, D. J., We-
distribution. Int. J. Technol. Manage. gryn, G. W., Blending OR/MS, judg-
28 (2004), pp. 151–171. ment, and GIS: Restructuring P&G’s
4 Brooke, A., Kendrick, D., Meeraus, supply chain. Interfaces 27 (1987), pp.
A., Raman, R., GAMS: A User’s Guide, 128–142.
7 Cohen, M. A., Lee, H. L., Resource 18 Kallrath, J., Combined strategic and
deployment analysis of global manu- operational planning – an MILP suc-
facturing and distribution networks. cess story in chemical industry. OR
J. Manuf. Oper. Manage. 2 (1989), pp. Spektrum 24(3) (2002), pp. 315–341.
81–104. 19 Klose, A., Drexl, A., Facility location
8 Lee, H. L., Billington, C., Manag- models for distribution system de-
ing supply chain inventory: Pitfalls sign. Eur. J. Oper. Res. 162 (2005), pp.
and opportunities. Sloan Manage. Rev. 4–29.
Spring (1992), pp. 65–73.
20 Levis, A., Papageorgiou, L. G., A hier-
9 Geoffrion, A. M., Powers, R. F., archical solution approach for multi-
Facility location analysis is just the site capacity planning under uncer-
beginning. Interfaces 10 (1980), pp. tainty in the pharmaceutical industry.
22–30. Comput. Chem. Eng. 28 (2004), pp.
10 Geoffrion, A. M., Graves, G. W., 707–725.
Multicommodity distribution system
21 Papageorgiou, L. G., Rotstein, G. E.,
design by benders decomposition.
Shah, N., Strategic supply chain op-
Manage. Sci. 20 (1980), pp. 822–844.
timization for the pharmaceutical
11 Gjerdrum, J., Shah, N., Papageor-
industries. Ind. Eng. Chem. Res. 40
giou, L. G., Fair transfer price and
(2001), pp. 275–286.
inventory holding policies in two-
enterprise supply chains. Eur. J. Oper. 22 Pirkul, H., Jayarama, V., Production,
Res. 143 (2002), pp. 582–599. transportation, and distribution plan-
12 Goetschalck, M., Vidal, C. J., Do- ning in a multi-commodity tri-echelon
gan, K., Modeling and design of system. Transport. Sci. 30 (1996), pp.
global logistics systems: A review 291–302.
of integrated strategic and tactical 23 Shapiro, J. F., Challenges of strategic
models and design algorithms. Eur. J. supply chain planning and modelling.
Oper. Res. 143 (2002), pp. 1–18. Comput. Chem. Eng. 28 (2004), pp.
13 Grossmann, I. E., Enterprise-wide op- 855–861.
timisation. A new frontier in process 24 Tsiakis, P., Shah, N., Pantelides,
systems engineering. AIChE J. 51(7) C. C., Design of multi-echelon supply
(2005), pp. 1846–1857. chain networks under demand uncer-
14 Guillen, G., Mele, E., Bagajewicz, tainty. Ind. Eng. Chem. Res. 40 (2001),
M. J., Espuna, A., Puigjaner, L., Mul- pp. 3585–3604.
tiobjective supply chain under un- 25 Williams, J. F., A hybrid algorithm
certainty. Ind. Eng Chem. Res. 60(6) for simultaneous scheduling of pro-
(2005), pp. 1535–1553. duction and distribution in multi-
15 Guinet, A., Multi-site planning: A echelon structures. Manage. Sci. 29
transhipment problem. Int. J. Prod. (1983), pp. 77–92.
Econ. 74 (2001), pp. 21–32.
26 Vidal, C., Goetschalck, M., Strate-
16 Harrison, T. P., Global supply chain
gic production-distribution models:
design. Informat. Syst Frontiers 3(4)
A critical review with emphasis on
(2001), pp. 413–416.
global supply chain models. Eur. J.
17 Kaihara, T., Multi-agent based supply Oper. Res. 98 (1997), pp. 1–18.
chain modelling with dynamic envi-
ronment. Int. J. Prod. Econ. 85 (2003),
pp. 263–269.
Appendix 185
Appendix
Nomenclature
Indices/Sets
i products
j plants
k possible distribution centers
l customer demand zone
Parameters
DH
Cik unit handling cost for product i at distribution center k
CkD,e annualized fixed cost of establishing a distribution center at location k
CkD,s annualized fixed cost of shutting down a distribution center at location
k
Cijd k unit duty cost for product i leaving production plant j to distribution
center k
d
Cikl unit duty cost for product i leaving distribution center k to customer l
CjP ,e annualized fixed cost of establishing a production plant at location j
CjP ,s annualized fixed cost of shutting down a production plant at location
j
CijP unit production cost for product i at plant j
CiS unit production cost for product i supplied to the network from third
parties
CijT k unit transport cost for product i from plant j to distribution center k
T
Cikl unit transport cost for product i from distribution center k to customer
zone l
Dkmin , Dkmax minimum/maximum distribution center capacity
Dil demand for product i in customer zone l
Hj operating horizon of production site j in days per annum
Mj annual days of maintenance of production site j
Nijc number of campaigns of product i at production plant j
Pijmin , Pijmax minimum/maximum production capacity of plant j for product i
Qmin max minimum/maximum rate of flow of materials transferred from plant
j k , Qj k
j to distribution center k
Qmin max minimum/maximum rate of flow of materials transferred from distri-
kl , Qkl
bution center k to customer zone l
Tijmin /Tijmax minimum/maximum availability of plant j for the production of prod-
uct i per annum in days
rjd expected daily production rate of plant j
αj k duty coefficient relating production plant j to distribution center k
αkl duty coefficient relating distribution center k to customer l
β coefficient applied to production cost for duty purposes
γ coefficient applied to transportation cost for duty purposes

δik coefficient relating capacity of distribution center k to flow of product
i handled
ζ utilization parameter in days
τ changeover coefficient in days
Continuous Variables
Dk capacity of distribution center k
Oik out-sourced product i delivered to distribution center k
Pij production rate of product i in plant j
Qij k rate of flow of product i transferred from plant j to distribution center
k
Qikl rate of flow of product i transferred from distribution center k to cus-
tomer zone l
Tij days allocated for the production of product i in plant j
Uj utilization of production plant j as production days per year
U maximum allowed difference in utilization between plants
Binary Variables
Yj 1 if production plant j is to be established, 0 otherwise
Yk 1 if distribution center k is to be established, 0 otherwise
Xj k 1 if production plant j is assigned to distribution center k, 0 otherwise
Xkl 1 if distribution center k is assigned to customer zone l, 0 otherwise
Wij 1 if production plant j is to produce product i, 0 otherwise
187
7
Design and Planning of Closed-Loop Supply Chains
Ana Paula Barbosa-Póvoa, Maria Isabel Gomes Salema, Augusto Q. Novais
7.1
Introduction
Traditionally, supply chains (SC) have been defined as logistics systems that start
with the supply of raw materials and end with the sales and distribution of goods
to final consumers. An efficient management of the supply chain is a major chal-
lenge, which has captured the academia attention. For the last 20 years, the number
of published papers has been growing consistently.
Even with all this awareness, the supply chain is far from being studied. There
are new problems appearing, new ways of regarding companies as organizations
and social partners, new forms of collaboration amongst companies, as well as
new commitments and responsibilities of companies toward the modern society.
Consumers and companies are changing their objectives. These are not strictly
related to cost, quality, and level of service but also include environmental aspects.
Business is recognizing the importance of the environmental performance of their
products and processes. All these factors contribute to change the way of looking
into the supply chain systems.
Most products once used, continue to exist and, consequently, it is necessary
to decide how to handle them. Until the 1980s, the most frequent decision was
simply a disposal option. More recently, this has been changing and a new way
of thinking began to emerge, where attitudes toward waste come under scrutiny.
Although, product recovery is not a new area (glass and metal scraps have been
recycled for many decades now), the last 10–15 years show a consistent increase
in the volume of products to be collected for recovery. In addition, the diversity
of products has spread. A new flow of products has emerged (the reverse flow)
leading to new and challenge problems within the supply chains. The integration
of reverse logistic systems must, therefore, be accounted for. Goetschalckx et al. [1]
consider, within the logistics elements of a supply chain, recycling centers for used
products and returned packaging. Savaskan and Van Wassenhove [2] point out that
the interaction between the forward and reverse channel decisions is particularly
important for channel coordination, and highlights the benefits of an integrated
ISBN: 978-3-527-31693-9
188 7 Design and Planning of Closed-Loop Supply Chains
closed-loop chain. One may think that this reverse flow could be studied using tools
developed for the traditional flow (forward flow). However, it has many particular
characteristics that need to be accounted for.
One of the first definitions of Reverse Logistics was provided by Stock [3], where
Reverse Logistics was defined as: “the term often used to the role of logistics in recy-
cling, waste disposal and management of hazardous materials; a broader perspec-
tive includes all issues relating to logistics activities carried out in source reduction,
recycling, substitution, reuse of materials and disposal”. Since then several authors
have proposed other definitions for reverse logistics.
Within this context, three major types of supply chains can be identified: forward
supply chain, the reverse supply chain, and the closed-loop supply chain. Forward
supply chains correspond to the networks that start at the raw materials and end at
the final customer. On the other hand, reverse supply chains involve the chain that
starts with the release of products by consumers and ends in any kind of recovery.
Finally, when these two chains are linked together, the closed-loop supply chains
(CLSC) is formed. Although, in this latter case the name may suggest that products
return to the original manufacture that is not strictly true since products may be
recovered at different sites within a global structure.
We look into the reverse logistics context and focus on the design and planning
of closed-loop supply chains. A review on the published work on the area is per-
formed. We do not intend to be exhaustive, but rather to refer works that we con-
sider interesting. Additionally, and due to the significance that traditional supply
chains have on closed-loop chains, major reviews on the subject are also under-
taken. Finally, we propose a model for the simultaneous design and planning of
closed-loop supply chains and conclude with some remarks and identification of
research directions.
7.2
Reverse Logistics
Several papers have been published aiming at creating a work basis for the reverse
logistics and CLSC problems [4–6]. However, due to these being recent issues, there
is no formal framework for the area and their contexts are not yet stable and well
defined.
In this section, reverse supply chains will be defined and the causes behind the
release of end-of-use products will be analyzed. Aspects such as what kind of prod-
ucts one should account in the reverse process, what are the structure of the closed-
loop supply chains and what are the major differences between forward and reverse
supply chains are discussed.
7.2 Reverse Logistics 189
7.2.1
Defining the Closed-Loop Supply Chain
After the definition proposed by Stock [3], several others have followed. One widely
accepted is the one given by the European Working Group on Reverse Logistics,
where Reverse Logistics is defined as “The process of planning, implementing and
controlling backward flows of raw materials in process inventory, packaging and
finished goods, from a manufacturing, distribution or use point, to a point of re-
covery or proper disposal” [6]. This definition contains all kinds of returns, leaving
out products that do not have any kind of value to be recovered (e.g., organic waste).
It also emphasis several points where the reverse chain may start and, therefore,
products may be returned for several reasons. De Brito [7] classified the returns in
three major groups:
• Manufacturing returns – after production phase there are
products that remain behind and do not proceed to the
distribution phase. These may be either unused raw
materials, products that failed quality tests, or by-products
[8, 9].
• Distribution returns – there are several reasons for products to
be returned after the distribution phase such as products or
packages may have been damaged, health, and safety
restrictions were not respected, amongst others. In addition,
unsold products and stocks adjustments are also included in
this group. Finally, there are products whose function is to go
back and forth in the supply chain, as pallets, containers [10].
• Customer returns – in this group are included the end-of-life
and end-of-use products. Examples are bottles or leasing
products.
In this chapter, we devote our attention to this last group of returns. We review
works where the supply chain for customer returns is modeled. In this case, prod-
ucts are returned for several reasons that can be associated with three major groups
of reasons:
• Companies perceive new business opportunities. Paper and
metal have been collected for many years now, due to the
value they generated. Nowadays, electronic products have a
very short life cycle. Therefore, many disposed products have
a large intrinsic value that is being recovered by companies.
• Legislative initiatives by governments. In Europe, the
disposal of products has been taken very seriously by the
European Union. Legislation sets targets aiming at the
minimization of the environmental impacts, which include a
mix of waste prevention, material recycling, energy recovery,
and disposal options. Extended producer responsibility (EPR)
is one of the major consequences of these legislative actions.
Now, original manufacturers are being held responsible for

the entire product life cycle, especially for their take-back,
recovery/recycling, and disposal [11]. Being firstly related
with packaging waste, EPR has now spread to other kinds of
waste. Electric and electronic equipments, car and tyres are
amongst products that now face EPR policies [12]. The
recovery of end-of-life products has been ignored by most
manufacturers, as they believe that there were no business
opportunities in the recovery of these products [13]. However,
nowadays companies to remain competitive, have to develop
strategies in order to adapt their supply chains to this new
flow of products [14]. In many cases, the reverse chain is not
created independently from the existing forward chain. This
is specially true if product recovery is being imposed by
legislation as discussed by Fleischmann et al. [15].
• Increasing environmental awareness form customers.
Companies are now more customer focused, which makes
them more responsive toward customers. Companies are
feeling that taking some action concerning the recovery of
their products can make their image more “green” to their
customers [16]. Example of this is the creation of a 100%
recyclable kayak developed by Walden Paddlers, Inc. [17].
Guide and van Wassenhove [18] show that companies should “pro-actively control
the quality, quantity, and timing of products’ returns” in order to make the most
profit out of this imposed flow. The authors alert for the lack of a framework that
helps decision making in this area and state that research on these aspects should
be considered.
The integration of the reverse channel into the supply chain structure leads to
an extended supply chain, the so called closed-loop supply chain as represented in
Fig. 7.1. This is characterized by two main flows: forward (descending flow) and
reverse (ascending flow). The forward flow is the traditional supply chain, where in
the production phase raw materials are transformed into final products that after
distribution reach the final customer. The reverse flow starts at the moment when
products are returned. After a collection phase, returns go through a combined
inspection/sorting process. Some end-of-use products just need some minor activ-
ities (e.g., bottles need an inspection and cleaning phases and afterwards they can
be refilled). After the sorting phase, products may have three destinations: they can
go directly to distribution, they can suffer some degree of re-processing (recycling,
recover, repair, . . .), or may undergo proper disposal or energy recovery.
For a more detail description of reverse processes and activities refer to Thierry
et al. [13] and Fleischmann et al. [19].
Therefore, closed-loop supply chains include traditional forward supply chain ac-
tivities and the additional activities of the reverse chain. One may think that for the
management of the closed-loop supply chain only the concepts of the forward sup-
Raw-materials
Repair
Recover
Recycle
Production Re-processing
Disposal
Distribution Sorting
Cunsumption
Collection
Fig. 7.1 Extended supply chain for end-of-life and end-of-use products.
ply chain are needed. However, this is not true as both forward and reverse flows
differ in several important characteristics. The Reverse Logistics Executive Council
(RLEC) [20] identified several differentiable factors. In the reverse process forecast-
ing is more difficult, many distribution points exist; product packaging and quality
is not uniform (packages are often damaged); destination/routing is unclear; pric-
ing depends on many factors and costs are less visible; inventory management is
not consistent and product lifecycle issues are more complex. Therefore, the over-
all reverse process has different sources of uncertainties and its visibility is less
transparent due to the structures involved.
As referred by Fleischmann [21] for the distribution, “reverse distribution is not
necessarily a symmetric picture of forward distribution”. This can be definitely ex-
tended to the other levels of the chain.
Concluding and as stated by Guide et al. [5], closed-loop supply chains require
careful design, planning, and control, where the differences between both forward
and reverse flows should be considered.
7.3
Literature Review
A literature review on the most important work published on supply chain with
reverse flows is given below. Different areas can be identified within this broad
area, although the present chapter will focus essentially on the treatment of the
design and planning of supply chain using optimization approaches.
Although our primary objective is to review models for the reverse flow, it is also
important to firstly analyze the context of the forward supply chain. This is done
through the analysis of some of the most important published reviews on forward
supply chain modeling. Following this analysis, a review of the published work on
the reverse supply chain is provided. Finally, works on closed-loop supply chain
models are presented.
7.3.1
Forward Supply Chain Models
Beamon [22] presents a review on the design and analysis of supply chains. Four
categories of models are analyzed: deterministic analytical models where all the
variables are known and specified, stochastic analytical models where at least one
variable is unknown and follows a probability distribution, economic models, and
simulation models. As future research, the author proposes four research areas.
Two are related with the definition and the integration of performance measures
in supply chain models, and the other two, concern the systematization of issues
that affect supply chains and the classification of the associated systems.
Later on, Min and Zhou [23] reviewed works in supply chain modeling. They ar-
gue that the supply chain is customer oriented, which brings new problems that are
ill-structured, strategic, and behavioral. These are considered by the Operations Re-
search community as “soft” issues. Traditional analytical tools are more concerned
with “hard” issues, which are more structural and technical, and thus easily trans-
lated by constraints. The authors conclude that there is a need for diversification
of the analytical techniques used for supply chain modeling. As future research
directions, they propose the development of models that integrate different supply
chain functions (e.g., production/distribution, production/sourcing, . . .), in more
complex structures (multi-echelon, multi-product, . . .). They also state that there is
a need for the development of multi-objective models that will be able to evaluate
real management trade-offs.
Goetschalckx et al. [24] analyze the work developed in the strategic design of
global supply chains. The authors argue that there are no analytical tools or
methodology to support “global, integrated, dynamic, and stochastic decisions” that
are common in global supply chain management. These decisions have been an-
alyzed in a “hierarchical-and-sequential manner” leading to smaller profit results
than if they were supported by adequate integrated strategic and tactical models. As
a main conclusion, the authors state that existing models have not addressed some
important characteristics of global supply chains such as exchange rates, transfer
prices, duties, and taxes amongst others.
More recently, Shah [25] reviews published work within the supply chain opti-
mization for the process industry. The author splits its review considering three
categories of SC problems: (i) network design, (ii) analysis and policy formulation,
and (iii) planning and scheduling. This review is very extensive as the author not
only refers works that use mathematical programming techniques, but also several
simulation techniques. The review is kept within the process systems engineering
area. He points out some particular features of process industry, namely flexibil-
ity, responsiveness of the production process, and the effective design of supply
chains “for the future”. In this latter group, the author mentions the need of con-
sidering the recovery and the remanufacturing of products within the supply chain
structure and operation.
Klose and Drexl [26] review the location/allocation models for supply chain
design. Several formulations are presented, discussed, and compared. For each
model, the authors also discuss complexity problems and algorithms that are used
to solve them. Although this review had as primary objective to discuss models for
distribution systems design, the authors also analyzed several other applications
that might fit the proposed formulations.
Meixell and Gargeya [27] performed a review on decision support models for the
design of global supply chains. The authors conclude that although the existing
models deal with some of the most difficult global supply chain issues, few models
address outsourcing, integration, and strategic alignment in global supply chain
design. Regardless of the fact that not a single work concerning reverse flows is
referred along the chapter, the emerging fields in global supply chains should also
be included in the research agenda of reverse supply chain.
7.3.2
Reverse Supply Chain Models
In 1997, Fleischmann et al. [28] published a very good review on quantitative model
for reverse logistics. They classified papers in three basic domains: distribution
planning (strategic models), inventory management, and production planning (tac-
tical/operational models). Concerning the distribution planning, authors state that
the interaction between the forward and reverse flows adds complexity to the sup-
ply chain management. They conclude that research on this matter has focus on
single issues rather than integrated analysis. This work is taken as reference and
therefore our work analyzes the published work after 1997.
Ammons et al. [29] developed an MILP model for the design and planning of a re-
verse production system for used carpets. The model considers a two-echelon net-
work with carpet collection sites, processing sites, and end-use points. The model
is a single product warehouse location model. A sensitivity analysis is performed
in order to evaluate the impact of several parameters on the final results. Namely,
the impact that collection fees have on the collected volume, the impact of sort-
ing costs, the attractiveness of opening additional processing sites and/or a raw
material recycling plant are analyzed. This was the first of a series of works devel-
oped by the authors on reverse production systems applied to carpet recycling. In
2000, Ammons et al. [30] proposed a new model for a very similar problem. Models
differ mostly in their formulation as this second one is based on the State-Task Net-
work (STN) formulation developed by Kondili et al. [31]. This model was applied to
a second case, the recycling of network router from an OEM (Original Equipment
Manufacturer). Again, this work was complemented with some sensitivity analysis.
In the same year, a different approach to this problem was proposed by Realff et al.
[32]. Using a robust optimization framework the objective was, as described by the
authors, “the minimization of the maximum deviation of the performance of the

network from the optimal performance under a number of different scenarios”.
Krikke et al. [33] present a mixed integer linear programming (MILP) model
for the re-design of the reverse logistics network for a particular copying machine
(Océ). The model was tailor made for this problem. The network was designed
considering three separate scenarios, and results were compared with three other
designs provided by the company managers. Among the decisions were processing
volumes for each facility, flow amounts, facilities supplying plans, and inventory
volumes in all network nodes. The authors stated that although only the reverse
chain was designed it could be useful to develop closed-loop models for an integral
optimization of the network.
Shih [34] develops an MILP model for the design of recycling network of home
appliances in Taiwan. Six scenarios were created and solved by the model in order
to evaluate the impact of different take back rates and of operational storage condi-
tions. This work was motivated by an EPR legislation. A four-echelon network was
modeled, considering collection points, storage sites, disassembly/recycling plants,
and material market or disposal option. Only collection sites had fixed location. All
other facilities were subjected to a choice. The model was based on a capacitated
location-allocation model. Most used data had to be estimated since this problem
was new in Taiwan.
Sodhi and Reimer [35] proposed a group of models for the recycling of end-of-life
electronic products, together with models that included recovery decisions and the
dismantling process integration. Each model assumed a different decision maker
within the supply chain context. So, this work did not address the supply chain op-
timization as a whole, but it rather proposed solutions focusing on single partners.
Considerations were made concerning the solving techniques that better suited
each model.
Jayaraman et al. [36] proposed both a strong and a weak formulation for the
reverse distribution problem. Their research addressed the design of reverse net-
works and considered product recall, product recycling and reuse, product disposal,
and hazardous product return. The weak formulation, in particular, was solved us-
ing a heuristic procedure, specially developed for this model. The authors stated
that this heuristic found optimal solutions for a significant proportion of problems
and, within a reasonable amount of time, it also solved a very large number of other
problems that conventional MIP tools could not handle.
Schultmann et al. [37] studied the problem of recycling spent batteries from
steelmaking plants in Germany. Based on facility location–allocation problem, the
model designed a two-echelon reverse network considering three main activities:
collection, sorting, and recycling. Among the decisions are the creation of new
sorting facilities or the capacity expansion of existing ones. Two scenarios based on
the type of collected batteries are set. The results show that the reverse network is
strongly influenced by the recovery/disposal options taken.
Le Blanc et al. [38] study the problem of redesigning a recycling system for LGP
tanks. In this case study, two alternatives are analyzed: a central facility and a set of
regional depots. The proposed model is an integer programming model based on
a standard location-allocation model where additional constraints were considered.

As costs are some time difficult to determine, the authors applied a heuristic pro-
cedure to estimate them as accurately as possible. As a conclusion it is stated that
although the results were “quite robust, the effects of estimation errors in the cost
can be large at an absolute level”.
Listes and Dekker [39] propose a stochastic approach to the case study of recy-
cling sand from demolition waste, in The Netherlands. The uncertainty in this case
is related to the demand sources and quality, i.e., from which locations the sand
to be recycled should originate and its characteristics. The scenario approach was
used to extend a previous published work. The authors concluded that “given the
existing computational power and using an adequate modeling it is nowadays pos-
sible to apply stochastic programming techniques to practical situations of logistic
networks design”.
7.3.3
Closed-Loop Supply Chain Models
Having analyzed the forward and the reverse supply chains we are now going to
focus on the design and planning of the integration of both structures – the closed-
loop supply chains. These are referred by Guide et al. [5] to require a careful design,
planning, and control due to its level of complexity.
Jayaraman et al. [40] developed a 0–1 mixed integer programming to model a
closed-loop supply chain. The network is modeled as a multiproduct warehouse
location model, with capacity constrains for a single period. Besides the location-
allocation features of the model, it also contains the stocking of the remanufactured
products. The model was tested on a set of problems based on the parameters
of an existing electronic equipments remanufacture firm. A sensitivity analysis is
performed to evaluate the impact that inbound and outbound costs have on the
network structure.
Fleischmann et al. [15] proposed a model for the location of logistic facilities. An
MILP formulation is proposed that constitutes an extension of the traditional ware-
house location problem where two such models are integrated: one for the forward
chain connecting factories to customers through warehouses, and the other for the
reverse chain connecting customers to factories, through disassembly centers. The
two chains are integrated by means of a balance constraint that assures for each
factory that its total return is not greater than its total production. The authors dis-
cuss the main differences between traditional logistics network and a combined
forward and reverse network. Two case studies already published were used to ex-
plore the model application and their analysis concludes that, for some cases, the
reverse chain must be accounted for simultaneously with the forward chain but,
for other cases, it is better to handle them separately. Their work was the first to
propose a general model formulation that simultaneously optimizes the reverse
and the forward networks.
Krikke et al. [41] proposes an MILP model for the design of a closed-loop supply
chain where both location-allocation decisions and product design are considered.
The objective function includes both supply chain costs and environmental im-
pacts using a performance indicator (based on LCA approach). As product design
is involved, an assembly and disassembly of products is explicitly modeled. The
model is single period. The product design is split in three major forms: modular-
ity, recyclability, and repairability. A sensitivity analysis is performed to evaluate the
relation between product and network design, and environmental issues.
In Beamon and Fernandes [42], a model for a single product closed-loop supply
chain design problem is proposed and used to analyze the impact of several para-
meters on the network structure. In order to reflect different costs structure, four
scenarios are created using data generated for the “purpose of experimentation and
analysis”. A sensitivity analysis is performed in several other parameters.
Fandel and Stammen [43] propose a strategic model for the supply chain de-
sign. The authors extend the traditional supply chain to account for the recycling
of products released by customers. Dynamic aspects of the network are modeled by
the use of a two-step time structure. Although promising, this work did not present
any application and, therefore, neither the adequacy nor efficiency of the proposed
approach to treat real world problems was tested, nor was the model proven as
solvable.
Salema et al. [44] study the problem of designing simultaneously the forward and
reverse networks. The proposed model is fairly general as it incorporates facility ca-
pacity limits, multiproduct, and uncertainty. A scenario approach is used to model
uncertainty. A case is presented and solved using standard B&B techniques, which
allowed the model generality to be corroborated within very satisfactory compu-
tational times. Later on, the same authors [45] proposed a second model where a
different approach was used to model flow variables. The previous case study was
adapted in order to fit the model, which allowed the comparison of both formula-
tions. As main conclusion the authors state that the second variable formulations
are more adequate to this kind of formulations as they originate faster computa-
tional results.
Salema et al. [46] propose a strategic and tactical model for the design and plan-
ning of supply chains with reverse flows. The authors considered the network de-
sign as a strategic decision, while tactical decisions are associated to production,
storage and distribution planning. The integration of these two kinds of decisions
is achieved by considering two interconnected time scales: a macro and a micro
time. The model was applied to an already published case study, using standard
Branch and Bound techniques. The obtained results confirm the model adequacy
to real problems.
Most of the previous models are strategic, tactical or both strategic and tactical
in nature. In terms of operational models concerning reverse logistics the authors
have focused their attention mainly into specific activities. In the last 10 years, at
least six surveys were published concerning disassembly [47–52]. This topic is of
extreme importance as it has a major impact not only on the efficiency of reman-
ufacturing/recovery processes but also in product design. Until very recently, en-
gineers and designers learned how to design products that were easily assembled.
However, as companies are becoming more aware of the need to recovery prod-
uct/components/materials, a new way of looking into product design is emerging

[53]. In terms of global structure operation, few works have been reported. Amaro
and Barbosa-Povoa [54] developed a very detailed model to the scheduling of closed-
loops supply chain where different product recovery policies are analyzed. A real
case study of the pharmaceutical industry has been solved.
7.3.4
Research Agenda
As referred above the increased global competition, with a parallel depletion of

natural resources lead to an expanding stricter environmental legislation. Also, an
increased awareness of the customers perception toward the need of incrementing
products return so as to recover as much as possible its added value to the environ-
ment is a fact. In this context, companies are realizing new business opportunities
associated with a life cycle approach to their products. To respond to this challenge
companies must invest on the design, planning, and operation of their logistic
structures, forward and the reverse supply chains, in order to take advantages of
all types of products returns and to include final environmental friendly disposal
structures and operations.
This context was reviewed previously and although the presented review is far
from being completed one can say that many aspects are still open for further
research in the area of the design and planning of closed-loop supply chains.
Most of the published work looks into the closed-loop supply chains as reflecting
the solution of particular case studies. As a result, the developed models appear
essentially as case study oriented rather then generic, which may restrict their ap-
plicability. Therefore, there is still a need for the development of generic formula-
tions where important aspects describing the closed-loop supply chains should be
considered. Some of these are described below:
1. Integration of international factors – within global
economies the international aspects within the logistics
structures must be accounted for. The usage of international
taxes as well as other economical or operational issues
should be included within the models formulations.
2. Integration of different levels of decisions – having in mind
that interactions between the different levels of decisions
exist in real cases when modeling CLCS, these must be
considered simultaneously in order to better sustain the final
decisions. Therefore, the formulations may result intractable
in terms of solution. Then there is a need for further
research on the algorithmic solution approaches.
3. Multiobjective decisions – closed-loop supply chains are
complex structures characterized by multiple objectives.
This situation must be evaluated within generic
multi-objective formulations that should allow different
objectives evaluation to be associated to the final decision.
4. Environmental impacts assessment – the models should

consider the flows and activities, with environmental
assessment providing the possibility of qualifying and
quantifying environmental aspects within the CLSC systems.
5. Uncertainty analysis – closed-loop supply chains as referred
above are subject to a high level of uncertainty. This should
be taken into account when modeling such systems since it
definitely will influence the final chain structure as well as
associated planning and scheduling activities. Also in this
case the associated computational burden will grow
accordingly and once more the investment in efficient
solution algorithms is required.
Having analyzed some of the possible aspects that should be considered when ad-
dressing the design and planning of CLSC this chapter proceeds with a proposal for
a generic model. This simultaneously establishes the CLSC structure and defines
the associated planning.
7.4
Closed-Loop Supply Chains Modeling
A generic model for the design and planning of closed-loop supply chains will be
presented. The problem will be firstly described followed by the definition of the
model formulation. The section will end with the solution of an illustrative case
study.
7.4.1
Problem Description
The system considered is defined in terms of four different agents that interact
amongst themselves (Fig. 7.2). These are factories, warehouses, customers and
disassembly centers. The forward flow links factories to customers through ware-
houses, whilst the reverse flow links customers to factories through disassembly
centers. No direct link is allowed between customers and factories. Customers are
characterized by a known demand that needs to be satisfied.
In the closed-loop context, products once used are collected, disassembled and
sent back to factories. The disassembly is performed in disassembly centers, where
subassemblies are assessed and sent to factories and/or disposal. Three different
groups of products actually flow in the network: forward products from factories to
customers through warehouses, reverse products from customers to disassembly
centers and subassemblies from disassembly centers to factories. Inventories are
allowed in all facilities and are limited to their maximum levels. Maximum and
minimum limits are also imposed on production levels and distribution flows.
Fig. 7.2 Network representation.
When designing and planning such systems, two levels of decision are consid-
ered: strategic level where the system structure is defined and a tactical level where
the planning of the resources is performed.
In this context, the model considers two levels of decisions within three inter-
connected time scales: a strategic time scale (time horizon) and a tactical time scale
(macro and micro time scales). Within the strategic time scale a time horizon (e.g.,
5 years) is defined over which all the facility locations are defined. Based on this
time horizon a macro time scale is defined which involves the division of the time
horizon provided into a set of intervals of equal duration (e.g., 1 year) where cus-
tomers’ demands and returns are satisfied (tactical decisions). These tactical deci-
sions are further enriched through the definition of a micro time scale where all
the CLSC activities are planned: production, storage, distribution, sorting and re-
covery. These time scales can be years/months, years/trimester, months/days or
whichever combination suits the problem. Travel times are modeled between net-
work levels. They are defined as the number of “micro” time units needed for a
product to flow from its origin to its destination. If travel times are set to zero then
a multiperiod location/allocation network model is obtained.
Finally, a cost function is minimized. Costs include investment and operational
costs and production, storage, distribution, disposal, and penalty costs (for non-
satisfied demand or return).
In short, the problem in study can be stated as follows:
Given:
• the investment costs,
• products’ bills of materials,
• the relation between forward and reverse products,
• travel time between each pair of interacting network agents,
• the minimum disposal fraction,
• forward product return fractions,
for each macro period and product:

• customer’s demand volume,
• the unit penalty costs for non satisfied demand and return,
and in addition, for each micro period and product:

• the unit transportation cost between each pair of interacting
network agents,
• the maximum and minimum flow capacities,
• the factory production unit costs,
• each facility unit storage cost,
• the unit disposal cost,
• the maximum and minimum production capacities,
• the maximum storage capacities, and
• the initial stock levels.
Determine the network structure, the production levels and storage levels, the flow
amounts and the non-satisfied demand and return volumes.
So as to minimize the global supply chain cost.
7.4.2
Time Modeling
As referred above the model developed considers two main levels of decisions: the
design level and the planning level. These involve the definition of three intercon-
nected time scales, a time horizon, a macro time scale, and a micro time scale as
depicted in Fig. 7.3. Consider H as the time horizon, t ∈ T an element of the macro
time set and t ∈ T an element of the micro time set. For each t ∈ T , n elements
exist in T . Both sets have to be ordered.
In order to establish the relation between the time scales, an operator is defined
where a current time unit is related to one that has already taken place (backward
t t+1
... ...
0
1 ... t'-1 t' ... n-1
0
Fig. 7.3 Interconnection between macro and micro time scales.

operator). It allows a connection between two different time units, within the same
constraint. This is required when there is an event on the chain that starts in a
certain macro period and continues over the next macro period.
Consider t ∈ T and t ∈ T , let (t, t ) be the current time instance. Suppose that
one wants to relate the current time unit with one that has occurred τ micro time
units before. The backward time operator ϒ is defined as

(t, t − τ ), if t − τ 0
ϒ(t, t − τ ) =
(t − ω, ωn + t − τ ), if t − τ < 0 ∧ t > ω,

where ω ∈ Z is the smallest integer greater or equal than (τ −t )/n, ω = [(τ −t )/n]
and n = #T .
For instance, let T = {1, 2, . . . , 12} and T = {0, 1, . . . , 21}, and suppose that the
current time unit is (t, t ) = (6, 10). Consider τ = 5 then ϒ(6, 10 − 5) = (6, 5) as
t − τ 0. Suppose now τ = 30, then ω = 30−10 22 = 1. Thus,
ϒ (6,10 − 30) = ϒ (6 − 1,1 × 22 + 10 − 30) = (5, 2).

To simplify notation and unless there is ambiguity, the generic micro time element
(t, t ) will be from now on denoted, simply by t . Note that although not written
explicitly, all the variables and or parameters defined within t are associated with
a macro time t.
7.4.3
Model Formulation
Based on the problem description presented a generic model for the design and
planning of a closed-loop supply chain is developed. This involves a set of indices
parameters and variables as given in the appendix.
All continuous variables are non-negative.
All the variables are represented schematically in Fig. 7.4 where the four echelons
of the supply chain are shown.
Apart from these variables a set of auxiliary variables, which are not used as
model decisions variables, are also defined to facilitate the constraint formulation.
These are the stock variables at customer sites and the dummy binary variables
associated with flows. The former describe the amount of product mr kept by cus-
tomer k, over micro period t and is represented by Sm c
r kt
. The latter are used to
define a set of disjunctive constraints. The following auxiliary variables are then
defined:
f
Eij1t = 1, if the flow between factory i and warehouse j exists, at micro time
t ; 0 otherwise,
f2
Ej kt = 1, if the flow between warehouse j and customer k exists, at micro
time t ; 0 otherwise,
r1
Eklt = 1, if the flow between customer k and disassembly center l exists, at
micro time t ; 0 otherwise, and
r2
Elit = 1, if the flow between disassembly center l and factory i exists, at
micro time t ; 0 otherwise.
X mf1f ijt ' X mf2f jkt '

a
Y j
Yi p Ykc
Yl r
X r2
mc lit ´
X mr1r klt '
Intermediate
Factory Customer
facility
Fig. 7.4 Schematics of the interdependence between flow and binary variables.
Finally, ϒ(t, t ) is a time operator that establishes the connection between macro
and micro time units as defined in the Section 7.4.2.
7.4.3.1 Constraints
Constraints definition involves firstly the definition of the constraints related with
the macro times (strategic decisions) followed by the characterization of the con-
straints defined over the micro times (tactical decisions).
Demand Constraint Each customer k, in each macro period t, has a specific demand
for product mf that needs to be satisfied. In order to assure that this demand can be
(totally or partially) satisfied by all the inbound flows that reach customer k, during
the time interval t, the following constraint is written:
f
Xm2f j kϒ(t,t −δj k ) + Umf kt = dmf kt ∀mf , k, t. (7.1)
j ∈J t ∈T
Return Constraint The return constraint has an analogous expression. Each cus-
tomer k has, in each macro period t, an amount of used products that need to be
collected, which can be totally or partially satisfied by the outbound flows:
r f
Xm1r klt + Wmr kt = βmf mr Xm2f j kϒ(t,t −δj k ) (7.2)
l∈L t ∈T mf ∈Mf j ∈J t ∈T
∀mr , k, t.
Disposal Fraction As referred above the disposal option is considered within the
closed-loop. This is done through the definition of a fraction of recovered sub-
assemblies that are estimated not to meet the quality standards for recovery. Thus,
a part of the inbound flow of disassembly center l can be sent to a fictitious factory,
factory i = 0, that represents any facility out of this supply chain – corresponds to
a disposal option:
r1
r
γ βmr mc Xm
r klϒ(t,t −δkl )
Xm2c l0t ∀mc , l, t, (7.3)
mr ∈Mr k∈K t ∈T t ∈T
where γ is assumed as a global parameter, thus not product dependent. However,

it can easily be formulated as product dependent.
Having established, in the global chain, the relation between aggregated quanti-
ties and flows, it is important to relate the latter with the operational aspects, such
as production, storage and transportation.
Production Constraints These constraints are established for two main purposes:
to assure the connection between production factories and associated inbound and
outbound flows and to guarantee factory capacity production limits.
The production planning contemplated in this model is of a tactical nature and
thus no consideration is given to differences between manufacturing and remanu-
facturing. However, products’ bills-of-materials are considered, allowing for a more
detail plan for production. Nevertheless, factories may still have inbound flows,
which need to be handled. These are subassemblies that were sorted and inspected
by disassembly centers, which are assumed to go directly into production.
For each micro period t , subassembly mc and factory i, the following constraint
assures that the amount of new subassembly mc units sent by suppliers plus the
used ones sent by disassembly centers (inbound flow) and the existing inventory
(which was set in the previous micro period) is equal to the total of subassemblies
(new and used) needed to produce all forward products mf delivered by the factory
(outbound flow) plus the remaining stock. The connection between subassemblies
and final products is achieved by the use of parameter βmc mf that reflects final
product mf bill-of-materials. Note that each factory is assumed to produce only
what is sent as outbound flow (just-in-time production).
r p
Zmc it + Xm2c liϒ(t,t −δli ) + Smc iϒ(t,t −1)
l∈L

∀mc , i ∈ I, (t, t ).
f p
= βmc mf Xm1f ij t + Smc it (7.4)
mf ∈Mf j ∈J
If ϒ(t, t − 1) = (1, 0), then Smc iϒ(t,t −1) = smc i0 which is the initial stock level at
p p
factory i.
Factory Capacity Constraints In any factory i, the production level is, in general,
bounded. These bounds are assumed constant for each factory over the model
horizon, however, they may differ amongst factories. The production capacity is
defined in terms of amounted produced over time t . Furthermore, a factory can
only be used if installed. Constraints (7.5) and (7.6) model these situations:

Xm1f ij t gi Yi ∀i, (t, t ),
f p p
(7.5)
mf ∈Mf

∀i, (t, t ).
f p p
Xm1f ij t hi Yi (7.6)
mf ∈Mf
Acquisition contracts are also bounded. Thus, constraint (7.7) defines a minimum
supplying volume for each factory.

Zmc it hit Yi ∀i, (t, t ).
p
(7.7)
mc ∈Mf
s
Also, as storage is allowed in factories, an upper bound is set (gi p ) for the storage,
since it is assumed that no product is stored but only the raw materials (subassem-
blies):
p s
Smc it gi p Yi , ∀i, (t, t ).
p
(7.8)
mc ∈Mc
Warehouse Storage Constraints At the warehouses, a material balance is also estab-

lished where the inbound flow plus the existing stock equals the outbound flow
plus the remaining stock. This must be assured in every micro period, for every
existing product:
f f
a a
Xm1f ij ϒ(t,t −δij ) + Sm
f j ϒ(t,t −1)
= Xm2f j kt + Sm f jt
, (7.9)
i∈I k∈K
∀mf , j, (t, t ).
Again, if ϒ(t, t − 1) = (1, 0), then Sm a

f j ϒ(t,t −1)
= sm
a
f j0
which is the initial
stock of product mf kept in warehouse j .
Furthermore, every warehousej has a maximum storage capacity (gjsa ), if in-
stalled (i.e., Yja = 1).

Sma
f jt
sa a
gj Yj , ∀j, (t, t ). (7.10)
mf ∈Mf
Disassembly Center Storage Constraints As previously mentioned, products’ bill-of-

materials are modeled. The disassembly centers collect, disassemble, and inspect
used products released by customers. The subassemblies that are still in good con-
dition are sent for remanufacturing. The remainder are sent to proper disposal,
which is assured by constraint (7.3).
Equation (7.10) establishes the material balances, which relate storage volumes
with the inbound/outbound flows. Thus, in each period t , the inbound flow and
the existing storage must equal the outbound flow plus the new storage volume.
r1 r
βmr mc Xm
r klϒ(t,t −δkl )
+ Sm
c lϒ(t,t −1)
mr ∈Mr k∈K

= r2
Xm c lit
r
+ Smc lt , ∀mc , l, (t, t ). (7.11)
i∈I ∪{0}
Again, if ϒ(t, t − 1) = (1, 0), then Sm

r

c lϒ(t,t −1)
= sm
r
r l0
, which represents the
initial stock level.
For each disassembly center l, a maximum storage capacity (glsr ) is also consid-
ered if the facility is used (Ylr = 1).

r
Sm c lt
sr r
gl Yl , ∀l, (t, t ). (7.12)
mc ∈Mc
Network Balance This model can be seen as generalization of two interdependent

warehouse location/allocation models, where both networks are connected. Focus-
ing on customers, one must assure that each of them, in every micro time instance,
cannot return more products than the amount received until then. Note that it was
assumed that a customer cannot introduce in the supply chain products originated
from a different supply chain.
A “material balance constraint” on customers is therefore defined, assuring that
for each customer and each return product mr in each micro period, the inbound
flow plus the existing “stock” must equal the outbound flow plus the new created
“stock”.
f c
βmr mf Xm2f j kϒ(t,t −δj k ) + Sm
r kϒ(t,t −1)
mf ∈Mf j ∈J

= r1
Xm r klt
c
+ Sm kt ,
r
∀mr , k, (t, t ). (7.13)
l∈L
If ϒ(t, t −1) = 0, then Sm

c

r kϒ(t,t −1)
= sm
c
r k0
, which is the initial stock on customer.
Transportation Flows Constraints Within the transportation flows, both maximum

and minimum capacities exist. The latter models the case of the minimum
amounts required to be transferred, when the flow is set to occur (E = 1).
Flows between factories and warehouses are limited to upper and lower bounds:

Xm1f ij t gij1t Eij1t ∀i, j, (t, t ),
f f f
(7.14)
mf ∈Mf

∀i, j, (t, t ).
f f f
Xm1f ij t hij1t Eij1t (7.15)
mf ∈Mf
Flow between warehouses and customers have also upper and lower bounds:

∀j, k, (t, t ),
f f2 f2
Xm2f j kt gj kt Ej kt (7.16)
mf ∈Mf

∀j, k, (t, t ).
f f f
Xm2f j kt hj 2kt Ej kt
2
(7.17)
mf ∈Mf
The same applies to flows between customers and disassembly centers

r r1
Xm1r klt glkt r1
Elkt ∀l, k, (t, t ), (7.18)
mr ∈Mr
r1
Xm r klt
r1 r1
hlkt Elkt ∀l, k, (t, t ), (7.19)
mr ∈Mr
and to flows between disassembly centers and factories.

r r2
Xm2c lit gilt r2
Eilt ∀i ∈ I ∪ {0} , l, (t, t ), (7.20)
mc ∈Mc
r2
Xm c lit
r2 r2
hilt Eilt ∀i ∈ I ∪ {0} , l, (t, t ), (7.21)
mc ∈Mc
Finally, for each network level, auxiliary constraints are defined to assure inte-
grality of the network. These state that if a certain site is not used, the occurrence
of the associated flows is precluded:
∀i, j, (t, t ).
f p
Eij1t Yi (7.22)
Constraint (7.21) enforces the flow to occur only if the factory site is chosen (when
p
Yi = 1). The same is assured by constraints (7.22)–(7.24) for the remaining flows.
∀j, k, (t, t ),
f a
Yj
2
Ej kt (7.23)
r1
Eklt Ylr ∀k, l, (t, t ), (7.24)
r2 p
Elit Yi ∀l, i, (t, t ). (7.25)
Objective Function The objective function is defined in terms of minimization of

costs.
The investment costs considered are related to the opening/usage of facilities
(factories, warehouses and disassembly centers) in each possible location. Thus,
we have:
p p
fi Yi + fja Yja + flr Ylr . (7.26)
i∈I j ∈J l∈L
On the other hand, the operational costs represent the costs faced by each type of fa-
cility when performing its core business. So for factories, raw material acquisition,
production, and storage costs are considered,
p f
r
cm Zmc it + cmf it Xm1f ij t
c it
mc ∈Mc i∈I t ∈T mf ∈Mf i∈I t ∈T j ∈J
s p
+ cmpc it Smc it , (7.27)
mc ∈Mc i∈I t∈T
whilst storage costs are defined for warehouses and disassembly centers:
s s
a r
cmaf j t Sm + cmr r lt Sm . (7.28)
f jt r lt
mf ∈Mf j ∈J t ∈T mr ∈Mr l∈L t ∈T
Forward and reverse transportation costs are also considered:

f f
f f
cm1f ij t Xm1f ij t + cm2f j kt Xm2f j kt , (7.29)
mf ∈Mf i∈I j ∈J t ∈T mf ∈Mf j ∈J k∈K t ∈T
r1 r1
r2 r2
cm Xm klt + cm Xm lit , (7.30)
r klt r c lit c
mr ∈Mr k∈K l∈L t ∈T mc ∈Mc l∈L i∈I0 t ∈T
Extra costs, referred as penalty costs, are also considered. These translate the
penalty for a non-satisfied demand or return:

u w
cm U
f kt mf kt
+ cm r kt
Wmr kt . (7.31)
mf ∈Mf k∈K t∈T mr ∈Mr k∈K t∈T
Summing up, the objective function is the minimization of all costs defined above.
7.4.4
European Case
This case was created based on a company that operates in Europe (adapted from
a previous published case, Salema et al. [46]). This company aims to determine the
network design for a supply chain that will involve three forward products (F1 , F2
and F3 ), two return products (R1 and R2 ) and four subassembly components (C1 ,
C2 , C3 and C4 ).
At the strategic level customers are grouped into 28 clusters, where each cluster is
named after the city it represents. Customers’ clusters (from now on will be desig-
nated simply as customers) are located in Amsterdam, Barcelona, Berlin, Brussels,
Copenhagen, Dublin, Düsseldorf, Essen, Frankfurt, Glasgow, Hamburg, Helsinki,
Lille, Lisbon, Liverpool, London, Lyon, Madrid, Nuremberg, Oslo, Palermo, Paris,
Rome, Rotterdam, Stockholm, Turin, Valencia, and Vienna.
Twelve of these cities are possible sites for warehouses and/or disassembly cen-
ters location: Amsterdam, Barcelona, Berlin, Brussels, Copenhagen, Essen, Ham-
burg, Lille, Liverpool, Paris, Turin, and Vienna. For factories, there are only three
possible locations: Amsterdam, Brussels, and Vienna.
In terms of time, a time horizon of 5 years is considered that is divided into equal
amounts of 1 year (macro). Each macro time unit is divided into a micro period of
four trimesters: macro period = “year” and micro period = “trimester”. Since the
model considers a horizon of 5 years, some data has to be estimated. These include
the demand volumes as well as variations in costs over the years. These estimations
were based on some assumptions:
1. transportation costs are proportional to the distance between
each city;
2. after the first year an actualization rate of 3% (or some other
convenient value) is applied to all costs;
3. in the first year, customers’ demand is made equal to a
fraction of the city inhabitants (a value between 0.04 and
0.055) while for the remaining years, this value is modified
by a variation factor (ranging from 0.98 to 1.05), allowing for
an increase or decrease in the demand volumes.
4. after use, products F1 and F2 are returned as R1 and product
F3 as R2 . In terms of return fractions, only 60% of F1 is
collected while F2 and F3 have a return fraction of 80%;
5. zero initial stock levels are assumed;
6. and the disposal fraction is set to 0.1, which means that at

least 10% of returns have to undergo disposal.
Minimum and maximum capacities are defined for production (1 ×106 and 2.5
×106 units, respectively). Maximum and minimum limits are imposed on all flows
in the network. A minimum level is imposed on acquisition volumes of new com-
ponents. Travel time is set to nil, which seems a reasonable assumption given
the chosen time scale (years/trimester) and the particular geographical area under
study.
7.4.4.1 Results
The resulting MILP model was solved through the use of GAMS/CPLEX (built
21.1), in a Pentium 4, 3.40 GHz.
The model is characterized by 54,617 variables (11,027 binary) and 37,710 con-
straints, and took about 4 CPU hours to be solved for a 6% optimality gap. The
optimal value found is 3.5 ×109 currency units.
The optimal network (Fig. 7.5) is characterized by two factories located in Brus-
sels and Amsterdam. Six warehouses are opened and located in Barcelona, Brus-
sels, Essen, Hamburg, Paris, and Turin. Seven disassembly centers are opened.
Four locations match the locations of warehouses: Barcelona, Essen, Paris, and
Turin. The remaining three are Amsterdam, Berlin, and Copenhagen.
Amsterdam factory supplies Essen and Hamburg warehouses and receives
returns from Amsterdam, Berlin, Copenhagen, and Essen. Brussels supplies
Barcelona, Brussels, Paris, and Turin and receives returns Barcelona. Essen, Paris,
and Turin. For the sake of simplicity, these connections are not shown in Fig. 7.5.
The optimal forward and reverse networks are presented in Figs. 7.5(a) and (b),
respectively. One can observe that factories have very close locations. Due to the way
transportation costs were estimated, those two sites act as geographical centers.
Fig. 7.5 (a) Forward network and (b) reverse network.

2500000
2000000
1500000 F3
Units
F2
1000000 F1
500000
0
12 3 4 1 2 3 4 1 2 3 4 1 2 3 4 12 3 4 12 3 4 12 3 4 12 3 4 12 3 4 12 3 4
1 2 3 4 5 1 2 3 4 5
Amsterdam Brussels
Factory / Year / Trimestre
Fig. 7.6 Production plan of both factories.
Both networks assure that all 28 customers have their demand and return satis-
fied.
Concerning the tactical level of decision four different analyses can be made,
respectively, for acquisition, production, storage, and distribution. As the model
produces a large amount of information, only some examples will be presented.
In terms of production, all three products are produced in both factories. In
Fig. 7.6, the production plan for the complete time horizon is shown. Brussels
factory is larger than the Amsterdam one. There are some trimesters that Brussels
factory produces at its maximum established limit, while Amsterdam factory has
its production leveled to the minimum capacity.
The acquisition plan of new components for both factories is shown in Fig. 7.7.
Amsterdam keeps this supplying to the minimum contractual level. Brussels fac-
tory is much less regular. These components together with the ones that came
from disassembly centers are assembled creating the production shown in Fig. 7.6.
Regarding inventories, a zero stock policy is foreseen in all facilities.
An example of the distribution plan between the Amsterdam factory and its
warehouses is shown in Fig. 7.8. Although not imposed, this flow occurs in all
trimesters and most of the times it is greater than the minimum flow level im-
posed. Only three times the minimum flow has happened, and it occurred only in
the flow to Hamburg in the first and fourth year.
In Fig. 7.9, one can see the flows from Barcelona warehouse to its customers.
The most striking supplying plan is the one to Madrid. This customer is only sup-
plied with products F2 from this warehouse. The major customer is Barcelona. It
is supplied every trimester and frequently this is of a single product. In three out
of the 5 years, Lisbon customer is supplied only three times. The main reason is
the distance between both locations.
8000000
6000000 C4
Units
C3
4000000 C2
C1
2000000
0
12 3 4 12 3 4 12 3 4 12 3 4 12 3 4 12 3 4 12 3 4 1 2 3 4 1 2 3 4 12 3 4
1 2 3 4 5 1 2 3 4 5
Amsterdam Brussels
Factory / Year / Trimestre
Fig. 7.7 Supply of new components.
1000000
750000
F3
Units
F2
500000
F1
250000
0
12 3 4 1 2 3 4 12 3 4 12 3 4 1 2 3 4 12 3 4 12 3 4 1 2 3 4 12 3 4 12 3 4
1 2 3 4 5 1 2 3 4 5
Essen Hamburg
Warehouse / Year / Trimestre
Fig. 7.8 Flow between Amsterdam factory and the warehouses.
Figure 7.10 shows the supplying plan of Madrid customer. As mentioned above,
this customer is supplied as product F2 supplied by Barcelona, and F1 and F3 by
Paris warehouse. Due to the distance, in some years there is one trimester where
there is no shipment.
In Fig. 7.11, another customer supplying plan is shown. London is supplied by
Brussels warehouse, and in each trimester for the entire time horizon some of the
demand is met. This supplying plan is very different from the previous one.
Two very different collection plans are shown in Figs. 7.12 and 7.13. The former
one belongs to Lyon customer and the latter to Brussels customer. Lyon has its
returns collected by two different disassembly centers, in one product/one center
relation. All R1 is sent to Turin whilst R2 is sent to Paris.
400000
300000
F3
Units
F2
200000
F1
100000
0
1234123412341234123134123123412341231231234123412341231234123412341234123
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5
Barcelona Lisbon Madrid Valencia

Customer / Year / Trimestre
Fig. 7.9 Flow between Barcelona warehouse and its customers.
280000
210000
F3
Units
F2
140000
F1
70000
0
1 2 3 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 1 2 3 4 1 2 3 13 4 1 2 4 1 2 3 4
1 2 3 4 5 1 2 3 4 5
Barcelona Paris
Customer / Year / Trimestre
Fig. 7.10 Supplying plan of Madrid customer.
Brussels customer has its returns collected only by Amsterdam. The number of
times returns are collected vary from once to four time a year. The return volume
of product R1 is larger than R2 . The reason is that product F1 and F2 are returned
as product R1 , whilst product R2 is the return of product F3 .
Figure 7.14 shows the supplying plan of Barcelona disassembly center. The vol-
ume of disposed products is much larger than the one that is sent to Brussels
factory. The reason is mainly the distance between both locations. However, 100%
of components C2 and C4 are disposed, and for C1 and C3 , the volume is of 10%
(which is the assumed disposal fraction). The reasons are that both components C1
and C3 have smaller transportation costs than the other two, and new components
of type C3 have the higher acquisition value.
400000
Units F3
F2
200000 F1
0
1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4
1 2 3 4 5
Brussels
Warehouse / Year / Trimestre
Fig. 7.11 Supplying plan of London customer.
60000
Units
40000 R2
R1
20000
0
2 3 4 12 3 4 12 3 4 12 3 4 12 3 4 12 3 4 12 3 4 1 2 3 4 13 4 12 3 4
1 2 3 4 5 1 2 3 4 5
Paris Turin
Disassembly centre / Year / Trimestre
Fig. 7.12 Collection plan of Lyon customer.
7.5
Conclusions
A new way of looking into supply chain has been emerging. The traditional defini-
tion of the chain that starts with raw materials and ends up in end-user consump-
tion is changing to include also the return to factories of end-of-life products. In
this context both reverse and closed-loop supply chains have appeared as new ways
of looking into the supply chain systems. Such systems are however quite complex
and should be carefully designed, planned and controlled. This implies the need of
efficient decision making tools to help the decision making process associated.
7.5 Conclusions 213
90000
60000
Units
R2
R1
30000
0
4 1 4 1 2 3 4 1 3 4 1 2 3 4
1 2 3 4 5
Amsterdam
Dis. centre / Year / Trimestre
Fig. 7.13 Collection plan of Brussels customer.
1500000
C4
1000000
Units
C3
C2
C1
500000
0
1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4 2 3 4 2 3 4 2 3 4 1 2 4 12 3
1 2 3 4 5 1 2 3 4 5
Disposal Brussels
Destination / Year / Trimestre
Fig. 7.14 Outbound flows of Barcelona disassembly center.
In this chapter, a review on the work developed on the design and planning of
both reverse and closed-loop supply chains was undertaken. As a main conclusion,
it can be stated that further research is required in order to create generic model
formulations that can be used actively in the type of decisions under study. A num-
ber of topics not yet enough explored should be further investigated such as the
integration of international factors; the integration of different decisions levels; the
analysis of multiobjective decisions; the incorporation of environmental impacts
assessment; last but not the least, uncertainty analysis.
Based on this, a generic model formulation for the design and planning of
closed-loop supply chain was presented. This considers the simultaneous analysis
of both strategic and tactical decisions where the system structure is defined taking
into account the planning of the related logistic activities. The model presented,
although reasonably generic, does not contemplate some of the points referred
previously that should be added.
In conclusion, closed-loop supply chain is a challenging area that opens up a
new and interesting set of issues to be addressed by both academia and industry
where the interdisciplinary nature of the problems involved should lead to generic
frameworks.
Acknowledgment
The authors gratefully acknowledge the support of the Portuguese National Science
Foundation through the project POCTI/AMB/57566/2004.
References
1 Goetschalckx, M., Vidal, C. J., Do- 8 Gupta, T., Chakraborty, S., Looping
gan, K., Modeling and design of in a multistage production system.
global logistics systems: A review Int. J. Prod. Res. 22 (1984), pp. 299–
of integrated strategic and tactical 311.
models and design algorithms. Eur. J. 9 Inderfurth, K., Janiak, A., Kovalyov,
Oper. Res. 143 (2002b), pp. 1–18. M. Y., Werner, F., Batching work and
2 Savaskan, R. C., Van Wassenhove, L. rework processes with limited deteri-
N., Reverse channel design: The case oration of reworkables. Comput. Oper.
of competing retailers. Manage. Sci. 52 Res. 33 (2006), pp. 1595–1605.
(2006), pp. 1–14. 10 Vlachos, D., Dekker, R., Return han-
3 Stock, J. R., Reverse Logistics, Oak dling options and order quantities for
Brook, IL, 1992. single period products. Eur. J. Oper.
Res. 151 (2003), pp. 38–52.
4 Dowlatshahi, S., Developing a the-
ory of reverse logistics. Interfaces 30 11 OECD, EPR programme implementa-
(2004), pp. 143–155. tion and assessment. In: Proceedings of
OECD Seminar on Extended Producer
5 Guide, V. D. R., Harrison, T. P., Van Responsibility, 2003.
Wassenhove, L. N., The challenge of
12 Hanisch, C., Is Extended Producer
closed-loop supply chains. Interfaces
Responsibility Effective? Environ. Sci.
33 (2003), pp. 3–6.
Technol. 34 (2000), pp. 170A–178A.
6 de Brito, M., Dekker, R., A Frame-
13 Thierry, M., Salomon, M., van
work for Reverse Logistics, in: Quan-
Nunen, J. A. E. E., Van Wassenhove,
titative Approaches to Reverse Logistics,
L. N., Strategic issues in product re-
Dekker, R., Inderfurth, K., Van
covery management. California Man-
Wassenhove, L. N., Fleischmann,
age. Rev. 37 (1995), pp. 114–135.
M. (eds.). Springer, Berlin, Germany,
2004, pp. 3–27. 14 Geyer, R., Jackson, T., Supply loops
and their constraints: The industrial
7 de Brito, M., Managing Reverse Logis- ecology of recycling and reuse. Califor-
tics or Reversing Logistics Management, nia Manage. Rev. 46 (2004), pp. 55–73.
Ph.D. Thesis. Erasmus University of
15 Fleischmann, M., Beullens, P.,
Rotterdam, Rotterdam, 2003.
Bloemhof-Ruwaard, J. M., Van
References 215
Wassenhove, L. N., The impact of 27 Meixell, M. J., Gargeya, V. B., Global

product recovery on logistics network supply chain design: A literature re-
design. Product. Oper. Manage. 10 view and critique. Transportation Re-
(2001), pp. 156–173. search Part E – Logistics and Transporta-
16 Toffel, M. W., Strategic management tion Rev. 41 (2005), pp. 531–550.
of product recovery. California Man- 28 Fleischmann, M., Bloemhof-
age. Rev. 46 (2004), pp. 120–+. Ruwaard, J. M., Dekker, R., van der
17 Farrow, P. H., Johnson, R. R., Lar- Laan, E., van Nunen, J. A. E. E., Van
son, A. L., Entrepreneurship, inno- Wassenhove, L. N., Quantitative mod-
vation, and sustainability strategies els for reverse logistics: A review. Eur.
at Walden Paddlers, Inc. Interfaces 30 J. Oper. Res. 103 (1997), pp. 1–17.
(2000), pp. 215–225. 29 Ammons, J. C., Realff, M. J., New-
18 Guide, V. D. R., van Wassenhove, ton, D., Reverse production system
L. N., Managing product returns for design and operation for carpet recy-
remanufacturing. Product. Opr. Man- cling. Working paper, Georgia Insti-
age. 10 (2001), pp. 142–155. tute of Technology, Atlanta, 1997.
19 Fleischmann, M., Krikke, H. R., 30 Ammons, J. C., Realff, M. J., New-

Dekker, R., Flapper, S. D. P., A char- ton, D., Decision models for reverse
acterisation of logistics networks for production system design. In: Hand-
product recovery. Omega-Int. J. Man- book on Environmentally Conscious
age. Sci. 28 (2000), pp. 653–666. Manufacturing, Madu, C. N. (ed.),
Kluwer, Bonton, 2000, pp. 341–362.
20 RLEC, Reverse Logistcs Executive
Council. In: http://www.rlec.org/ (last 31 Kondili, E., Pantelides, C. C., Sar-
access in 2006). gent, R. W. H., A general algorithm
for short-term scheduling of batch-
21 Fleischmann, M., Quantitative Mod- operations. Part 1. MILP formulation.
els for Reverse Logistics, Springer, Ger- Comput. Chem. Eng. 17 (1993), pp.
many, 2001. 211–227.
22 Beamon, B. M., Supply chain design 32 Realff, M. J., Ammons, J. C., New-
and analysis: Models and methods. ton, D., Strategic design of reverse
Int. J. Prod. Econ. 55 (1998), pp. 281– production systems. Comput. Chem.
294. Eng. 24 (2000), pp. 991–996.
23 Min, H., Zhou, G. G., Supply chain 33 Krikke, H. R., van Harten, A.,
modeling: Past, present and future. Schuur, P. C., Business case oce:
Comput. Ind. Eng. 43 (2002), pp. 231– Reverse logistic network re-design for
249. copiers. OR Spektrum 21 (1999), pp.
24 Goetschalckx, M., Nair-Reichert, 381–409.
U., Ahmed, S., Santoso, T., A review 34 Shih, L. H., Reverse logistics system
of the state-of-the-art and future re- planning for recycling electrical ap-
search directions for the strategic pliances and computers in Taiwan.
design of global supply chains. In: Resour. Conserv. Recyc. 32 (2001), pp.
MHRC, Portland, Maine, 2002a. 55–72.
25 Shah, N., Process industry supply 35 Sodhi, M. S., Reimer, B., Models
chains: Advances and challenges. for recycling electronics end-of-life
Comput. Chem. Eng. 29 (2005), pp. products. OR Spektrum 23 (2001), pp.
1225–1235. 97–115.
26 Klose, A., Drexl, A., Facility location 36 Jayaraman, V., Patterson, R. A., Rol-
models for distribution system de- land, E., The design of reverse dis-
sign. Eur. J. Oper. Res. 162 (2005), pp. tribution networks: Models and solu-
4–29. tion procedures. Eur. J. Oper. Res. 150
(2003), pp. 128–149.
37 Schultmann, F., Engels, B., Rentz, for the design and planning of sup-
O., Closed-loop supply chains for ply chains with product return, in:
spent batteries. Interfaces 33 (2003), Computer Aided Chemical Engineering,
pp. 57–71. Marquardt, W., Pantelides, C. (eds.),
38 Le Blanc, H. M., Fleuren, H. A., Vol. 21B, Elsevier, Amsterdam, 2006a,
Krikke, H. R., Redesign of a recycling pp. 2129–2134.
system for LPG-tanks. OR Spektrum 47 Zhang, H. C., Kuo, T. C., Lu, H. T.,
26 (2004), pp. 283–304. Huang, S. H., Environmentally con-
39 Listes, O., Dekker, R., A stochastic scious design and manufacturing: A
approach to a case study for product state-of-the-art survey. J. Manuf. Syst.
recovery network design. Eur. J. Oper. 16 (1997), pp. 352–371.
Res. 160 (2005), pp. 268–287. 48 Gungor, A., Gupta, S. M., Issues in
40 Jayaraman, V., Guide, V. D. R., Sri- environmentally conscious manufac-
vastava, R., A closed-loop logistics turing and product recovery: A survey.
model for remanufacturing. J. Oper. Comput. Ind. Eng. 36 (1999), pp. 811–
Res. Soc. 50 (1999), pp. 497–508. 853.
41 Krikke, H., Bloemhof-Ruwaard, J., 49 Erdos, G., Kis, T., Xirouchakis, P.,
Van Wassenhove, L. N., Concurrent Modelling and evaluating product
product and closed-loop supply chain end-of-life options. Int. J. Prod. Res. 39
design with an application to refriger- (2001), pp. 1203–1220.
ators. Int. J. Prod. Res. 41 (2003), pp. 50 Desai, A., Mital, A., Review of liter-
3689–3719. ature on disassembly algorithms and
42 Beamon, B. M., Fernandes, C., design for disassembly guidelines for
Supply-chain network configuration product design. Int. J. Ind. Eng.-Theory
for product recovery. Product. Plan. Appl. Practice 10 (2003), pp. 244–255.
Control 15 (2004), pp. 270–281. 51 Lambert, A. J. D., Disassembly se-
43 Fandel, G., Stammen, M., A general quencing: A survey. Int. J. Prod. Res.
model for extended strategic supply 41 (2003), pp. 3721–3759.
chain management with emphasis on 52 Kang, J. G., Xirouchakis, P., Disas-
product life cycles including develop- sembly sequencing for maintenance:
ment and recycling. Int. J. Product. A survey. Proceedings of the Institution
Econom. 89 (2004), pp. 293–308. of Mechanical Engineers Part B-Journal
44 Salema, M. I. G., Barbosa-Povoa, of Engineering Manufacture, Vol. 220,
A. P., Novais, A. Q., An optimization 2006, pp. 1697–1716.
model for the design of a capacitated 53 Bras, B., McIntosh, M. W., Product,
multi-product reverse logistics net- process, and organizational design
work with uncertainty. Eur. J. Oper. for remanufacture – an overview of
Res. 179 (2007), pp. 1063–1077. research. Robotics Comput.-Integrat.
45 Salema, M. I., Póvoa, A. P. B., No- Manuf. 15 (1999), pp. 167–178.
vais, A. Q., A warehouse based design 54 Amaro, C., Barbosa-Povoa, A., Close
model for reverse logistics. J. Oper. loops supply chains: Managing prod-
Res. Soc. 57 (2006b), pp. 615–629. uct recovery portfolios. In: Escape,
46 Salema, M. I., Barbosa-Póvoa, A. P., Marquardt, W., Pantelides, C. C.
Novais, A. Q., An integrated model (eds.), Vol. 16, 2006, pp. 1875–1880.
Appendix 217
Appendix
List of Symbols
Indices
i, j, l, k supply chain entities: factory, warehouse, disassembly center, and
customer, respectively
i=0 allows products to leave the supply chain – corresponds to a disposal
option
mf , mr , mc products: forward, reverse, and subassemblies
t, t time units: strategic and tactical, respectively
Sets
I, J, L, K potential locations for the supply chain entities: factory, warehouse,
disassembly center, and customer, respectively
Mf , Mr , Mc products: forward, reverse, and subassemblies, respectively
T,T strategic and tactical time sets, respectively
Parameters
γ is the minimal disposal fraction, that allows the network
to send some return products to an external processing
facility; it takes values in the range [0, 1]
p a r c
smf i0 , sm f j0
, sm r l0
, sm r k0
is the initial stock of product in each type of network en-
tity for time (t, t ) = (1, 0)
s
gi p , gjsa , glsr maximum storage capacity of factory i, warehousej , and
disassembly center, l, respectively
p p
gi , hi maximum and minimum production of factory i
hit minimum supplying volume to factory i, for macro pe-
riod t
f f
gi 1 , hi 1 upper and lower bound value for flows leaving factory i
gir2 , hri 2 upper and lower bound value for flows arriving at factory
i
f f
gm2f kt , hm2f kt upper and lower bound value for flows reaching cus-
tomer k, at time t
r1 r1
gm r kt
, hm kt
r
upper and lower bound value for flows leaving customer
k, at time t
δab travel time, i.e., the number of micro time units required
to travel from locations a to b
βmn bill-of-materials between product m and product n; this is
a matrix that relates all three types of products (forward,
reverse and subassemblies)
dmf kt demand of product mf to be supplied to customer k, over
macro period t
p
fi , fja , flr investment cost of factory, warehouse and i disas-
sembly center, respectively
u
cm f kt
unit variable cost of non satisfied demand of prod-
uct mf to customer k, for macro period t
w
cm unit variable cost of non satisfied demand of prod-
r kt
uct mr to customer k, for macro period t
r
cm unit variable cost of new sub-component mc ac-

c it
quired by factory i, at time t
p
cmf it unit production cost of product mf , manufactured
in factory i, at time t
s sa sr
cmpf it , cm f jt
, cm lt
r
unit storage cost of one unit of product kept in fac-
tory i, warehouse j , disassembly center l, at time
t
f f r1 r2
cm1f ij t , cm2f j kt , cm r klt
, cm lit
r
transportation cost of one unit of product from fac-
tory i to warehouse j , warehouse j to customer k,
customer k to disassembly center l, from disassem-
bly center l to factory i, respectively, at time t
Variables
p
Yi , Yja , Ylr , Ykc = 1 if factory i, warehouse j , disassembly center l, or
customer k integrates the supply chain; 0 otherwise
Umf kt (Wmr kt ) non-satisfied demand (return) of product mf (mr ) of
customer k, over macro period t
f f r1 r2
Xm1f ij t Xm2f j kt Xm r klt
Xm lit
c
amount of product that flows between two supply
chain entities, over period t
p a r
Smc j t , Sm f jt
, Sm lt
c
amount of material stocked in factory i,
warehousej , and disassembly center l and,
over period t
Zmc it amount of new subassembly components mc con-
sumed by factory i, over period t
219
8
Combining Strategic Design and Operative Planning in the
Process Industry
Josef Kallrath
In this chapter we address various conceptual issues in strategic, multisite design

and planning problems in the process industry, and particularly, their combination
in one combined model. This applies to the chemical and pharmaceutical industry,
food industry, but also parts of the metals industry. The design aspects include the
opening of new units and the shutdown of existing units (reactors, tanks) or whole
production sites to be purchased or sold in strategic design planning, outsourcing
or in-sourcing decisions, the mutual design of production facilities, the design of
distribution networks, the design of customer or product portfolios, and the finan-
cial design of global supply chains in which the transfer prices are variables. We
explain the need for combined design and operative planning models, favor indi-
vidual tailor-made modeling and solution approaches, and provide a model frame-
work and references to real world problems in which some of these features are
consistently integrated with an operative planning approach. Finally, we focus on
problems of supply chain planning and design in large companies related to the
needs of appropriate modeling and maintainable IT structures.
8.1
Introduction
Operative planning is most often understood as mid-term planning (a month to a

year) and involves various supply chain issues such as production planning, sales
planning, and shipment. The major focus is usually tactical planning – master
planning (cf. [41]) in the parlance of some advanced planning systems1) (APS) –
with horizons of a few months, followed by production planning and scheduling
(cf. [47]) (more precisely, short-term or detailed scheduling). We use it here more
in the sense of any kind of planning producing plans and schedules used oper-
1) APS may include demand planning (demand forecast), production planning, and scheduling, dis-
tribution, and transport. Planning and scheduling may also support optimization across multi-
warehouse, multiplant, multisite, and multicountry environments.
ISBN: 978-3-527-31693-9
220 8 Combining Strategic Design and Operative Planning in the Process Industry
atively in a production department or by marketing forces of a company. We as-

sume that it applies to an underlying multisite production network. It includes the
purchase of raw materials, the selection of suppliers, as well as distribution deci-
sions. We prefer to discuss it here from the perspective that operative planning
establishes the kernel and core model; it includes master planning and production
planning and, possibly, scheduling as defined in APS. Note that APS may distin-
guish master planning from production planning and scheduling; we do not. BASF
implemented such operative planning system in some production divisions long
before the existence of APS such as SAP APO already in the early 1990s. A typical
case of such a model is provided by Timpe and Kallrath [49] – for a given multi-
site network production, sales, shipment, and storage plans are computed for time
buckets within a time-discrete model formulation. Another example to mention is
Matsuoka and Mukari [35] who present a model for production planning for mul-
tisite batch plants.
While most APS contain strategic planning, planning, and detailed scheduling
as separate modules, our view is more holistic. We will give arguments for this
later. Independent of subtle differences in the definition we stress that operative
planning – Shah [46] calls this supply chain planning – considers a given and fixed
infrastructure over a time horizon of weeks, months, or quarters of a year to ex-
ploit the production, distribution, and storage resources in the supply chain in or-
der to respond to demand forecast or specific orders. The details of the production
process (rates, capacities, etc.) are also given and fixed. In short we define opera-
tive planning as determining production levels, raw material requirements, selection of
customers, and distribution schemes. Operative planning can also be characterized as
medium-term planning having a time horizon between a month and a year.
Strategic planning or multisite design planning comes in several variants. Goet-
schalckx [15] calls it strategic network planning with a focus on the long-term strate-
gic planning and design of the supply chain. His contribution is a pleasure to read
and gives a clear motivation why a supply chain is subject to redesign from time to
time: respond to market changes, or mergers and acquisitions producing a need to
consolidate duplicated elements in the supply chain.
Kouvelis and Munson [31] approach design planning to structure global supply
chains. Their mixed-integer linear programming (MILP) model involves the eco-
nomics of scale, complexity costs, transportation costs with and without tariffs and
is used to design the global network structure.
In the context of the chemical industry we find Shah [46] who calls it supply chain
infrastructure (network) design or redesign of the logistics network. A very elaborated
model is described by Lasschuit and Thijssen [32], which, at the strategic and global
planning level for a network of production plants, represents decisions on feed-
stock procurement and distribution, utilization of production capacities, utilization
of modes of transport and demand allocation. It is important to note and typical for
strategic planning that not only existing capabilities have to be considered, but also
new opportunities in all areas have to be evaluated. Their resulting mathematical
programming model is a mixed-integer nonlinear programming (MINLP) model:
integer aspects arise because of, for example, fixed costs/investment costs, tiered
pricing, and cargo costs. Nonlinear relations are mainly caused by multiplication of
quantity and economic variables. Ahmed and Sahinidis [1] present a mixed integer
programming (MIP) model for long-range chemical process planning. This model
combines investment decisions and operative features. A detailed mathematical
model of a combined strategic and operative planning in the process industry over
a time horizon of 15 years has been presented by Kallrath [24]. This model accounts
for chemical reactors of known capacity to be opened or shut down, and it accounts
even for sets of units belonging to a site subject to acquisition. Guillen et al. [16] si-
multaneously optimize process operations and financial decisions to enhance the
integrated planning/scheduling of chemical supply chains – they refer to the de-
sign or strategic part as financial or budget planning. They provide a mathematical
formulation combining a scheduling/planning model with a cash flow and budget-
ing formulation.
We note that design or strategic planning exists on its own, or with only little op-
erative planning aspects involved or, especially in the chemical process industry in
combination. We will favor the combined approach but at first we will briefly focus
on why the design and strategic planning and operative planning are separated.
This hierarchical view, strategic planning–master planning–production planning and
scheduling, is the basis for many implementations we find in APS, for example,
provided by SAP (SAP APO) or i2 Technologies. This hierarchical approach was
developed in the past because the combination of all aspects was too difficult to
handle. Nowadays, it is also often reflected in the companies where different de-
partments or project groups are responsible for strategic or design planning, oper-
ative planning or production planning (master planning the parlance of SAP APO),
and detailed scheduling. The usual approach to address these three time horizons
is a top-down methodology where the output of the master planning is used as in-
put for the detailed scheduling. Strategic and design planning is done more within
specific projects in which various scenarios considered in less details than in oper-
ative planning. The strategic design involves decisions with a long-lasting effect on
the supply chain of a company. This includes the supply chain design, i.e., comput-
ing the optimal network and the design of the embedded plants and facilities. The
disadvantage lies in the fact that the higher level has less details, i.e., constraints,
then the next level, and thus the more detailed level may become easily infeasible.
In the case of master planning (this is the tactical level producing long/medium
term management decisions typically updated at a rate ranging between once per
month, every quarter, and once every year resulting in overall production, purchas-
ing and sales decisions, inventory policies, and transport strategies) and detailed
scheduling (the deepest operative level referring to day-to-day decisions such as
sequencing, scheduling, tank issues, filling, and truck loading, etc.) one would ex-
pect benefits and more compatibility if either the master planning contains more
details (and thus being closer to detailed scheduling) or certain information from
scheduling is passed back to the master level planning.
Therefore, in this contribution we strongly support the idea that one develops
and runs a tailor-made mixed integer linear programming model involving the
design features, operative features, and, in the future, appropriate links to schedul-
ing as well. The scheduling problem itself should also be open to a design mode.
We call this the combined strategic and operative planning (CSOP) approach and
describe it in Section 8.2. We refer to the underlying MILP or MINLP model as
the SOP model which means strategic-operative planning model. Once somebody
comes up with an appropriate framework for integrated planning and scheduling
we would also include scheduling as discussed in Section 8.2.3.4. Algebraic model-
ing languages such as AIMMS, GAMS, LINGO, MPL, Mosel, or OPL studio are well
suitable to implement such models; cf. [26]. The SOP model could have a modular
structure which is well supported by the modeling languages. Finally, in Section
8.3 we discuss the appropriateness of commercial software and individual models
for the CSOP approach.
8.2
Combined Strategic and Operative Planning
In Section 8.1 operative planning and design/strategic planning have been briefly
reviewed separately. Especially, strategic planning is presented as an approach se-
lected on its own undertaken by the appropriate organizational units in companies.
We favor a combined strategic and operative planning (CSOP) approach based on
the SOP model, which involves strategic and operative planning decisions. The
starting point is an operative planning model in place with new strategic or deci-
sion questions coming up.
8.2.1
Operative Planning
In this contribution we define operative planning as any decision problem lead-

ing to production plans and schedule, sales plans, transport plans, etc. for given
demands, costs, prices, capacities of producing units and storage units within a
multisite production network. It is the purpose of operative planning to generate
these plans in such a way that the contribution margin, i.e., yield minus all vari-
able costs, is maximized subject to exact or partial demand fulfillment in time.
In this sense, operative planning may involve tactical planning (master planning)
over a few months but also at least aspects of scheduling (sometimes called detailed
scheduling) although this may work only for individual sites. Operative planning
supports the production process and operation within the production departments,
but also the planning activities in a marketing department solving the decision
problems which customers are delivered from which sales point, as well as how to
distribute the material flow within the multisite network. It includes the acquisition
of resources and raw materials, distribution decisions, and selection of suppliers.
Operative planning in the sense of our definition includes supply chain planning
and production planning. The operative planning model is the core model. Kallrath
[25] describes a generic model covering many features occurring in the chemical
and pharmaceutical industry, and even food industry.
8.2.2
Multisite Design and Strategic Planning
Preparing the proposition in Section 8.2.3, in this contribution we refer to multisite

design planning, or strategic planning, as any decision planning around an opera-
tive planning problem. It is a generalization of the operative planning core model
in the sense that fixed and given input data now might become variables – both
continuous and discrete variables. These new variables may represent the design
decisions with respect to the existence and capacities of the production and storage
units. It could also involve redesigning the topology of the multisite production
network, e.g., it may involve a depot-location problem for sales-storage sites. We
also consider product portfolio and customer portfolio problems as strategic plan-
ning problems. Another example of a strategic planning problem is the investment
decision of installing a new system of tanks. In production planning tanks play a
role only in terms of their capacity. In short-term scheduling we also need to con-
sider time aspects as charging time, extracting times, or extracting rates of the
tanks. Thus, this investment decision is a design decision around the underlying
short-term scheduling problem. A financial design problem is the after-tax profit
maximization problem in which the transfer prices are variables leading to MINLP
problems. They include bilinear terms because the transfer prices are multiplied
with the volumes transported.
Although it hardly can be a complete list of possible design decisions we try
to structure them as follows (some terms may appear both in the left and right
column):
Problem type Degrees of freedom/ decision variables
Capacity planning units, reactors, tanks, etc.

Facility location depots, warehouse, production sites
Financial design transfer prices, transport costs, total duties
Functional design production rates of reactors, process rates of filling stations
Portfolio optimization customers, products
Supply chain structure location variables for warehouses, sales points
Supply chain design topological variables, location variables
Topology connection of nodes in a network, possible routes of material
through nodes
Unit selection reactor shutdown/opening
Unit design capacity and production rates of the units
Unit investment new reactors, tanks, or filling stations
A final comment on strategic planning and design planning is about the ap-
proach sometimes found in strategic departments: scenario analyses. These may or
may not exploit mathematical optimization. The approach is rather based on simu-
lation in which the strategic experts of a company select appropriate scenarios to be
evaluated. If the approach is purely based on simulation there is no guarantee that
the optimal solution or even a solution close to the optimum will be found since
it covers only a part of the feasible decision space. This is especially troublesome
for complex problems, or those which require decisions of large financial impact.
This simulation approach only works if the decision variables can only take a few
values (ideally binary decisions), and if there are not too many decisions which can
be combined.
However, a valid way to exploit scenario analyses is in combination with mathe-
matical optimization. In a first step one would compute the optimum considering
all degrees of freedom. This is followed by a type of sensitivity analysis in which
some variables are fixed to realistic values in follow-up optimization runs to analyze
the consequences of allowing a nonoptimal fixing of a variable. This could be moti-
vated by the fact that a certain contract with a supplier cannot easily be cancelled, a
production plant cannot be sold immediately, a new production technology which
do the start-up could be more expensive than existing ones, or the choice of a sub-
optimal location because for the optimal location some contracting issue are still
unsolved.
8.2.2.1 The Need for Exact Optimization in Design Optimization

The results of strategic planning involving design decisions are expected to be im-
plemented and have an effect for a long time. Therefore, we should come as close
as possible to the exact optimum in order not to lose potential profit. One might
argue that the input data are subject to uncertainty and thus the model is not exact
anyway. However, in order to properly understand the implications of a model with
given data we should be sure that we do not lose information because we are not
close enough to the optimum.
Another argument is that for decision taking we should guarantee that we apply
the proper logic to compare competing scenarios (a scenario may be constructed
by some a priori fixed decisions or choice of parameters). In order to compare sce-
narios correctly, we need the exact optimum or at least safe bounds. The derivation
of safe bounds measuring the maximal possible distance from the exact optimum
is what distinguishes an exact optimization algorithm from improvement meth-
ods such as genetic algorithms, simulated annealing, or tabu search. To illustrate
the problem when using improvement methods instead of exact optimization al-
gorithms, let us work out the consequences of not having or having tight and save
bounds by considering the following example.
Assume that an industrial company wants to compare the economic impact mea-
sured as the net present value (NPV) of two different investment decisions. Each
decision would involve a combination of building new production facilities and dis-
continuing the use of old, less competitive assets. The investment decision to be
taken is significant as it involves the potential of spending several million US$ and
impacts many people at the manufacturing locations.
After modeling and solving the problem using an exact optimization algorithm,
the results presented for each alternative are as follows:
Description Alternatives
Scenario A B
Net present value (NPV) in units of million US$ 110 120
Integrality gap, max. deviation from optimal solution 2.1% 1.8%
Maximum possible NPV 112.31 122.16
If we want to maximize the net present value of the two alternatives, then clearly,
scenario B is better and preferable. Note that scenario A is bounded by 112.31,
while in B we have already a feasible point with objective function value 120 and a
1.8% integrality gap. The reader could ask why the gap is not zero. We have chosen
this case of a nonzero gap to illustrate and explain the value of safe bounds.
A method which does not generate bounds might produce the following results:
Description Alternatives
Scenario A B
Net present value (NPV) in units of million US$ 112 110
Integrality gap, max. deviation from optimal solution n/a n/a
Maximum possible NPV ? ?
Since such methods do not provide the maximum possible deviation from the op-
timal solution, one might conclude that alternative A is more profitable. Note that
in scenario A we found an even better feasible point than the exact method. This
demonstrates even more that computing feasible points is only half the story. The
other half is bounds or the proof of optimality.
How can we explain this result? The improvement method returned a value that
is 99.7% optimal for alternative A and a value that is 90% optimal for alternative B.
There is no control or guarantee for the solution quality of improvement methods.
In this example case we could only quantify the deviation from the optimal solution
because we computed the comparison basis with an exact optimization algorithm.
Therefore, we believe that it is in the clients’ best interest to either prove optimal-
ity or at least provide reasonable tight bounds on the value of the objective function.
In addition, the use of exact optimization methods can be coupled with sensitivity
analyses, analyses of robustness, and stochastic optimization described in [27]. If
the data we used in the example scenarios A and B are so uncertain that the ob-
jective function may vary by more than 5% or 6%, we cannot really distinguish
between scenarios A and B. This is a fair statement saving us from the trouble to
have lengthy meetings and discussions. If the input data are more accurate, the
scenarios can be distinguished giving a quantitative basis for further discussion.
But note that the dependence of the uncertainty of the objective function on the
uncertainty of the data can quantitatively be established by only mathematical tech-
niques and a model. Thus we conclude that improvement methods can lead to
believing to have the best possible solution when in fact there is no way to prove
this claim. One might be attempted to conclude that safe bounds or exact optimiza-
tion is relevant only for strategic decision problems. However, it is not always wise
to distinguish between strategic and operative (tactical) planning. Each operative
planning problem can easily be modified to be of more strategic spirit by adding
appropriate design degrees of freedom. But even in pure operative planning prob-
lems, it might pay off to implement exact methods depending on the quality of the
solution generated by exact optimization or improvement methods.
There is another important difference between optimization algorithms and im-
provement methods: exact fulfillment of hard constraints and the proof of infeasi-
bility of a problem. Optimization algorithms guarantee constraints to be satisfied
within specified tolerances. In improvement methods they are sometimes imple-
mented by penalty methods (this is actually a method to treat soft constraints). An
alternative approach is to reject infeasible points completely. If an improvement
method fails to determine a feasible point we cannot safely conclude that the prob-
lem is infeasible.
8.2.2.2 Design Planning under Uncertainty

In strategic design planning the time horizon could easily involve several years. The
problem treated in [24] covers 15 to 20 years. Most strategic design problems exploit
demand input data as the driving force. These data are based on demand forecast.
Costs and sales prices are also subject to forecast. Therefore, at least the amount
of demand, costs, and prices are subject to uncertainty leading to an optimization
problem under uncertainty.
Kallrath [27] reviews a variety of techniques used to approach such problems.
Among them are scenario-based stochastic optimization (cf. [12], [21], [22], [43], [8],
[9], [10], [30], the Handbook of Stochastic Programming [42], [44], [2], [45], or [36]), and
chance-constrained programming (cf. [11], [40], [18], [17], [38], [3], [4]). Optimization
under uncertainty is not lacking a sufficient number of solution approaches. Most
of them involve an equivalent deterministic formulation (EDF) of the probabilistic
optimization problem (POP). However, there are almost no commercial solvers,
especially when it comes to solving the EDF problem when the original POP is
a MILP or even a MINLP problems. The field also suffers from the fact that it is
very difficult to check the underlying assumption regarding the nature of the un-
certainty and its stochasticity. The probability functions are more guesses than any
quantitative help. The interpretation of the solution is another difficult problem.
In strategic or design planning one usually would like to hedge against certain
risks while taking a design decision. An example is the reactor size in a design de-
cision. With the new reactor capacity added the predicted demand should be met
but not exceeded by a large factor. Thus, the design should be robust with respect to
deviations from the predicted demand. Therefore, when dealing with uncertainty,
robust optimization approaches are recommended if data are available to quantify
the uncertainty and an appropriate model and solution approach is available. If the
data and/or the computational resources are not available, however, then one is en-
couraged to take a scenario-based approach in order to determine how sensitive the
system is to changing conditions (forecasted demand quantities, costs, and prices
– this is similar like in weather forecast to determine the sensitivity with respect
to the initial condition of the underlying differential equation model). The term
robust optimization was originally coined by Ben-Tal et al. [6] analyzing numeri-
cal consequences of perturbed input data of LP problems. Bertsimas [7] presented
some equivalent robust formulations for various structured MILP problems.
A completely independent development is that by Floudas and co-workers; cf.
[33] and [34]. They developed a complete theoretical framework for general MILP
problems and the scheduling and planning problems are a subclass of the prob-
lems that the theory can address. They also provide concrete evidence on large-scale
applications of robust optimization approaches for scheduling.
This robust optimization formulation is based on a min–max framework and
when applied to MILP problems, produces “robust” solutions that are immune
against data uncertainty. Uncertainty is considered in the coefficients of the ob-
jective function, as well as the coefficients and right-hand-side parameters of the
inequality constraints in MILP problems. Robust optimization techniques are de-
veloped for uncertain data described by several known distributions including a
uniform distribution (interval uncertainty), a normal distribution, the difference
of two normal distributions, a general discrete distribution, a binomial distribu-
tion, and a Poisson distribution. The robust optimization formulation introduces
a small number of auxiliary variables and additional constraints into the original
MILP problem, generating a deterministic robust counterpart problem which pro-
vides the optimal/feasible solution given the (relative) magnitude of the uncertain
data, a feasibility tolerance, and a reliability level.
While in stochastic programming the number of variables increases drastically,
in this robust optimization approach the number of variables and constraints ap-
proximately only doubles (for interval uncertainties). An argument sometimes used
against robust optimization is that is too restrictive and pessimistic missing better
opportunity becomes invalid because the approach presented by Lin et al. [33] al-
lows us to specify the probability at which probabilistic constraints have to be ful-
filled. Another advantage of the Floudas robust optimization approach is that for
uniform distributions the type and the complexity of the problem do not change –
MILP problems remain MILP problems.
After all, one should keep in mind that the systematic trends embedded in the
forecast of demand, costs, and prices over more than 5 or 10 years are usually
overruled by political factors or natural disasters. Astrology looks probably like
precise science when its positive forecasts ratio is compared to the tons of com-
mercial predictions made in 1925, 1946, 1983, and 1995. All economic assump-
tions made in these years were drastically violated in 1929, the after war years and
Wirtschaftswunder in Germany, the fall of the Soviet empire in 1990, and the dis-
aster of September 11, 2001. We do not know when the next unexpected events
destroy the current expected economic development, but it is certain there will be
strong perturbations; imagine the meteorite in 1908 would not have hit the Russian
Tunguska but a major city on Earth. This is not to say one should not try to pro-
duce some forecast or try to justify quantitatively the design at hand. However, one
should not overdue. This is consistent with our preference for robust optimization
when it comes to design planning.
8.2.3
Combined Approach: Design and Operative Planning
A detailed mathematical model of a combined strategic and operative planning in

the process industry has been presented by Kallrath [24]. This model accounts for
chemical reactors of known capacity to be opened or shut down, and it accounts
even for sets of units belonging to a site subject to acquisition. Therefore, the mul-
tisite production network is subject to design decisions but not the coupled layer
of the multisite customer and sales networks. It also includes the design of swap
contracts, a typical instrument in the process industry. Uncertain future demands
are modeled by simple scenarios; this can be improved by, for instance, stochas-
tic programming or robust optimization. More features which one might want in a
planning tool for the process industry are listed in [25], Section 2. Model extensions
for customer portfolio analysis are formulated in [27]. This individual realization of
a project provides the mathematical starting point and backbone of our discussion.
This discussion is supported by a more recent model by Guillen et al. [16] which
covers already budget planning on the one hand, and planning and scheduling on
the other hand. Furthermore, Tsiakis et al. [50] developed a MILP model to deter-
mine the optimal configuration of a production and distribution network subject to
operational and financial constraints. Operational constraints include quality, pro-
duction, and supply restrictions, and are related to the allocation of the production
and the work-load balance between production sites. Financial constraints include
production costs, transportation costs, and duties for the material flowing within
the network subject to exchange rates. As a business decision the outsourcing of
production is considered whenever the organization cannot satisfy the demand.
Their model addresses both infrastructure and operational decisions. The three
publications [25], [16], and [50], as well as additional references in [50], indicate that
combined approaches are really knocking at our doors.
In this section, we address various issues such as the motivation, the choice of
objective function, time horizon and time resolution, and the need for an integrated
tailor-made approach.
8.2.3.1 Motivation for the Combined Approach

The key argument for the combined approach is the fact that the strategic or de-
sign part and the operative core are strongly connected by the capacity or other
coupling quantities. In the combined approach one determines the optimal capac-
ity or other coupling quantities corresponding to the design degrees of freedom.
The corresponding quantity to the investment decision to add larger or more re-
actors to the production network is the production capacity. Tank design problems
have the storage capacity as the corresponding quantity.
Design planning without the operative planning kernel bears the risk to work
with the wrong effective capacity. Intelligent operative planning can increase the
effective production capacity of the production network and can decrease the in-
vestment costs. Pure operative planning can help to prepare design decisions or
lead to design planning. A key question could be: what is the maximum demand of
a certain demand spectrum which can be satisfied with the given production net-
work or, more general, what is the maximum production capacity under specified
conditions? To answer this question we require that the operative planning model
allow us to maximize the total production, i.e., it is flexible enough to define a new
objective function and to operate without explicit demands (objective function 7
in [24]) or to operate with certain demand spectra (objective function 8 in [24]). A
similar approach leads to customer or product portfolio optimization.
In these cases the operative planning model is used to explore the limits of the
capacity. In a scheduling problem one could apply a similar approach to explore
the punctuality of fulfilling orders. If one is content with the current situation and
capacity, nothing else needs to be done. If the perspective of the business exceeds
the current capacity limits, one is naturally led to a design planning problem where,
for instance, the production capacity after solving the design planning problem
needs to match the projected business. To summarize this: the operative planning
model alone can be used to explore the limits. If the projected business expectations
exceed the limits we need to add additional capacity (units, topological connections,
etc.) to a given supply chain design. Only the combinations of design planning and
operative planning can ensure that this is done in an optimal way.
8.2.3.2 The Choice of the Objective Function

In operative planning, i.e., production and sales planning, contribution margin is
one of the most reasonable objective functions to select. Of course, multicriteria
objectives combining contribution margin with minimum storage volume, for in-
stance, could also make sense.
In strategic planning the choice of the objective function is more subtle. Kallrath
[24] used a discounted net present value objective function which appeared reason-
able for the investment decision involved. The longer the planning range the more
important to include discount factors and interest rates to represent the cost and
price terms properly. The investment decision of a new tank system or set of filling
stations could, however, target for providing just about sufficient tank capacity in
such a way that the tanks or filling stations will never become a bottleneck in the
scheduling problem. Analyzing this situation requires us to model the short-term
scheduling problem sufficiently accurate and also to have the input to this schedul-
ing problem computed accurately enough in the production planning phase.
Portfolio optimization problems (customer- or product-wise) may in addition to
the contribution margin consideration involve complexity costs.
The maximization of the profit after tax of a globally operating company with the
transfer prices and transportation costs among the decision variables is considered
by Vidal and Goetschalckx [51].
8.2.3.3 Time Horizon and Time Resolution

Strategic planning may lead to a longer time horizon involving several years over
which the net present value of a strategic decision is investigated. The time reso-
lution might be coarse but this depends more on the character of the decision. If
the strategic decision is connected to the implementation of a new tank system,
the process of charging or extracting to tanks needs to be modeled. This requires a
fine time resolution. For the scheduling module a continuous-time formulation is
recommended. In the SOP model time-discrete and time-continuous formulations
might coexist.
8.2.3.4 Design Planning, Operative Planning, and Scheduling

Most aspects addressed also hold if we include scheduling into the picture. The in-
vestment decision of installing a new system of tanks would be a good example. In
short-term scheduling we need to consider time aspects as charging time, extract-
ing times, or extracting rates. Thus, this investment decision is a design decision
around the underlying short-term scheduling problem; for a review on short- and
medium-term scheduling see, for instance, [14].
The operative model ideally would combine a discrete-time formulation for the
planning part and a continuous-time formulation for the scheduling module. Such
integrated models, although not yet available, are knocking at our doors (private
communication Peter Verderame, Munawar Shaick, and Floudas – On the Integra-
tion of Planning and Scheduling). They establish the compatibility between plan-
ning and scheduling. The underlying MILP or MINLP model is a strategic-operative-
scheduling model. Nowadays, this is a real challenge. Nevertheless, Guillen et al. [16]
cover already budget planning on the one hand, and planning and scheduling on
the other hand.
8.2.4
The Need for an Integrated Tailor-Made Approach
In Section 8.2.3 we referred to three case studies in the literature, in which design
and operative planning had been combined. In this section we strongly argue that
CSOP needs a unique MILP or MINLP model for each particular industrial case (a
product division of a larger chemical company), i.e., we express doubts about the
usefulness of commercial software or standard approaches. The recent case studies
by Guillen et al. [16] and Tsiakis et al. [50] support this claim. This is not to say that
such a model should include the whole supply chain of the global players such
as BASF, Bayer, Dow, Proctor & Gamble, etc., but it should describe a reasonable
business unit or division of such companies, or the whole production of a mid-size
enterprise.
Could APS be an alternative? It seems difficult and is not recommended for the
following reasons.
1. APS providers target for having a generic software for all
industries. They expect the software to scale well for all sizes
of problems. This is understandable from the commercial
point of view but also indicates that this is a low-level

compromise.
2. The data model is often not really appropriate for a
mathematical optimization model. The reason is that often
when the data model had been developed the decision or
optimization problem was not yet part of the picture at all.
3. SAP APO had an integrated strategic network planning
module but discontinued this from release 4.0. If this
module had been a great success, it would probably not have
been abandoned.
4. It is not easy to foresee which design analysis one would like
to perform in the next few years. That makes it difficult to
develop standard software for strategic design planning
problems which rather require a case-to-case- and
project-to-project-based approach.
Particularly, the third issue makes the advantage of a tailor-made CSOP approach
obvious. It implements a culture in a company or division which considers MIP-
based decision support as a natural way to approach the company’s design deci-
sions.
8.2.4.1 Solution Approach

Most of the planning problems in the process industry lead to MILP models, some
to MINLP models. Among the relevant building blocks are tracing the states of
plants, modeling production, balance equations for material flows, transportation terms,
consumption of utilities, cost terms, variables and constraints involving the design deci-
sions, and special model features. Mode changes, start-up and cancellation features,
and nonlinear cost structures require many binary variables. Minimum utilization
rates and transportation often require semicontinuous variables. Special features
such as batch and campaign constraints across periods require special constraints
to implement the concept of contiguity. The model, however, remains linear in all
variables. Only if the pooling problem occurs, e.g., in the refinery industry or the
food industry, we are really facing a MINLP problem. For a review on algorithms
used in LP, MILP, NLP, and MINLP the reader is referred to [37], [13], or [23].
It is very convenient and saves a lot of maintenance work if the planning model
is implemented in an algebraic modeling language. In modeling languages one
stores the knowledge about a model. A model coded in an algebraic modeling lan-
guage defines the problem; it usually does not specify how to solve it. Modeling lan-
guages are declarative languages and they contain the problem in a declarative form
by specifying the properties of the problem. Algebraic modeling languages are a spe-
cial class of declarative languages, and most of them are designed for specifying
optimization problems of the form
min f (x), (8.1)

s.t. g(x) = 0, (8.2)
h(x) 0, (8.3)
where x denotes a subset of X = Rm
× Zn .
Usually, in operative planning all
functions involved are linear.
The problem is flattened, i.e., all variables and constraints become essentially
one-dimensional, and the model is written in an index-based formulation, using
algebraic expressions in a way which is close to the mathematical notation. Typically,
the problem is declared using sets, indices, parameters, and variables.
In an algebraic modeling language, the model can be written in a form which is
close to the mathematical notation. The formulation of the model is independent of
solver formats. Different solvers can be connected to the modeling language, and the
translation of models and data to the solver format is done automatically. This has
several advantages. The formerly tedious and error prone translation steps are done
by the computer, and after thorough testing of the interface errors are very unlikely.
There is a clean cut between the modeling and the numerical, algorithmic part. In
addition, for hard problems different solvers can be tried, making it more likely that
a solution algorithm is found which produces a useful result.
In a modeling language, model and model data are kept separately. There is a
clear cut between the model structure and the data. Thus, many different instances
of the same model class with varying data can be solved. Many systems provide
an ODBC (open database connectivity) interface for automatic database access and
an interface to the most widely used spreadsheet systems. This relieves the user
from the laborious duty of searching for the relevant data every time the model
is used. A second advantage of this concept is that during the development phase
of the model (in the cycle) the approach can be tested on toy problems with small
artificial data sets, and later the model can be applied without change for large-scale
industry-relevant instances with real data.
Modern algebraic modeling languages such as GAMS [52] or Mosel [53] use
state-of-the-art commercial solvers, e.g., XPressMP [54] or CPLEX [55], and allow
us to solve even huge MILP problems with several hundred thousand variables
and constraints quite efficiently. In the case of MINLP, the solution efficiency de-
pends strongly on the individual problem and the model formulation. However,
we recommend that for both problem types, MILP and MINLP, the full mathemat-
ical structure of a problem is exploited, that appropriate reformulations of models
are made, and that problem-specific valid inequalities or cuts are used. Software
packages may also differ with respect to the ability of presolving techniques, default
strategies for the Branch&Bound algorithm, cut generation within the Branch&Cut
algorithm, and last but not the least diagnosing and tracing infeasibilities, which is
an important issue in practice. Modeling languages also provide an elegant means
to incorporate decomposition approaches (cf. [13], Chapter 6, as well as time de-
compositions (cf. [19], [20]), column generation and Branch&Price (cf. [5]), or hy-
brid techniques [39] in which constructive heuristics and local search improvement
methods are coupled with exact MIP algorithms to produce feasible points and
tight lower and upper bounds.
8.2.4.2 Software Design Issues

The core operative model needs to satisfy the customer’s expectation regarding
time horizon, time resolution and degree of reality, and computing time. This
model is ideally maintained by an inhouse crew, or in other cases by long-term
contracted external consultants.
If this model is coded in an algebraic modeling language it is relatively easy to
change fixed input data into variables. One should not expect that a design problem
occurs only once in the lifetime of an operative model. This year one might want to
investigate capacity adjustments, in three years an acquisition, and at some other
time tax laws are changed drastically. Therefore, the design modules of the model
might not be activated all at once but may increase in size over the lifetime of the
operative planning problem.
One may anticipate that multisite design models, and even more if they are com-
bined with operative planning models, become large easily. Goetschalckx [15] gives
some impressive example. Production rates and transfer prices, once they become
variables, turn a MILP problem into a MINLP problem. Therefore, it is important
to have a modeling language at hand which offers MILP as well as NLP or MINLP
solvers. It is also important to maintain the SOP model as modular as possible.
Various design modules are arranged around the core operative model as satellites;
one could also consider them as components of a tool set. It is rare that all de-
sign module are activated all at once. More likely these modules are options which
are used in certain cases. The great advantage of this structure is to have full data
consistency and compatibility. Mathematical optimization models – unlike human
users – are not very tolerant against data inconsistencies. Therefore, using such
models ensures that a company’s or division’s database is clean and consistent.
8.2.4.3 Limits of the Approach

The CSOP approach can be limited by several reasons such as
1. data availability,
2. model complexity if more and more satellite modules are
added,
3. the numerical complexity to solve the resulting model,
4. character and structure of the model (MILP, NLP, MINLP,
convex versus nonconvex), and
5. maintaining the model.
The lack of data is probably the largest problem in model-based design planning
combined with operative planning. If design aspects are included in the model,
one needs all data (consistent with the operative approach) for the design objects
as well. These refer to capacities, production rates, transport connections, etc. If
we are dealing with portfolio analysis, additional data might be necessary charac-
terizing the customers or products. If a company considers to purchase existing
units, these data exist; obtaining them might be difficult, especially, if the units
are planned to be purchased in the course of a takeover. If the new units to be
integrated do not yet exist but are rather free design units, some properties (e.g.,
production rates) could also be degrees of freedom. That way not only the overall
capacity (a static property of a plant) but also a functional property (the production
rates) is results of the design study. The fact that it might be difficult or expensive
to obtain or provide the data needed should be balanced against the potential gains
obtained by the approach. If the design involves the investment of several Million
Cypriot pounds it seems well justified to spend a few hundred thousand Cypriot
pounds to get a better database to support the decision.
The numerical complexity can be increased due to the number of variables
caused by a fine grid in a discrete-time formulation. But even in the case of a very
large MILP model there are techniques such as column generation or hybrid tech-
niques producing good quality solutions and bounds [28].
Another reason is that some design decisions may turn the operative model as-
sumed to be mixed-integer linear so far into a mixed-integer nonlinear problem.
This is the case if the specific selling prices, specific costs per ton, or production
rates become decision variables. These problems are not easy to solve. However,
the human decision taker will face the same problem and complexity. His situa-
tion is even worse, because human-based approaches without the support of math-
ematical optimization suffer strongly from obtaining feasible solutions. Usually
only a few points in the parameter space are investigated in typical simulation ap-
proaches. Eventually, the decision should be based on balancing the cost to support
the decision against the implications and scale of the design decision.
In Section 8.2.4.2 a possible scenario for maintaining the core model and extend-
ing design modules has been described.
8.3
Combined Strategic and Operative Planning – Commercial Software versus Individual
Models
In Section 8.2 we discussed the combination of design and operative planning from
the needs of obtaining high quality decisions for a design and operative planning
problem. The focus was on the technical side.
In this section we elaborate more on the discussion of commercial software ver-
sus individual models from the viewpoint that this affects companies’ policies on
where and how to allocate supply chain planning and design in a company’s orga-
nizational matrix. These are typical problems of supply chain planning and design
in large companies related to the needs of appropriate modeling (the innovative
side) versus maintainable IT structures (the standardization side).
8.3.1
Commercial Software for Operative Planning and Design Planning
Commercial software for strategic planning and operative planning is or was avail-
able through the large APS providers such as SAP or i2 Technologies (cf. [48]);
other providers are Aspen Technology, Inc., Manugistics, J. D. Edwards, Visopt,
8.3 Combined Strategic and Operative Planning – Commercial Software versus Individual Models 235
Logictools, Baan, Peoplesoft Inc., Planning Systems Inc., Mapics, CAPS Logistics,
Epicor, Tata Technologies, and Kiran Consulting Group. They offer separated mod-
ules for those tasks; actually SAP stopped offering strategic planning in SAP APO
4.0. Especially, the large APS providers take advantage of the deep integration with
the company’s transactional data systems. Integrity and software with the same
look-and-feel may appear attractive and even allows us to reallocate personal easily
from one department to another one. Below we provide a list of providers [56] offer-
ing solutions to design planning, production planning, and scheduling. If strategic
or design planning is offered, it is not in a simultaneous approach with operative
planning.
Name Founded Web address Type Product P SD
Adv. Proc. Comb 1993 combination.com PS VirtECS +

AI Systems aisystems.be PS SteelPlanner +
AspenTech 1981 aspentech.com PF, APS MiMi +
ASPROVA 1994 asprova.com PS ?
AXXOM 2001 axxom.com P ORion-Pi + +
Delta Software 1980 deltasw.com CP DeltaPlan +
Haverly Systems 1962 haverly.com PS GRTMPS +
Honeywell Process S. see [56] R, PS +
ILOG 1987 ilog.com PS PowerOPS +
Ingenious ingenious.cc PS ProPlan, ProSched +
Insight 1978 insight-mss.com ND SAILS + +
i2 Technologies 1988 i2.com APS +
Logic Tools 1995 logic-tools.com APS, P SupplyPlanner +
OM Partners ompartners.com PS OMPDesigner,Planner + +
Optience 2003 optience.com PS SMartSuite +
OR Soft 1990 orsoft.net PS Schedule++ + +
PSE 1997 psenterprise.com PS OSS Scheduler +
Quintiq 1997 quintiq.nl PS +
River Logic 2000 riverlogic.com P Enterprise Opt ? +
SAP 1972 sap.com APS SAP APO + (+)
Shell Global Solutions 2001 shell.com GMOS + +
Supply Chain C. 1993 supplychain.com P Zemeter +
Technologix Inc. 1993 technologix.ca P OptiNet +
VisOpt Inc. visopt.com APS +
Wam Systems 1987 wamsystems.com P Picaso +
In this table type resolves whether the companies offer consultancy (C), and
whether the software is for planning (P), scheduling (S), or refinery scheduling
(R); P is a flag checking the suitability of the product for the process industry while
the flag SD indicates whether the providers offer solutions for strategic or design
planning.
Finally, a very elaborated approach was followed by Shell Global Solutions when
they developed their software GMOS/NetSim [32], which stands for Global Man-
ufacturing and Logistics Optimization System/Network analysis and Supply chain op-
timization System. GMOS covers the end-to-end supply chain, from raw material to
intermediates, to finished product grades to the end customer. It solves both NLP
and MIP problems, with graphical user-interaction and constraint-analyzing facil-
ities. The GMOS/NetSim modular approach allows optimization between multi-
ple manufacturing plants, packaging locations, and distribution modes for both
inhouse and third-party operations, taking joint venture/ownership aspects into
account. The model can offer optimization of grades production at a compo-
nent/specification level. Movements and storage of products can be tuned to re-
duce tax, import tariffs, and duties. A structured navigator manages the data re-
siding within workbooks and database files, and therefore the user has full control
over the data. It includes operative planning features also not yet fully down to
include sequencing problems and setup changes. Shell uses GMOS to undertake
projects for customers but in exceptional cases it can also be purchased. It needs to
be checked for each case whether the operative planning aspects are covered to the
level of detail required.
8.3.2
The Role of Individual Software for the CSOP Approach and APS
CSOP can exist on its own, or simultaneously with tight connections to an APS. In
the latter case the APS could provide the database. However, for the CSOP we need
all data also for design units. These are usually not available in the APS. There-
fore, it would be better to have a GUI and a local database for the CSOP approach
which is interfaced to an APS. Extracting only data from an APS should not lead
to problems. Some APS also have a module for strategic planning implemented.
However, the definition of strategic planning differs from ours, which includes all
features of the APS but offers many more. Interfacing the CSOP to an APS can be
approached as described in [29].
It is not a surprise that many discussion points arise when talking about the
CSOP and the role of APS or other commercial software, such as
• Why use an individual tailor-made model and not an APS?
• What are the benefits of the CSOP approach?
• At which point one should prefer a CSOP versus an APS?
• What are the dependences on APS and modeling specialists?
We discuss these points in the next subsections.
8.3.3
When to Select Individual Tailor-Made Methods?
Let us first discuss this only with respect to the operative model. An obvious an-
swer is, of course, when the APS is not sufficient to support the operative decision
problem one has in mind. There are other criteria such as follows.
1. What is the value of having a transparent inhouse model
versus a black box model, for instance, a standard APS? May
8.3 Combined Strategic and Operative Planning – Commercial Software versus Individual Models 237
be it does not matter for the end user, but for the company
this makes a difference.
2. How well suitable is the APS standard model? Does it fit the
reality or is it necessary to change the reality? What is the
purpose of the model?
3. What should be the lifetime of the model and do we expect
modified request toward the model over the years. The
upfront investment might be significant but can easily pay
out over the years if more requests have to be served.
A more subtle argument mainly for larger companies with many production divi-
sions might be to evaluate the standardization aspects higher than the individual
appropriateness. History answers the questions clearly for us: standardization is a
type of consolidation producing stability for a period of time but also blocking new
innovations. In the end each standard is replaced by a better one. Eventually, inno-
vation always wins – the question is only whether we will become old enough to ex-
perience this. What we observe now in supply chain planning is that APS systems
have difficulties with the existing reality when it comes to master and production
planning in the process industry and metals industry. For strategic planning and
scheduling their acceptance seems to be shrinking anyway; again the problem is
that they do not meet the level of detail required.
Now we discuss the same alternatives, individual approach versus standard soft-
ware, with a focus on the strategic design decisions. Here the aspects to be consid-
ered are as follows.
1. Can the APS or any other commercial software be extended
to cover the design decision problem addressed?
2. What is the time scale to realize such projects?
3. What are the costs involved in this?
There are not only the APS but also other software tools (SAILS, AXXOM,
OMPDesigner) such as GMOS developed by Shell Global solutions, which tar-
gets on strategic planning including some aspects of operative planning. GMOS is
very suitable for the process industry. This package checks whether the operative
planning part is detailed enough. If not one would have two separate pieces of op-
timization packages: one for strategic and one for operative planning. This is not
what we are promoting in the chapter.
There is another group of arguments why individual solutions for strategic plan-
ning and design planning are more appropriate than commercial solutions.
Strategic planning is where the business history (customers, products, design)
is coded. Design planning also reflects the company’s development over time. It
would be very surprising if any commercial software would be suitable to such a
situation without any modification. Thus the key issue is whether the commercial
software is open enough to understand precisely the embedded model and to mod-
ify it to the own needs.
With an individual solution the benefits of the CSOP approach are obvious:
• combine and integrate strategic and operative planning

precisely as wanted,
• get optimal (or bound-proven) decisions of design and
strategic decisions in the context of an underlying supply
chain planning problem,
• be fully open to new strategic or design decision problems,
• keep flexibility and react promptly on new requests, and
• control the implementation speed (it may take much longer
to add new aspects to a standard package controlled by a
commercial vendor).
SAP’s decision to abandon the strategic network module from SAP APO 4.0 on
indicates that commercial supply chain software focuses more on the operative
aspects. Another reason could be that multisite global networks are not necessar-
ily available in only one SAP system but could be in several. Detailed scheduling
coupled with design aspects is not available at all. This is not a surprise: even in de-
tailed scheduling alone a lot of very specific know-how of a manufacturer is hidden
and it is difficult for an APS or commercial software to model this. In addition, the
problems SAP APO has with meeting their clients requests related to individual
storage entities have been known for many years. Therefore, asking for a coupled
analysis of a detailed scheduling and investment decision on a new tank system is
probably too much.
8.3.4
Developing and Maintenance Costs of the Inhouse Model
The development effort and costs scale with the complexity and richness of the
operative core planning, or planning-scheduling model. One should not expect to
get this cheaper than 150 kEuro and experience from many projects shows that 500
kEuro cover already most of the features. If the model is properly built, strategic
design aspects can be added module per module at project-based costs.
For maintaining this inhouse model built for one product division one or two
skilled people could suffice. If we allow for one or two additional persons, the group
can do the operative planning, maintain the software, and produce new design
modules. If the whole model has a lifetime of at least 10 years, the development
cost of, for instance, 500 kEuro requires us to spend 50 kEuro/year as investment
cost into this inhouse model now owned by a product division. These investment
costs plus, for instance, three permanent people for doing the operative planning,
maintaining the software, and producing new design modules should be compared
to the purchasing or development cost and maintenance fees of usually 15% to
be paid to a commercial provider and the personnel cost of the people using the
commercial software to do the planning.
8.4 Summary 239
8.3.5
Dependences on Software Providers or Inhouse Modeling Specialists?
With any piece of software obtained to be used over a longer time horizon depen-
dences are established. The commercial software providers often also operate as
consulting firms. Large players such as SAP or i2 Technologies can easily develop
monopolistic behavior and dominate the price to be paid. The dependence on in-
house specialists can at least be better controlled. However, the number of inhouse
specialists should not be too small (risk to lose the know-how due to personal fluc-
tuation) nor should the group of such people be too large and have to much power
(politics in large organizations). A well functioning inhouse optimization group is
also very helpful to obtain external software or appropriate consulting at a reason-
able price.
8.4
Summary
We have presented a review of multisite design planning studies in the literature,

commented on operative planning, and proposed a combined strategic and opera-
tive planning (CSOP) approach based on a tailor-made mixed-integer-model involv-
ing strategic, design, planning, and possibly scheduling components (SOP model).
In a maximizing net profit scenario the client reports cost saving of several millions
US$ using a combined model [24]. CSOP attempts to integrate strategic and oper-
ative planning. By combining both models into a single framework, one ensures
the efficient utilization of plant resources, and design decisions are only imple-
mented when an increase or decrease in production levels is warranted. The key
motivation for the integration of strategic and operative planning is the fact that the
two planning problems are strongly connected to another one. For example, design
decisions have a profound effect on the plant’s ability to meet production levels.
The SOP model could coexist with software such as SAP APO, or it could replace
its planning modules completely (Section 8.3). Hence, it is always flexible to incor-
porate new features or to support strategic studies (product portfolio or customer
portfolio studies) or design questions (more units, units of different size). In the
coexistence scenario, an appropriate date interface to SAP APO could always en-
sure the data integrity. Using the SOP model would not disturb the daily operative
needs of master planning. It would also allow planning under uncertainty and a
complete scenario analysis. If the SOP model is used in an organization without
an APS, i.e., on a standalone basis, of course, one could also have two implemen-
tations of the SOP model: one for the daily operative needs, and another one for
separate studies.
Among the design features we addressed are supply chain design and analysis,
the distribution and capacity design of production or storage facilities, as well as
financial design optimization, e.g., the maximization of after tax profit with trans-
fer prices and transport costs as decision variables. The combined strategic and
operative models are expected to be large in size and difficult to solve. They may re-
quire tailor-made solution strategies. This makes it not very likely that commercial
software packages are available which can cover this functionality. When compar-
ing generic software packages to tailor-made modeling software packages, the use
of tailor-made packages is recommended due to their inherent flexibility and abil-
ity to adapt to changing system requirements. Tailor-made solution strategies may
include decomposition techniques (rolling time decomposition, column genera-
tion, and Branch & Price) or hybrid techniques in which constructive heuristics
and local search improvement methods are coupled with exact MIP algorithms to
produce feasible points and tight lower and upper bounds. The ideal case would
be a commercial software which is generic enough to serve as a real standard but
is also sufficiently open to map a client reality in the desired details and to allow
tailor-made solution techniques to be added if necessary. Modeling languages such
as GAMS or Mosel, solvers such as CPLEX or XpressMP, and experienced consul-
tants are available giving hope that the next few years bring us much closer to this
ideal case.
Acknowledgments
The comments and corrections by Bert Beisiegel (b2st GmbH, Mülheim an der
Ruhr, Germany), Gerard Debeuckelaer (UTI, Bucarest, Romania), Ulrich Eber-
hard (BASF Aktiengesellschaft, Ludwigshafen, Germany), Thomas I. Maindl (SAP,
Waldorf, Germany), and Steffen Rebennack (University of Florida) are gratefully
acknowledged. Especially, Beate Brockmüller’s (BASF Aktiengesellschaft, Lud-
wigshafen, Germany) careful comments and iterated feedback as well as the con-
structive suggestions by Chris A. Floudas (Princeton University) helped to improve
this contribution.
References
1 Ahmed, S., Sahinidis, N. V., Chem- 3 Arellano-Garcia, H., Martini,

ical process planning, in: Encyclope- W., Wendt, M., Li, P., Wozny, G.,
dia of Optimization, Vol. 1, Floudas, Chance-constrained batch distilla-
C. A., Pardalos, P. M. (eds.), Kluwer, tion process optimization under un-
Dordrecht, The Netherlands, 2001, pp. certainty, in: Proc. 4th Intl. Conf. on
252–254. Foundations of Computer-Aided Process
2 Andrade, R., Lisser, A., Maculan, Operations (FOCAPO), Grossmann,
N., Plateau, G., BB strategies for I. E., McDonald, C. M. (eds.), OMNI
stochastic integer programming, Press, Wisconsin, 2003, pp. 609–612.
in: Special Volume of Annals of OR: 4 Arellano-Garcia, H., Martini, W.,
State-of-the-Art IP and MIP (Algo- Wendt, M., Wozny, G., Robust opti-
rithms, Heuristics and Applications), mization process design optimization
Spielberg, K., Guignard, M. (eds.), under uncertainty, in: Proc. 6th Intl.
Kluwer, Dordrecht, The Netherlands, Conf. on Foundations of Computer-
2005. Aided Process Design (FOCAPD),
References 241
Floudas, C. A., Agrawal, R. (eds.), optimization of process operations

CACHE Corp., Austin, TX, 2004, pp. and financial decisions to enhance
505–508. the integrated planning/scheduling
5 Barnhart, C., Johnson, E. L., of chemical supply chains, Comput.
Nemhauser, G. L., Savelsberg, Chem. Eng. 30 (2006), pp. 421–436.
M. W. P., Vance, P. H., Branch-and- 17 Gupta, A., Maranas, C. D., Manag-
price: column generation for solving ing demand uncertainty in supply
huge integer programs, Oper. Res. chain planning, Comput. Chem. Eng.
46(3) (1998), pp. 316–329. 27 (2003), pp. 1219–1227
6 Ben-Tal, A., Nemirovski, A., Ro- 18 Gupta, A., Maranas, C. D., McDon-
bust solutions of linear programming ald, C. M., Mid-term supply chain
problems contaminated with uncer- planning under demand uncertainty:
tain data, Math. Progr. 88 (2000), pp. customer demand satisfaction and in-
411–424. ventory management, Comput. Chem.
7 Bertsimas, D., Sim, M., Robust dis- Eng. 24(12) (2000), pp. 2613–2621.
crete optimization and network flows, 19 Janak, S. L., Floudas, C. A., Kall-
Math. Progr. B 98 (2003), pp. 49–71. rath, J., Vormbrock, N., Production
8 Birge, J. R., Stochastic programming scheduling of a large-scale indus-
computation and applications, IN- trial batch plant: I. Short-term and
FORMS J. Comput. 9 (1997), pp. 111– medium-term scheduling, Ind. Eng.
133. Chem. Res. 45 (2006), pp. 8234–8252.
9 Birge, J. R., Louveaux, F. V., Introduc- 20 Janak, S. L., Floudas, C. A., Kall-
tion to Stochastic Programming, Vol. 10, rath, J., Vormbrock, N., Production
Springer, New York, 1997. scheduling of a large-scale industrial
10 Carœ, C. C., Schultz, R., Dual de- batch plant: II. Reactive scheduling,
composition in stochastic integer pro- Ind. Eng. Chem. Res. 45 (2006), pp.
gramming, Oper. Res. Lett. 24 (1999), 8253–8269.
pp. 37–45. 21 Kall, P., Stochastic Linear Program-
11 Charnes, A., Cooper, W. W., Chance- ming, Springer, Berlin, 1976.
constrained programming, Manage. 22 Kall, P., Wallace, S. W., Stochastic
Sci. 5 (1959), pp. 73–79. Programming, Wiley, Chichester, 1994.
12 Dantzig, C. B., Linear programming 23 Kallrath, J., Mixed integer optimiza-
under uncertainty, Manage. Sci. 1 tion in the chemical process industry:
(1955), pp. 197–206. experience, potential and future per-
13 Floudas, C. A., Nonlinear and mixed- spectives, Chem. Eng. Res. Des. 78(6)
integer optimization: fundamentals and (2000), pp. 809–822.
applications, Oxford University Press, 24 Kallrath, J., Combined strategic and
Oxford, UK, 1995. operational planning – an MILP suc-
14 Floudas, C. A., Short-term and cess story in chemical industry, OR
medium-term scheduling, in: Supply Spektrum 24(3) (2002), pp. 315–341.
Chain Optimization, Papageorgiou, 25 Kallrath, J., Planning and schedul-
L. G., Georgiadis, M. C. (eds.), Wiley- ing in the process industry, OR Spek-
VCH, Wiesbaden, Germany, 2007. trum 24(3) (2002), pp. 219–250.
15 Goetschalckx, M., Strategic network 26 Kallrath, J., Modeling Languages in
planning, in: Supply Chain Manage- Mathematical Optimization, Kluwer,
ment and Advanced Planning, 2nd Norwell, MA, USA, 2004.
edn., Stadtler, H., Kilger, C. (eds), 27 Kallrath, J., Solving planning and
Springer, Berlin, Deutschland, 2002, design problems in the process indus-
pp. 105–121. try using mixed integer and global op-
16 Guillen, G., Badell, M., Espuna, timization, Ann. Oper. Res. 140 (2005),
A., Puigjaner, L., Simultaneous pp. 339–373.
28 Kallrath, J., Modeling difficult opti- 37 Nemhauser, G. L., Wolsey, L. A., In-
mization problems, in: Encyclopedia of teger and Combinatorial Optimization,
Optimization, Vol. 2, Floudas, C. A., Wiley, New York, 1988.
Pardalos, P. M. (eds.), Springer, Dor- 38 Orçun, S., Altinel, I. K., Hortaçsu,
drecht, The Netherlands, 2008, in O., Scheduling of batch processes
press. with operational uncertainties, Com-
29 Kallrath, J., Maindl, T. I., Real Op- put. Chem. Eng. 20 (1996), pp. S1215–
timization with SAP-APO, Springer, S1220.
Heidelberg, Germany, 2006. 39 Pochet, Y., Wolsey, L. A., Production
30 Klein-Haneveld, W. K., van der Planning by Mixed Integer Program-
Vlerk, M. H., Stochastic integer pro- ming, Springer, New York, 2006.
gramming: general models and algo- 40 Prékopa, A., Stochastic Programming,
rithms, Ann. Oper. Res. 85 (1999), pp. Kluwer, Dordrecht, The Netherlands,
39–57. 1995.
31 Kouvelis, P., Munson, C. L., Using 41 Rohde, J., Wagner, M., Master plan-
a structural equations modeling ap- ning, in: Supply Chain Management
proach to design and monitor strate- and Advanced Planning, 2nd edn.,
gic international facility networks, Stadtler, H., Kilger, C. (eds.),
in: Handbook of Quantitative Supply Springer, Berlin, Deutschland, 2002,
Chain Analysis: Modeling in the E- pp. 144–160.
Business Era, Simchi-Levi, D., Wu, 42 Ruszczyński, A., Shapiro, A., Sto-
S. D., Shen, Z. M. (eds.), Springer, chastic Programming, volume 10 of
New York, 2004, pp. 681–709. Handbooks in Operations Research and
32 Lasschuit, W., Thijssen, N., Sup- Management Science, Elsevier, North-
porting supply chain planning and Holland, Amsterdam, 2003.
scheduling decisions in the oil and 43 Schultz, R., On structure and stabil-
chemical industry, Comput. Chem. ity in stochastic programs with ran-
Eng. 28 (2004), pp. 863–870. dom technology matrix and complete
33 Janak, S. L., Lin, X., Floudas, C. A., integer recourse, Math. Program. 70
A new robust optimization approach (1995), pp. 73–89.
for scheduling under uncertainty: I. 44 Schultz, R., Stochastic programming
Bounded uncertainty, Comput. Chem. with integer variables, Math. Program.
Eng. 28 (2007), pp. 1069–1085. B 97 (2003), pp. 285–309.
34 Lin, X., Janak, S. L., Floudas, C. A., 45 Sen, S., Algorithms for stochastic
A new robust optimization approach mixed-integer programming models,
for scheduling under uncertainty: II. in: Handbook of Discrete Optimization,
Uncertainty with known probability Aardal, I., Nemhauser, G. L., Weis-
distributions, Comput. Chem. Eng. 31 mantel, R. (eds.), Elsevier, North-
(2004), pp. 171–195. Holland, Amsterdam, 2004.
35 Matsuoka, S. S., Mukari, M., Pro- 46 Shah, N., Process industry supply
duction planning for multi-site batch chain: advances and challenges, Com-
plants with the MILP method, J. Japan put. Chem. Eng. 29 (2005), pp. 1225–
Pet. Inst. 47(5) (2004), pp. 318–325. 1235.
36 Mitra, G., Poojari, C., Sen, S., 47 Stadtler, H., Production planning
Strategic and tactical planning mod- and scheduling, in: Supply Chain
els for supply chain: an application Management and Advanced Planning,
of stochastic mixed integer program- 2nd edn., Stadtler, H., Kilger, C.
ming, in: Handbook of Discrete Op- (eds.), Springer, Berlin, Deutschland,
timization, Aardal, I., Nemhauser, 2002, pp. 178–193.
G. L., Weismantel, R. (eds.), Elsevier, 48 Stadtler, H., Kilger, C. (eds.), Sup-
North-Holland, Amsterdam, 2004. ply Chain Management and Advanced
References 243
Planning, 3rd edn., Springer, Berlin, fer pricing and transportation cost al-
2004. location, Eur. J. Oper. Res. 129 (2001),
49 Timpe, C., Kallrath, J., Optimal plan- pp. 134–158.
ning in large multi-site production 52 GAMS Corporation,
networks, Eur. J. Oper. Res. 126(2) http://www.gams.com
(2000), pp. 422–435.
53 Dash Optimization,
50 Tsiakis, P., Papageorgiou, L. G., http://www.dashoptimization.com
Georgiadis, M. C., Optimal design
54 Dash Optimization,
of supply chain networks using math-
http://www.dashoptimization.com
ematical programming, in: Supply
Chain Optimization, Papageorgiou, 55 ILOG, http://www.ilog.com
L. G., Georgiadis, M. C. (eds.), Wiley- 56 The Honeywell URL is
VCH, Wiesbaden, Germany, 2007. http://hpsweb.honeywell.com/Cultures/en-
51 Vidal, C. J., Goetschalckx, M., A US/Products/OperationsApplications/
global supply chain model with trans- PlanningScheduling/default.htm
245
9
Chemical Supply Chain Redesign
Pavan Kumar Naraharisetti, Iftekhar A. Karimi, Rajagopalan Srinivasan
9.1
Introduction
A supply chain in the manufacturing sector consists of a network of production

and inventory holding facilities, and these facilities link material, information, and
finance flows. In the network, the production facilities transform the raw materi-
als into intermediates and products. Intermediates are reprocessed into products
either at the same facility where they are manufactured or at a different produc-
tion facility. Raw materials, intermediates, and products may be transported from
one facility to another, sometimes over continents, and the products are sold to the
customers. Across the supply chain network, generally, material moves from the
raw material end to the customer end (not considering material recycle), finances
moves from the customer end to the raw material end, and information flows in
both directions. Sometimes, material and hence the finances may move in the re-
verse direction, in cases where there is a return of goods or buyback. A schematic
of the supply chain showing material movement (without considering return of
goods) is shown in Fig. 9.1.
Capacity is the maximum rate of output of a facility or the ability of the entire
supply chain to achieve certain tasks. It is an aggregate ability rather than the abil-
ity of individual resources and it deals with product families rather than individual
products. Capacity management in a supply chain involves the management of the
existing network of facilities, transportation, material supply contracts, and invest-
ments and disinvestments in the network, so as to maximize the shareholder value
and customer satisfaction. Supply-chain redesign deals with the investments and
disinvestments and relocation of facilities in the supply chain network. This leads
to a new network structure in terms of more or less number of production and
inventory holding facilities or the same network with different production and in-
ventory holding capacities. Relocation can be considered as disinvestment at one
location and investment at another location. The disinvestment of facilities can be
considered as an investment. This is because the organization incurs a cost due to
employee compensation, among others, so that they can have increased profits (re-
ISBN: 978-3-527-31693-9
246 9 Chemical Supply Chain Redesign
Fig. 9.1 Schematic representation of a chemical supply chain

of a multinational corporation. Suppliers (S); production
facilities (PF); distribution centers (DC).
duce losses) in the future [1]. Capacity management should consider the forecasts
of the raw material availability, product demand, and prices of goods while devel-
oping a plan. Since the future is uncertain, the plan that is developed should be
robust to the risk due to the uncertainty involved in the realization of the forecasts.
Capacity management in a supply chain can be classified as strategic, tactical,
and operational planning. Strategic planning deals with long-term planning, which
is performed for a period of several years to decades. The primary decisions in
strategic planning are related to investments and disinvestments in capacity, sign-
ing of contracts related to material supply, and third party logistics, import–export
regulations, tax, and so on. Tactical planning involves the utilization of the facilities
across the supply chain network that is given by the long-term capacity plan. In the
case where there is no change in the network due to the strategic plan, the tacti-
cal plan is performed with the available network structure. In the case where the
strategic plan calls for a change in the network structure early on in the planning
horizon, which is early enough that it falls under the planning horizon of the tacti-
cal plan, then these changes are to be included in the tactical plan. Tactical planning
is done for a period of a few months to one or two years. Distribution of material,
amount of inventory at the facilities, import–export regulations, tax are some of the
decisions that are considered while developing a tactical plan. Operational planning
deals with the use of the available capacity at a given production facility to meet
9.2 Chemical Supply Chains 247
the targets set by the tactical plan. The decisions involve assigning equipment to
the tasks that are to be performed, while considering such issues as the availability
of storage, resources such as electricity, cooling water among others. This is done
for timescales of few days to one or two weeks. In this chapter, we will primarily
focus on the strategic and tactical planning for the mathematical model; however,
we briefly describe capacity management at the various levels. For effective func-
tioning of the organization, the interaction between the three levels of planning is
important, that is there should be two-way communication between the strategic
and tactical planning, and tactical and operational planning.
In general, when the amounts of raw material purchased or products sold in-
crease, the price per unit material decreases and this is called economies of scale.
Several examples of economies of scale can be seen in the sale of consumer goods.
Organizations agree in advance on the price for purchase and sale of material and
this involves economies of scale. These agreements are in terms of contracts for
material supply and they are financially and legally binding. Contracts are signed
for extended periods of time in order to safeguard the organizations against un-
certainties that may affect material availability. While signing the contracts, the
decisions that the organizations consider are related to economies of scale, dura-
tion of contract, penalty for undersupply, or deviation in quality and others. In the
case where the raw material availability decreases or its price increases in the mar-
ket, contracts help in guaranteed material supply at nominal prices for the buyer.
In the case where the raw material availability is high and its prices are falling, the
interests of the supplier are protected as the buyer has to buy material at a price
that is greater than market price.
Mergers and acquisitions result in a new supply chain network and the manage-
ment has to obtain a supply chain redesign plan so as to eliminate duplication
or redundancy in facilities. Some examples of mergers and acquisitions in the
chemical industries include Chevron Texaco acquiring Unocal in 2005; PetroChina,
a subsidiary of China National Petroleum Corporation International, acquiring
PetroKazakhstan in 2005; Sanofi-Synthelabo acquiring Aventis to become Sanofi-
Aventis in 2004. Worldwide merger and acquisition volume was about $2.7 trillion
in 2005, an increase of 38% (approximately) from the previous year. This amount
was the highest since 2000 and the third largest amount in history (Thomson Fi-
nancial, Mergers & Acquisitions Review, Fourth Quarter 2005) and this gives us
an idea of the magnitude of the problem under consideration. A demerger also in-
volves a supply chain redesign because the new organizations must decide which
facilities each of them is going to own.
9.2
Chemical Supply Chains
Every organization in the manufacturing sector has a supply chain, no matter how
small. Small organizations such as computer assembly units or packaging units
may have only one facility; however, they need to optimize their manpower uti-
lization, inventory of raw materials and products, keeping in mind the variable de-
mand for products and availability of raw materials. Most of the supply chains we
see in our daily life handle goods that are itemized, that is we count them by num-
ber. They include automobiles, electronic items, clothes, and others. The process
of manufacturing these items is by assembling parts and the calculations are all by
the number of parts required to make one product. Different parts come to one ma-
chine and the product is assembled by this single machine and this type of setup
is called a job-shop. In other cases, a part goes to a machine and some work may
be done on it and the part then moves forward to other machines where additional
work is done on it. Such a type of setup is called flow-shop. The production can
sometimes be a combination of job-shop and flow-shop. Unlike many manufactur-
ing supply chains, which handle discrete materials that are assembled, chemical
supply chains involve combination of continuous, semicontinuous, and discrete
flow of raw materials and products. Continuous flows can be seen in pipelines,
inside a production facility, or between two production facilities in the neighbor-
hood, or between countries as is in the case of the supply of crude oil and natural
gas. Semicontinuous may be in the case of crude unloading from a tanker. There
is continuous flow when a tanker arrives, and when there is no tanker, there is no
flow. Discrete flow can be seen in batch production plants, where raw materials (in
a particular ratio) react in a reactor to form products.
In addition to the above fundamental difference between chemical versus other
supply chains, there are several other characteristics that distinguish chemical sup-
ply chains. For instance, chemical supply chains are long and divergent, are mostly
petroleum derived, and are hence highly sensitive to the oil prices and sociopolit-
ical uncertainties in the oil-producing countries. Complex processing equipments
are used and large capital investments are required in order to implement process
innovations, meet environmental regulations, and operate in a safe and healthy
environment. The variety of materials involved is huge, for example bulk gases
and liquids, clean petrochemical products and fuels, food and beverages, paints,
detergents and other household chemicals, and so forth. Sometimes they require
specific storage due to toxicity and quality requirements, and cannot usually have
shared inventory facility unless they are itemized in small containers. The long lead
time in delivery of raw materials (crude oil), pipeline transportation in addition to
very large crude carriers and multiparcel tankers, further add to the differences
when compared to other types of supply chains. Several of the products are sold
as raw materials to other chemical plants, and the number of chemical plants in-
volved before a final product reaches the consumer is large. The supply chain also
involves recycle, where the downstream plants sell their products as raw materials
to the upstream plants.
It is estimated that the total investments required to meet the energy needs
worldwide through 2030 will be $16 trillion and a considerable amount of these
investments will be in the oil and gas industries, which amount to $200 billion
per year (International Energy Agency’s World Energy Investment Outlook, 2003).
ExxonMobil invested a net (investment in property, plant, and equipment) of about
$100 billion in the year 2003; 55% of which was in the upstream oil and gas,
40% was in the downstream and chemicals business, and the remainder being in
“others” and discontinued operations (ExxonMobil Financial & Operating Review,
2003). In addition to the above examples, two expert opinion papers [2, 3] on the fu-
ture of process systems engineering have highlighted the importance of enterprise
and supply chain optimization in chemical supply chains.
In the chemical industry, the technical challenges arise from three dimensions.
In the physical scale, they are due to the need to consider the system at several phys-
ical aggregation levels: globally dispersed entities, production lines within a manu-
facturing site, and even individual equipment items. Second, in the temporal scale,
they are because of the different dynamics of the entities and operations such as
distribution/logistics, which may involve weeks; equipment residence times, which
involve hours; and individual processing steps, which may involve minutes. Third,
uncertainty, which is due to the fact that key factors such as market parameters
of product demands, feed stock availability and their prices; technical parameters
such as product yields, product qualities, processing times/rates; and facilities pa-
rameters such as reliability/availability, all have significant stochastic components.
9.3
Capacity Management
The world trade in merchandize for the year 2004 was $8.9 trillion in exports and
$9.3 trillion in imports (World Trade Organization: International Trade Statistics,
2005). As the human population around the world increases from 6.1 billion in
2000 to an estimated 9 billion in 2050 (World Population Prospects: The 2004 Revi-
sion, UN Population Division), this trade is only set to increase. The result is ever
changing business environment, which in turn leads to changes in the supplies of
raw materials and demands of products. Increase in demand may be due to the
opening of new markets and introduction of new and technologically better prod-
ucts. Reduced demand may be due to the introduction of trade barriers, products
becoming obsolete, and competition from other players. With the advent of global-
ization, organizations are venturing into previously unknown territories either by
locating new manufacturing and distribution facilities or by partnering with other
organizations in geographically distant locations, which further increase the com-
plexity of the supply chain and ultimately increase the difficulty in making optimal
decisions to maximize shareholder value and customer satisfaction [4, 5]. As new
markets open up, existing markets become saturated, and new sources of raw ma-
terials and demands are identified; there is a need for locating new production and
distribution facilities and shutting down old facilities. Examples of such decisions
abound in the media. For instance, Airbus in talks for locating assembly plant in
China; Ford, GM, and Merck shutting down plants in USA; Samsung investing in
new facilities in South Korea. There is also an increasing trend of relocating pro-
duction facilities to low cost centers, for example Nike moved from Indonesia to
China. Hence, organizations have identified the need for better business decision
support tools that allow for better capacity management [6–8].
Several enterprise resource planning systems are commercially available. They

enable improved coordination in the movement of materials, information, and fi-
nancial flows across the various entities in the supply chain of a particular orga-
nization and also among different organizations. Advanced planning systems are
also called decision support systems or analytical information technology tools [9],
and they are used to obtain a plan and make optimal decisions in the supply chain
in order to maximize the shareholder value. Some resource planning systems con-
sider planning; however, they do not consider sophisticated advanced planning.
Hence, advanced planning systems that make extensive use of mathematical pro-
gramming and meta-heuristics are being developed to bridge this gap [10]. It can
clearly be seen that supply chain redesign involves investments on a large scale
and the organizations should consider the various uncertainties and risks that may
arise in future and make decisions that are robust.
9.3.1
Operations Management
Various issues in operations management have been discussed before by several re-
searchers [11]. They include managing research and development, introducing new
products, planning for new capacity, managing and improving the operations in the
existing network of production and inventory facilities, increasing robustness, the
role of the management and strategy of organizations, among others. In addition,
several real case studies were presented. Readers who are more interested in the
theory related to these issues may find it worthwhile to read this book. We briefly
present these and other issues to make this chapter more complete.
An organization in the manufacturing sector has different departments such as
accounting, distribution, engineering, finance, human resources, marketing, and
operations. Operations Management integrates all the departments and helps to
manage the process of transforming raw materials into products. Capacity Man-
agement involves the management of material, production, and inventory holding
facilities, supply chain redesign, and risk associated with investments in the supply
chain. These fall under the preview of the distribution, engineering, finance, and
operations departments. Effective communication is of utmost importance in ob-
taining information and developing and implementing a plan for the future. The
plans also consider the forecasts of the raw material availability and demand for
products and their prices. Hence, communication should be established at differ-
ent levels of a department and across the departments wherever possible for the
effective functioning of the organization. Communication across various depart-
ments along the chain of events from the raw materials being transformed into
products is called horizontal communication and that along the different levels in
a given department or group of departments is called vertical communication.
It is common belief that strategic decisions are generally initiated and approved
by the various departments before reaching the higher management in an orga-
nization [12–15]. However, the process of strategic decision making varies widely
from organization to organization and even from one project to another project
within the same organization. As the size of the organization and the complex-
ity of the supply chain increases, it becomes increasingly difficult to obtain a plan
and the decentralized way of decision making would not be appropriate. Hence,
the management relies on an overall perspective of an empirical, mathematical, or
financial analysis and the leadership involved while making these strategic deci-
sions.
Customer satisfaction and shareholder value can be increased by efficient man-
agement of the supply chain through implementing the latest technology, reducing
risk by strategic material supply contracts, exploring new markets via new invest-
ments, and exiting from less attractive markets. As has been mentioned before,
having information and using the information to make good business decisions is
the key to managing the supply chain efficiently. An analysis of the requirements
of the customers, forecast of the demand of the various products, availability of raw
materials, and the prices of all the goods is necessary and this is obtained from mar-
ket analysis. In addition, the market analysis provides further analysis on relative
capabilities of all the players in the market. The information from the market analy-
sis is used by the organization to develop a corporate strategy where the goals are
defined and the means of achieving the goals are outlined. Having the knowledge
of our organizations capabilities benchmarked against the competitors capabilities
will give us an insight into the probable cost and quality of the product that we can
sell, the flexibility in production capacity that we may want to have, and the delivery
time we can offer, and this helps us in developing the investment and production
plan so that we meet the goals of the organization.
9.3.2
Factors Influencing Investment Decisions
Factors that influence investments and disinvestments in the supply chain can be
classified as those that are quantifiable and those that are qualitative in nature.
Hence, two types of models are generally found in literature. They are nonmathe-
matical models which take into consideration the factors that are qualitative and are
not directly quantifiable, and mathematical models which use directly quantifiable.
We loosely classify the factors into qualitative and quantifiable factors; however,
there may be overlap in some cases.
Various factors that influence the decision of new investments were identified
and the advantages of locating a production plant in Singapore instead of Malaysia
were demonstrated in the literature [16]. The factors that influence a location choice
are, briefly, quality of life in the community, cost of living for employees (this influ-
ences the availability of skilled work force at low cost), community, and government
view of the type of industry – whether they are friendly or averse to this type of in-
dustry, ease and speed of compliance to regulations, infrastructure for the work
force such as housing, schools for children, health care facilities, recreational facil-
ities, availability of skilled workforce from educational institutes nearby, presence
of competitors and the market share one can possible have, strength of unions and
culture of productivity of people in the region – whether they prefer fewer working
hours compared to other workers in other parts of the world, safety from terrorists
and other antisocial elements, reliability of transportation to the region, stability of
the government among others.
Readily quantifiable factors are: cost of investing and operating production and
distribution centers in the area, forecast of demand for products and availability of
raw materials in the neighborhood, cost of transporting goods, government regu-
lations on import–export and pollution, taxes and tax incentives given by local gov-
ernments if any, employee compensation, cost of utilities, area of site and layout,
unemployment compensation in case of disinvestment among others.
9.4
Supply-Chain Redesign
9.4.1
Nonmathematical Models
Several researchers have presented structured ways of analyzing the qualitative fac-
tors and have shown how the efficiency of the supply chain could be improved by
making good investment decisions. Analytical hierarchy process combined with
quality function deployment was used to make strategic decisions, and the analy-
sis of how an unnamed organization chose Singapore as opposed to Bangkok or
Beijing was presented by researchers [17]. This decision happened to be contrary
to popular belief that a facility should be located in or moved to a location where
operating costs are low and the study had thus demonstrated the need for analyt-
ical approaches in strategic decision making. A long-term capacity management
framework considering an empirical analysis of the relationship between strategic
planning and tactical planning was presented in literature [18]. In addition, stan-
dard operations management text books describe strategic planning and tactical
planning in greater detail.
The setup or relocation of a production facility is a lengthy project which re-
quires time-phased planning. There are different phases while searching for a new
location. Once a decision is made at the senior management level to expand pro-
duction capacities, focus of the analysis shifts toward identifying possible facilities
that could be expanded and/or locations where a new investment could be made.
After this, a master list of criteria that an investment should meet is made. For each
potential location, weights are given for all the criteria and the total weight for each
of the locations is calculated. The investment choice that obtains the greatest score
is selected. This analysis also gives results that can easily be compared against one
another. There are different types of weighing methodologies and describing such
methodologies is beyond the scope of this chapter. Concurrently, the cost benefit
analysis and feasibility study is carried out either for each location or for the few
important locations, depending on the insight that is required. Strategies for ca-
pacity management can be divided into those that are related to strategic planning,
tactical planning, and operational planning. We briefly present them below.
9.4.1.1 Strategic Planning

A long-term capacity plan considers the uncertain demand in the future, typically
in timescales of 5–10 years or even decades and considers decisions regarding ca-
pacity expansion, reduction, relocation, and new capacity planning. New capacity
planning is also called as facility planning and it is a study that determines how
much long-range production capacity is needed, when this additional capacity is
needed, where the production facilities should be located, and the layout and the
characteristics of the facilities. While making a long-term capacity plan, an excess
production capacity is usually allocated and the production facility is operated at a
capacity utilization that is less than the total capacity. This is done in order to have
a buffer capacity, which can be used in times of production losses or unexpected
increase in demand. If it happens that the plant is operating at full utilization con-
tinuously over several time periods, a new capacity should be considered. This pro-
cedure of wait-and-see to build a new capacity leads to loss of potential sales as
it does not consider the forecasts and hence gives a suboptimal solution. Other
decisions in long-term capacity planning include the strategic decisions which de-
cide the quality of the product and the process/technology that should be used –
whether a high technology or a low technology process is to be selected from the
list of available technologies.
The strategy for capacity can be classified as capacity leading demand, capac-
ity lagging demand, and capacity tracking demand. These strategies would be the
same whether an existing facility is expanding its capacity or a totally new facility
is being planned. For a capacity increase, the strategies are explained below. The
lead, lag, and track strategies are described in Fig. 9.2. Briefly,
Lead: At any given time, the capacity of a facility is greater
than or equal to the demand. There is flexibility in
production due to over capacity and the costs are high due to
the unutilized capacity. This strategy is followed by extreme
risk takers and the new capacity addition should be planned
well in advance and even when the utilization is not 100%.
By the time the demand would have reached the current
capacity leading to 100% utilization, the additional capacity
would become operational and hence, the net capacity would
still be greater than the demand and the utilization would be
less than 100%.
Lag: The capacity is always lesser than or equal to the
demand. Hence, utilization is the highest (100%) and when
the demand increases, there is loss of potential sales. A new
capacity addition is planned such that when it is available for
production, the demand would have increased to a value
greater than the new net capacity and the utilization would
still be 100%.
Track: The capacity keeps a close track of the demand by
means of small capacity additions. Hence, the utilization is
a)
Timing of capacity
addition Amount of capacity
change
Capacity
Demand curve
Time
b)
Demand
curve
Capacity
Timing of capacity Amount of

change capacity change
Time
c)
Timing of capacity
change Amount of
capacity change
Capacity
Demand
curve
Time
Fig. 9.2 (a) Capacity strategies – lead strategy, (b) capacity

strategies – lag strategy, and (c) capacity strategies – track
strategy.
close to 100% and there are fewer lost sales due to lower
capacity and times of underutilization of capacity.
In all the three strategies described above, the time at which the construction for
the capacity expansion starts depends on the rate of increase of demand and the
construction lead time for the facility. Intuitively, a track strategy would be ideal, if
we consider each facility individually. The capacity expansion problem deals with
the sizing problem in capacity, its type, the timing of capacity, and also the location.
Factors influencing the location decisions are discussed later. The sizing problem
involves determining the optimal step change in the capacity. Timing of capacity
is more difficult due to the difficulties associated with the forecasting of demand
and hence, tracking option is sometimes followed. If there is tracking, the empha-
sis shifts to sizing problem. Once the location is fixed, the timing and the sizing
are interrelated. Assuming that the demand is increasing continuously for long
periods of time, an early investment would require a small capacity addition and
another small addition would be required at a later time. However, if there is delay
in making a decision at the first instance, the decision maker may then decide to
have one single large capacity addition instead of two small additions. It should
be noted that the time when this single large capacity is added would be different
from the times at which the two small additions happen.
The capacity expansion decision should be made by firms during the short-
time window of investment opportunity preceding an uncertain growth in mar-
ket size. There are two kinds of information. First, an industry outlook on the
emerging market that is common knowledge to all the firms in the industry, for
example industry market report, government publications, and so on. Second, self-
understanding of the future demand derived from the organization’s own market
research. A capacity planner must decide how much capacity to build and when
to build it, keeping in view the response of his competitors. Planners must bear
in mind that his capacity choice and timing directly influence aggregate industry
capacity and excess capacity in the industry directly influences his capability and
the industry as a whole.
After the capacity choices are made, the market size is revealed (other players
will do an analysis) and the firms engage in capacity-constrained price competi-
tion. Two scenarios are possible. Simultaneously or sequence capacity expansions
are possible and the players should come under some sort of equilibrium in order
to avoid industry over capacity. On the basis of existing demand forecasts, different
levels of probability from high to low may be considered. First move may be advan-
tageous as this helps firms signal to others that they will start operations soon and
others may be taking a risk. Second mover then always does a follow up and adjusts
his capacity expansion calculations based on the first’s capacity. Or, it may turn out
that the second waits until demand is certain and then adds capacity if required.
The above process of decision making by considering one facility at a time does
not take into consideration the information about the rest of the supply chain, and
hence the affect that this facility may have on the rest of the supply chain is ig-
nored. Moreover, when following a track strategy, several small increments in ca-
pacity may be thought of as being appropriate. However, considering the amount
of work involved in building facilities and the construction lead times, among oth-
ers, organizations would prefer to have fewer investments distributed in time at
any single site. On the other hand, a single large investment at an early stage may
be inappropriate. These factors are called the economies and diseconomies of scale
and are crucial for the efficient management of the facility. For example, an air-
craft manufacturer may want to build one large facility where all the parts of the
aircraft are assembled, instead of satellite assembly units and in this context, large-
scale operations help in the reduction of costs and in better management. However,
the same is not true with hospitals. Building one large hospital instead of several
smaller ones would lead to long commutes to this central hospital and may even
lead to chaos as patients may find it difficult to locate the exact department/section
that they need to visit. Hence, small hospitals that are geographically distributed
are preferred. Organizations should hence weigh the options between expanding
an existing facility and building a new facility. It can be seen that a capacity addi-
tion should be in between a certain minimum and maximum and that the location
plays an important role. Organizations may also consider building a new facility
close to an existing facility, so that the operations are better managed. This would
also help the smooth integration of the new facility with the existing network. Ca-
pacity planning also involves the decisions regarding the disinvestment of facilities
and the strategies for disinvestments also are similar. However, in this work, we
consider that a facility is completely disinvested instead of in small steps, as this
is typical of chemical industries. In the strategic plan, aggregate data for the raw
material availability, product demands, and prices of the goods are considered in-
stead of detailed data. In other words, demand for all products under a family can
be grouped as one, for example different versions of a same model of a car.
9.4.1.2 Tactical Planning

Once the strategic plan is decided, the details of the capacities across the supply
chain network and the contracts that are in place for raw material supply and prod-
uct sales will be known. In the tactical plan more details with respect to the raw
material purchase, production rates, distribution of goods, work force levels and
inventory levels, capacity limits, demand for products, and any other customer re-
quirements are considered. The tactical plan has the targets of maximizing the
profit and customer satisfaction, minimizing inventory, while optimally using the
facilities with minimal changes in production rate and work force levels.
On the basis of market analysis and the strategy of the organization, the manage-
ment decides to either follow the push method or the pull method of production for
operation during the tactical decision-making process. Push method is a process
where the products are manufactured hoping that the customers would buy them.
The driving force comes from the production team and examples of a push sys-
tem would be all the products that are available on the shelf such as the consumer
goods. Pull method is a process where the manufacturing is done on demand and
is customer driven. An example of the pull method would be the manufacture of
cruise liners, aircraft among others. Further, the production of goods can be classi-
fied into different categories such as MTO, MTS, ETO, and ATO, which are based
on the type of industry that our organization belongs to, the strategy of the organi-
zation, and the arrival of the orders. Briefly,
MTO is make-to-order, for example a particular golden
pendant.
MTS is make-to-stock, for example consumer goods.
ETO is engineered-to-order, for example custom designed
furniture, houses, and so on.
ATO is assembled-to-order, for example personal computers.
There is another way of classifying the production strategy and this can be classified
into three types. They are as follows:
Level: Constant production rate over the planning horizon
irrespective of how the demand changes. This helps in
reducing the costs due to frequent changes in the rate of
production, but may not be able to capture sales due to
increasing demand and also has the risk of over production
when the demand is decreasing.
Chase: Production rate is frequently modified for it in order
to be in tune with the demand. In facilities where the cost of
changing the production rate is high, this may not be a good
option. In order to follow a chase case, the facility should
always have excess capacity, and this may sometimes be too
costly.
Mix: This is a case where both level and chase type of
production strategies are followed. This is particularly useful
when the organization chooses to neglect small fluctuations
in demand, or meet these small fluctuations by making use
of the inventory that may be present as a result of level type
of production, but wishes to change the production rate
when there is large deviation in expected demand. The
decision of when to switch between the chase and level case
depends on the magnitude and frequency of these
fluctuations in demand and also on the cost of changing the
production rate.
A somewhat loose correlation exists between the above two types of classifications.
A level case may be a make-to-stock type of strategy. A chase case may be a make-
to-order and engineered-to-order, where the production rate is in tune with the
demand. Assemble-to-order may be considered as a mix case, where part of the
facility that produces the spare parts can follow a level case and the part of the
facility that assembles the parts may follow the chase case.
9.4.1.3 Operational Planning

In order that the goals of the organization are realized, the operational planning
is of equal importance as the strategic and tactical planning. Operational plan-
ning assigns resources to get certain tasks done and this is also called schedul-
ing. There are two types of scheduling. First, work force scheduling – this deals
with the process of employees being assigned to any work while considering the
total skill set of the employees and not individual employees. Second, operations
scheduling – this assigns a particular employee to the work that needs to be done
and/or assigns particular job to particular resource. A work force schedule gives an
overall understanding of the employee skills available to get the tasks done. Oper-
ation schedules are short-term plans that are drawn in order to execute a master
production schedule obtained through tactical planning. It deals with how best to
use existing resources and capacity to perform the given tasks considering the var-
ious constraints. In the case where the schedule cannot meet the targets set by
the tactical plan, which may be due to the various constraints and/or uncertain-
ties that may arise, the tactical plan is recalculated and new schedules are obtained
for the future. This feedback between different planning levels is important. While
performing scheduling, the factors that are to be considered are the setup times,
process time, materials handling time between operations, waiting time, and the
resources/utilities available.
9.4.2
Mathematical Models
The focus of this chapter is developing tools that are quantitative in nature and we
briefly present a literature review on the mathematical models in capacity manage-
ment.
9.4.2.1 Strategic Planning

Several researchers have considered the long-term capacity management problem.
A comprehensive review on the design and operation of supply chains with a
greater emphasis on financial strategies was presented [19]. They discuss several
issues such as international tax planning, means of financing investments, trans-
fer pricing, and uncertainties. In addition, they discuss various modeling and so-
lution strategies such as simulation-optimization, real options, multistage models,
among others. A capacity expansion model that handles capacity increments in dis-
crete amounts (defined a priori) or multiples of this discrete amount was presented
[20, 21]. A capacity expansion model for the Korean petrochemical industry was
presented [22]. An integrated investment model with production distribution was
presented in [23]. Capacity expansion and production distribution models for the
pharmaceutical industries were presented [24–26]. A postimplementation analysis
of the plan for upgrading technological capabilities of a petrochemical firm was
presented [27]. Nonlinear models to consider capacity expansions were presented
[28, 29]. A strategic capacity planning model to analyze the effect of regulatory fac-
tors in this era of globalization was presented [30].
Partial relocation of facilities was considered by some and they allowed only those
facilities that existed at the beginning of the horizon to relocate [31]. Also, they dis-
cussed a single facility being relocated to more than one location. These are limita-
tions for chemical production facilities because partial relocation/disinvestments
and relocation to multiple new locations in small capacities are not practical.
A dynamic facility location model for combat logistics was studied [32]. The
characteristics of this problem are also different from those in chemical supply
chains. Hence, a different model is required. In our model, we consider the re-
location/disinvestment of future facilities also and allow only complete reloca-
tion/disinvestment.
9.4.2.2 Tactical Planning

Several researchers have considered the problem of short-term capacity manage-
ment that is production and distribution planning. Probably, the most comprehen-
sive model to address various issues in the supply chain of the electronics industry
was presented [33]. Duty drawbacks are becoming increasingly important due to
globalization and competition between organizations. A production distribution
planning model considering duty drawbacks was presented [34]. An integrated
planning model considering production capacity and inventory for a system with
seasonal demands was considered [35]. Some other production allocation and dis-
tribution models are presented [36–38]. A nonlinear model for transfer pricing in
the supply chain was presented [39]. A global logistics systems model considering
different logistics options was presented [40]. A comprehensive strategic planning
study on R&D pipeline management using a simulation-optimization framework
was presented [41–43]. A model for production-distribution planning of multiprod-
uct batch plants and applied model predictive control strategy on a rolling horizon
was presented [44]. A multiobjective supply chain design model considering capac-
ity additions was presented [45]. An enterprise-wide production and distribution
model considering various operational policies was presented [46]. A production
planning and distribution model by grouping products into families and consid-
ering transportation times was presented [47]. An integrated production and dis-
tribution model was considered [48] and this was solved by using aggregated time
periods.
9.4.2.3 Contracts for Material Supply

Contracts involving the use of computer and electronic equipment were consid-
ered [49]. A cash management/budgeting model for planning and scheduling was
considered [50]. A mathematical model to study options contracts in the chemical
industries for short-term planning was presented [51]. These works deal with oper-
ational/tactical issues in contracts. An empirical analysis of futures and contracts
in the chemical industries considering strategic decisions was presented [52]. A
comprehensive analysis on the contracts in the supply chain was given [53, 54].
Finally, research on contracts in the petrochemical industries was presented [55].
9.4.2.4 Remarks
Although there has been extensive research in the area of capacity management,
there are some gaps in the literature, which are one or more of the following:
1. Discrete amounts (multiples of fixed amount) are considered
for investments instead of continuous amounts. Continuous
amounts allow any possible capacity to be invested within
some limits.
2. Lower limits on new constructions and expansions of an
existing facility are the same. In reality, the lower limit on a
new construction is usually much larger than the lower limit
on expansion at an existing location.
3. Existing facility is assumed to always operate. This leads to

no allowance for shutdown for maintenance.
4. The lower limit on throughput is zero in order to account for
cases where no raw material is available. This is not realistic.
5. Highly nonlinear and complex formulations are used, which
do not guarantee global optima.
6. Only expansions are considered and not disinvestments.
7. Strategic material supply contracts and options to raise
capital via loans and bonds are not considered.
In addition to the above-mentioned methods of analysis, capacity management in-

volves forecasting data and using these data to make optimal business decisions.
Since forecasts may not always be accurate, a high amount of uncertainty and,
hence, risk is involved. In order to solve problems that have uncertain parameters,
the following methods are considered.
1. scenario-based approaches, which are characterized by
discretization over the uncertain parameter space,
2. multistage stochastic programming, where an algorithm is
used to transform the stochastic optimization problem into a
series of deterministic problems, and
3. mathematical models that account for robustness.
Uncertainties can be classified into three levels and scenario-based approaches

could be used to analyze the risk due to uncertainties [56, 57]. They are (1) general
environmental uncertainties – political, natural disasters among others; (2) indus-
try uncertainties – demand, supply fluctuations among others; and (3) company or
operational uncertainties. Scenario analysis is one of the widely used methods of
risk analysis, particularly when there are investments and disinvestments in supply
chains, while robust mathematical models are widely used for investments in the
stock market. In this chapter, we present a scenario analysis approach to analyze
how environmental uncertainties, which pose the highest level of uncertainties in
the chemical industries, could affect the supply chain.
9.4.3
This Chapter
In the present chapter, we present a novel deterministic mixed-integer linear pro-

gramming (MILP) model for supply chain redesign, which addresses all the above
issues. We also combine both the strategic and tactical planning models into a sin-
gle supply chain redesign problem, as we believe that this approach handles several
key decisions in an end-to-end supply chain of a multinational corporation (MNC)
to a finer detail, when compared to solving each problem individually. In addition,
the problem is formulated with the ability to incorporate production changeover
costs, depreciation, tax and regulatory factors, transportation costs, storage costs,
production costs, inflation rate, and loans and bonds for raising capital for invest-
ments. The model aims to maximize the net present value (NPV) of the total as-
sets, including the immovable properties such as facilities, thus maximizing share-
holder value.
9.5
Problem Description
A chemical supply chain of an MNC typically consists of raw material suppliers,

production facilities, input and output inventory facilities at the production facili-
ties, and inventory facilities at the distribution centers. The raw materials are pur-
chased from the suppliers, transported and stored at the input inventory facilities
of the production facilities. The production facilities process the raw materials into
products and intermediates, which are then stored at the output inventory facili-
ties. The recycled material is sent back to its corresponding input inventory facility,
while the intermediates are sent to the respective input inventory of the down-
stream production facilities. The products are sent to the distribution centers. The
MNC can sell its materials from any of its inventory facilities. Hence, the inventory
facilities at the production centers cater to the local markets whereas the distri-
bution centers cater to the geographically distant markets. There may be several
production facilities at a single location. In that case, each facility is considered to
be a distinct facility, while their logistics attributes would be similar.
Organizations usually rely on strategic contracts for both raw material purchase
and product sales. This is in order to reduce the risk associated with uncertain-
ties in the supply chain and markets. Contracts on raw material supply are signed
to hedge risk against unavailability or high price of raw materials, while those
on products are signed to hedge risk against falling prices or reduced demand.
Contracts consider quantity flexibility, accounting for economies of scale [38], and
penalties for under supply. Key decisions that the management considers before a
contract is signed are the time when the contract should come into effect and the
duration for which it should be signed.
Investing requires capital and MNCs have to consider how to raise it. They may
either invest their profits or consider raising capital through loans and bonds. In
our model, we include some constraints for raising capital via loans and bonds.
In addition, we consider the possibility of a facility undergoing a technology up-
grade in order to produce a better product, which may reduce the production costs
and/or increase the customer satisfaction. This technology upgrade may not lead to
capacity addition; however, organizations may choose to add capacity and perform
technology upgrade at the same time. As described previously, we include more
features like depreciation, production change over costs, production costs, storage
costs, transportation costs, import–export tariffs, interest rate for loans and bonds,
inflation rate, and after-tax profit. We develop a discrete-time formulation where the
entire planning horizon is divided into a fixed number of periods of equal length.
We now describe several aspects, features, and assumptions related to various
redesign decisions. To this end, an interval means a group of periods; construc-
tion refers to the addition at a previously nonexisting facility; expansion refers to

the capacity addition at an existing facility; and throughput factor is the ratio of
throughput to capacity.
Investments and disinvestments:

1. The lower limit on construction is different from that of an
expansion.
2. Investments and disinvestments have a fixed and variable
term for the costs/income. The variable term is proportional
to the capacity and the fixed term accounts for other possible
costs.
3. The time between the start of construction and start of
operation of each facility (construction lead time) is fixed and
is known.
4. Expansion does not affect the production. An expansion
cannot be started while a previous expansion or construction
is in progress, that is only one addition is allowed at any
given time at a given facility.
5. There are limits on the total amount of investment and
disinvestment in a given interval (e.g., the amount of
investment that can be made in a year is 500 million).
6. A facility has to be disinvested completely. Even when the
facility reached its current capacity by several small capacity
additions, the disinvestment results in selling off the entire
capacity.
7. A technology upgrade results in a change in the product
specifications. The stoichiometry for the upgraded facility
may be different when compared to the original facility.
Production allocation:
1. There is an upper and lower limit on throughput if a facility
is in operation and the throughput is zero if a facility is not
in operation. The throughput is the total weight of all the
material that the facility processes at any given time.
2. The upper limit on the sales is the demand. The excess
material that is produced can be sold at a discounted price,
which is user defined.
3. A production facility can operate for only a fixed number of
periods in any given interval. This is to allow time for
maintenance.
4. Starting and shutting down production (changeover) for
maintenance or otherwise involves fixed costs.
5. Back orders (negative inventory) are not fulfilled.
6. Once a facility undergoes a technology upgrade, the

throughput may change. The change in throughput as a
percent of the previous capacity is user defined.
7. Production costs have two terms. The first is proportional to
the amount of material processed and the second to the
capacity of the facility. Hence, the facility incurs a cost that is
proportional to the amount of products produced and a cost
even if it is not in operation.
Distribution and sales:

1. The demand and sales price for the technologically better
products are different from that of the original products.
2. Inventory holding facilities at the production facilities act as
sales locations catering to the local markets while the
distribution centers cater to the geographically distant
markets.
3. Transportation time for the movement of materials is
assumed to be negligible as each period is of the order of a
few months in this strategic decision-making formulation.
4. Once the material is sold, the seller does not bear any further
cost, which means that purchase of raw material and sales of
products are as free on board (FOB) [58]. Hence, cost,
insurance, and freight (CIF) and tax for raw materials are
borne by the MNC. CIF and tax after product sales are borne
by the customer.
5. Storage costs have two terms. The first is proportional to the
amount of material stored and second to the capacity of the
facility. Hence, the facility incurs a cost that is proportional
to the amount of material stored and a cost even if it is not
storing any material.
Contracts, loans and bonds:

1. Contracts include economies of scale and penalties for under
supply. The duration for which a contract is to be signed is a
decision variable.
2. A loan is repaid in small installments and these installments
account for the principal and the interest. A principal and the
interest for a bond are paid at the end of the maturity period.
3. The loan installments may be different from one period to
another and hence, any profile for the loan repayment can be
incorporated. This amount as a fraction of the actual loan is
user defined.
4. A material can be purchased either within or outside the
terms of the contract from a given supplier. No quantity
flexibility and duration of purchase exist if a contract is not

in place.
Miscellaneous:
1. The rate of depreciation varies between facilities and from
one period to another. It is different for an existing facility
and the capacity addition. However, all additions are
depreciated at the same rate.
2. Different countries have different interest rates to account
for inflation. This rate varies from one period to another.
3. There are limits on transportation and storage of material.
4. Import–export tariffs and other regulatory factors are
included in the transportation costs. Since transportation
costs are proportional to the amount of material transported,
only variable costs and no fixed costs are considered for
tariffs.
5. There are several types of uncertainties involved in the
supply chains as have been described previously.
9.5.1
Problem Statement
The management desires a capacity management and supply chain redesign plan.
This comprises plans for raw material purchase, sales; production, material distri-
bution, inventory plan; investment, disinvestment, technology upgrade plan; and
contracts, loans, and bonds plan.
The demand profiles for all products from all customers for both original and
technologically upgraded products, raw material availability profiles from all sup-
pliers; price forecasts for raw materials and products; transportation rate, insur-
ance, import–export tariffs; production rate for original and technologically up-
graded products, storage rate, production changeover costs; depreciation rate, in-
terest rate for loans and bonds, inflation and tax rates in different nations; con-
struction lead times, rate of investments and disinvestments, cost of technology
upgrade are assumed to be known.
Subject to the following constraints: capacity utilization limits, limits on investments

and disinvestments, constraints for material balance, inventory limits, and im-
port/export constraints. The objective is the maximization of the shareholder value
(NPV). We have previously presented preliminary versions of the formulation [59,
60] and present here the detailed model and results.
9.6
Problem Formulation
To formulate this capacity management and supply chain redesign problem, we

use a uniform discrete-time representation with a prefixed number (H) of iden-
tical periods. The first period begins at time 1 and all decisions are made at the
beginning of each period. We view the MNC’s multiechelon supply chain as a set
of material suppliers, a set of materials, a set of production facilities, a set of inven-
tory holding facilities at production sites, and a set of inventory holding facilities
at distribution centers. The complete formulation involves several other features
as described in the problem statement and hence, we present it as different parts.
They are constraints related to capacity, inventory, contracts, loans, bonds, costs,
and incomes.
9.6.1
Capacity Change
In order to model the capacity change, we include all the facilities in the net-
work – both the existing and the future possible facilities. At the beginning of the
planning horizon, some facilities exist and some do not. We use binary variables
FEYft , FSYft , FYft , and FXft to model the supply chain redesign. These variables are
defined in the section on nomenclature.
Due to the time required for construction of a facility, that is construction lead
time (CTf ), a facility cannot expand (FYft = 1) more than once in any interval of
length (CTf ).

t
FYft 1 t > CLTf . (9.1)
t−CLTf +1
If a facility f is not under construction at the beginning of the horizon, then no

capacity addition is possible for t CLTf . Hence, the above equation is written
only for t > CLTf . However, if a facility is under construction at the beginning of
the horizon, then FYft can be fixed accordingly.
In practice, a few large capacity additions would be preferred over several small
ones because a capacity addition involves a lot of construction work and it is con-
venient to have fewer additions. Although a technology upgrade may not lead to a
change in capacity; for ease of representation in the mathematical model, it can be
considered as being similar to capacity addition as some equipment are modified
and hence, the plant may have to be shutdown. Therefore, it may be desired to limit
the number of capacity additions and technology upgrades for a facility by using

FYft + TUFt NEf t > CLTf , F ∈ UP, t > TDf . (9.2)
t
Now, if a facility f exists (FEYft = 1) at time t, a capacity addition (FYft = 1) does

not change its status, but a disinvestment is possible at time t + 1 (FXft+1 = 1). If
a facility f does not exist at time t (FEYft = 0), then a capacity addition is possible
(if this facility is not a disinvested facility because a disinvested facility cannot be
purchased back – FSYft is used to represent this) at time t + 1 (FYf t+1 = 1), but
a disinvestment is not possible because the facility does not exist in the first place.
Next, we transform this explanation into constraints.
A facility f cannot change its status (exists or does not) for t min[CLTf , TDf ]
and this is given by Eq. (9.3). All equations with labels starting with “F” are not
written as explicit constraints, but they are meant for fixing variables.
FEYft = FEYf,1 t min[CLTf , TDf ]. F (9.3)
If a facility starts production due to a capacity addition at time t, but did not exist at
time (t − 1), then it should be taken as existing at time t. In other words, FYft = 1
and FEYf (t−1) = 0 ⇒ FEYft = 1,
FEYft FYft t > CLTf , (9.4)
FEYft FEYf (t−1) + FYft − FXft

t > min[CLTf , TDf ], t > CLTf , t > TDf . (9.5)
Equation (9.5) also ensures that a facility cannot be disinvested at t, if it does not
exist at (t −1), and it can exist at t, only if it either existed at (t −1) or it adds capacity
at t. Similarly, it also ensures that a facility cannot add capacity and disinvest at the
same time.
Also, a facility existing at time (t − 1) must either exist or be sold-off at time t
(Eq. (9.6)).
FEYf (t−1) FEYft + FXft t > TDf , t > min[TDf , CLTf ]. (9.6)
Once a facility is disinvested, it cannot be purchased back and should remain sold-
off for the remainder of the horizon. Equations (9.7–9.9) ensure this.
FSYft FXft t > TDf , (9.7)
FSYf (t+1) FSYft H > t > TDf . (9.8)
A facility cannot exist and be sold-off at the same time.
FSYft + FEYft 1 t > TDf . (9.9)
Once a facility undergoes a technology upgrade (TUft = 1 ⇒ UFft = 1), it is taken
as an upgraded facility (UFf (t−1) = 1 ⇒ UFft = 1). If a facility is not undergoing
an upgrade at the beginning of the horizon, then it cannot be upgraded before its
construction lead time for upgrade (TUPf ). However, if an upgrade is in progress
at the beginning of the horizon, then the variables must be initialized appropriately.
UP is the set of facilities that can undergo a technology upgrade.
UFft = UFf (t−1) + TUft t > TUPf , (9.10)
UFet = 0 t TUPf . F (9.11)
Capacity of a facility at time t is related to the capacity addition and the disinvest-
ment. To model capacity addition, we write
faft fcapft − fcapf (t−1) + fcapU

f FXft
t > min[CLTf , TDf ], t > TDf , (9.12)
faft fcapft − fcapf (t−1) t > min[CLTf , TDf ]. (9.13)
The last term in Eq. (9.12) is to accommodate the disinvestment of a facility at t,
which will make its capacity zero (fcapft = 0), and to ensure that capacity expan-
sion remains zero instead of becoming negative. Note that disinvestment is not a
negative capacity addition in our formulation. Investment and disinvestment are
modeled separately and “fa” is a positive variable.
Since no capacity addition is possible during the initial period of the horizon,
which is equal to the construction lead time of the facility, we write
faft = 0 t CLTf . F (9.14)
The capacity increase for a facility must be zero, when no capacity is added at time
t.
faft fcapU
f FYft t > CLTf . (9.15)
The lower limit on an addition at an existing site and that of a green-field invest-
ment is different and we model this in Eq. (9.16).
faft FAN O
F (FYF t − FEYF (t−1) ) + FAf (FEYf (t−1) + FYft − 1)
F ∈ PIN, t > CLTf . (9.16)
The capacity of a facility must be zero, when the facility does not exist, and must
be the same as that at the beginning of the horizon for the initial interval, during
which no investment or disinvestment is possible.
fcapft fcapU
f FEYft t > min[CLTf , TDf ], (9.17)
fcapft = fcapf,1 t min[CLTf , TDf ]. F (9.18)
9.6.2
Capacity Utilization

The amount of a material supplied ( m mdmsit ) by a raw material supplier to var-
ious input inventory facilities at production facilities must equal the amount of
material purchased.

mdmsit = dsmst + dcrmst im ∈ FM, msi ∈ MFF. (9.19)
i r<r c
In the above, subscript “m” refers to the raw material, “s” refers to the supplier, and
“i” refers to the receiving facility. For each contract, we have ranges for purchase
amounts with different prices. “r” is the index that denotes the range in which the
purchase amount lies, and “r c ” is the cardinality of “r.” MFF is the set of arcs that
“m” can take between “s” and “i,” FM is the set of materials that s can sell, and S
is the set of raw material suppliers. Furthermore, only one term on the right-hand
side of Eq. (9.19) is nonzero at any given time depending on whether a contract is
signed for that material or not. More on contracts is explained in Section 9.6.4.

The amount ( m mdmFft ) of materials that a facility can receive/process cannot
exceed facility throughput.

U
mdmFft TFU U U
ft fcapft + fcapf TU (TFf TU + TFUf TU )UFf TU t
mF
f TU ∈ UP, mF ∈ MF, t > TUPP , mFf ∈ MFF, (9.20a)

U
mdmFft TFUU U U
ft fcapft + fcapf (TFft + TFUft )(1 − UFft )
mF
mFf ∈ MFF, f m ∈ FM, f TU ∈ UP, t > TUPf , (9.20b)
where throughput factor (TFU
ft ) is defined as the ratio of throughput to capacity
(fcapft ). The last terms in each of Eqs. (9.20a) and (9.20b) account for technology
upgrade. UFft = 1 if a facility undergoes a technology upgrade, else UFft = 0. IIN
is the set of inventory facilities associated with a production facility. PI is the set of
facilities.
If a production facility is in operation (FZpt = 1), then the amount that it
processes must exceed the lower limit on allowable throughput (TFL ft fcapft ). The
first terms in the following equations ensure this. However, the throughput of a
facility may change after a technology upgrade. The last terms in these equations
account for the change in throughput after technology upgrade. P is the set of
production facilities.

U
mdmipt TFL pt fcappt − fcapp (1 − FZpt )
mi
− fcapUP (TFU U
P t + TFUP t )UFP t
pm ∈ FM, mip ∈ MFF, P ∈ UP, t > TUPP , (9.21a)

U
mdmipt TFUL
pt fcappt − fcapp (1 − FZpt )
mi
− fcapU U U
p (TFpt + TFUpt )(1 − UFpt )
pm ∈ FM, mip ∈ MFF, p ∈ UP, t > TUPp . (9.21b)
Equations (9.21a) and (9.21b) are meant for production facilities only, but one could
write similar equations for inventory facilities also. Equations (9.20b) and (9.21b)
are written only for the facilities that could undergo a technology upgrade.
Now, if a production facility is not in operation, then its throughput must be
zero. We assume that inventory facilities always operate, and hence, we do not
write similar equations.

U
mdmipt (TFU U
pt + TFUpt )fcappt FZpt pm ∈ FM, mip ∈ MFF. (9.22)
mi
A facility cannot operate (FZpt = 0) if it does not exist (FEYpt = 0).
FEYpt FZpt . (9.23)
A facility should be shutdown (FZp(t−1) = 0) for some time in order to integrate

the new added capacity with the old capacity, before it becomes operational (FYpt =
1). We assume that this time is an integer multiple of period length. Thus, if this
time is equal to one period length, then we can write
FZp(t−1) 1 − FYpt t > CLTp . (9.24a)
If it is longer than one period, then the above equation can be modified appropri-
ately. Similarly, a facility should shut down for integration, when there is a technol-
ogy upgrade. Assuming that this time is also equal to one period, we can write
FZf (t−1) 1 − TUft f ∈ UP, t > TUPf . (9.24b)
Facilities cannot operate nonstop and they must shut down for routine mainte-
nance. For example, a facility may shut down once every 18–24 months. If STp is
the maximum number of periods that a facility can operate in an interval of length
PTp periods, then

t
FZpt STp t PTp . (9.25)
t−PTp +1
An inventory facility may store more than one material, for example, when ma-
terials are containerized in drums and boxes and share an inventory facility. The

amount ( m mfmf t ) of materials that the facility holds at time t cannot exceed its
holding capacity (fcapft ).

mfmf t fcapft f m ∈ FM, f ∈ I. (9.26)
m
Note that Eq. (9.26) is written only for inventory and not for production facilities
because the latter does not store any material.
9.6.3
Material Balance
A mass balance on an inventory facility includes the holdup at time (t −1), material
received at time t, material shipped to other facilities at time t, and sales (at both
actual and reduced sales price) at time t. Initialization for holdups can be done
appropriately.

mfmft = mfmf(t−1) + mdmFft − mdmfFt − sMdt − wMot
F F
mfF ∈ MFF, mFf ∈ MFF, fm ∈ FM, M ∈ PM, o ∈ IO. (9.27)
Equation (9.27) includes an allowance for selling excess material at a lower price
than normal price or disposing it at some negative price (or positive cost). We
achieve this through variable wmot . It avoids the situation in which a material oc-
cupies storage unnecessarily. Products can be sold at the output inventory holding
facilities associated with the production facilities as well as the distribution centers.
However, in the example which we present later, we allow sales only at the distri-
bution centers. Logically, there is no incentive in transporting the waste/excess
materials to the distribution centers and hence, we do not write the waste disposal
term for the distribution centers. If the sale price of waste at distribution centers
is sufficiently high, then we have to write these terms for the distribution centers
also.
For a production facility, materials received from associated input inventory facil-
ities, processed, and sent to output inventory facilities are in stoichiometry ratios.
We write equations to relate input and output materials [Eqs. (9.28a–9.28e)] and to
relate input and recycle materials (Eq. (9.29)).
Equation (9.28a) is written for all the materials other than those that may un-
dergo a technology upgrade.

i mdmipt mdm pot
= o
ψmp ψm p
im ∈ FM, pm ∈ FM, m ∈ (M − UM − UMM). (9.28a)
Equations (9.28b) and (9.28c) are written for the materials that may undergo a tech-
nology upgrade and it corresponds to materials after undergoing a technology up-
grade.

i mdm ipt o mdmpot i mdm ipt
+ fcapU
p (1 − UF pt )
ψm p ψmp ψm p
− fcapU
p (1 − UFpt ) m ∈ UMM, im ∈ FM, im ∈ FM, (9.28b)

mdmpit fcapU
p UFpt m ∈ UMM, im ∈ FM. (9.28c)
i
Equations (9.28d) and (9.28e) are written for the materials that may undergo a tech-
nology upgrade and it corresponds to materials before undergoing a technology
upgrade.

i mdm ipt o mdmpot i mdm ipt
+ fcapU
p UF pt − fcapU
p UFpt
ψm p ψmp ψm p
im ∈ FM, m ∈ UM, im ∈ FM, (9.28d)

mdmpit fcapU
p (1 − UFpt ) pm ∈ FM, m ∈ UM, (9.28e)
i

mdmipt i∈IINR mdm pit
i
= pm ∈ FM, m ∈ RCM, pm ∈ FM. (9.29)
ψmp ψm p
Here, i refers to the input and o to the output inventory facilities that are linked to
a given production facility p. m refers to the set of materials that a given produc-
tion facility can process. ψmp is the stoichiometry coefficient of material m that is
processed in a given production facility p. IIN is the set of all input inventory fa-
cilities and OIN is the set of all output inventory facilities that are linked to a given
production facility. IINR is the set of input inventory facilities that take recycled
material from a given production facility.
9.6.4
Supply Contracts
A material supply contract with a supplier specifies several prices based on the
amount of material purchased. The number of (r c −1) ranges in which the material
can be purchased is predefined and the total amount of material purchased must
fall into one of the ranges. We use binary variable Vrmst to identify purchase within
a range r and (r + 1), and the amount dcrmst of purchase in that range must fall
within the quantity limits of that range.
Vrmst QCrmst dcrmst Vrmst QC(r+1)mst r < r c , m ∈ RM. (9.30a)
If a contract does not exist for a material with a supplier, then the material must be
purchased from outside the contract from that supplier. Organizations may wish
to limit purchases without contracts, so the amount (dsmst ) of such purchases must
not exceed some prespecified limits.

dsmst QSmst 1 − Vrmst m ∈ RM. (9.30b)
r<r c
A contract for a material with a supplier becomes effectual, when material is pur-

chased for the first time under the terms of the contract, that is r<r c Vrmst = 1

and r<r c Vrms(t−1) = 0. Therefore,

CSmst = Vrmst t = 1, m ∈ RM, (9.31a)
r<r c

CSmst Vrmst − Vrms(t−1) 1 < t, m ∈ RM, (9.31b)
r<r c r<r c

CSmst Vrmst 1 < t, m ∈ RM, (9.31c)
r<r c

CSmst 1 − Vrms(t−1) 1 < t, m ∈ RM. (9.31d)
r<r c
In addition, there can be limits on contract durations. Once a contract with a sup-
plier begins, it must remain in effect for some minimum duration (unless it has
been signed toward the end of the horizon) and any material purchased from that
supplier must follow the contract terms. A maximum duration also exists for the
contract duration.

CSmst Vrmst m ∈ RM, t + CLL s > t > t, (9.32)
r<r c

CSmst + Vrmst CLU L
s − CLs + 1 m ∈ RM, t < H − CLU
s + 1,
t r<r c

t + CLU L
s t t + CLs . (9.33)
Note that contracts that are already in effect at the beginning of the horizon can be
enforced by initializing the variable CSmst appropriately.
9.6.5
Loans and Bonds
The MNC must pay back the principal and interest on bonds issued for raising
capital at the time of maturity. No interest or installment of the principal is paid
before maturity.
cbbt = BPb(t−BLb ) nbb(t−BLb ) (1 + RBbt )BLb t > BLb , (9.34)
where nbbt is the number of bonds b issued at time t, BPbt is a bond’s principal
value, and RBbt is the interest rate. We allow an upper limit on the capital that can
be raised through bonds. No bonds can be taken toward the end of the horizon
unless the length of the bond is small enough to be repaid before the end of the
horizon.
A loan is repaid in installments over the loan period and no lump sum is paid at
the end. All the installments account for the repayment of the principal along with
the interest. We allow the loan amount (lalt ) to be variable with time and between
some limits. With that, the total payment on loans that are active at time t is:

ialt = lalt RLlt t (t − LLl ) < t t. (9.35)
t
There may be an upper limit on the total amount of loans. Installments for a loan
that is taken during the end of the horizon may not account completely for the
loan. Hence, we assume that a sum that is equivalent to the unpaid loan is paid as
installment at the end of the horizon. To account for this, the parameter RLlt H is
given accordingly.
9.6.6
Costs and Incomes
The purchase of raw material in any period incurs two costs. One is fixed and
the other is variable with the amount of purchase. However, both depend on the
time and contract in which a purchase is made. Hence, even when no material
is purchased under the contract, the organization would need to pay a minimum
sum.

crmmst = KCmst Vrmst + dcrmst δcrmst + dsmst δsmst m ∈ RM. (9.36)
r<r c r<r c
The production cost for a facility is given by Eq. (9.37a). The production cost in-
volves a fixed cost that is proportional to the capacity of the facility and variable
costs that are proportional to the amount of material produced.

cprpt = γ cpt fcappt + mdmipt γ tpt
m i
pm ∈ FM, i ∈ (MF ∧ IIN). (9.37a)
For a technologically upgraded facility (UFft = 1), the production costs may be
different from that of the nonupgraded facility (UFft = 0). The difference in this
production cost is given in Eq. (9.37a–9.37e) and this can be either positive or neg-
ative.

TU
cprtupt γ cpt fcappt + mdmipt γ tptTU
m i
+ fcapU TU TU
p (γ cpt + γ tpt )(1 − UFpt )
pm ∈ FM, p ∈ UP, i ∈ (MFF ∧ IIN), t > TUPp . (9.37b)

TU
cprtupt γ cpt fcappt + mdmipt γ tptTU
m i
− fcapU TU
p (γ cpt
TU
+ γ tpt )(1 − UFpt )
pm ∈ FM, p ∈ UP, i ∈ (MFF ∧ IIN), t > TUPp . (9.37c)
If the facility does not undergo a technology upgrade (UFpt = 0), then the addi-
tional cost involved is zero.
TU
cprtupt (γ cpt + γ tptTU )fcapU
p UFpt
pm ∈ FM, p ∈ UP, i ∈ (MFF ∧ IIN), t > TUPp , (9.37d)
TU
cprtupt −(γ cpt + γ tptTU )fcapU
p UFpt
pm ∈ FM, p ∈ UP, i ∈ (MFF ∧ IIN), t > TUPp . (9.37e)
For transportation, we assumed that the costs are borne by the receiving inventory
facilities.

ctsft = mdmFft αmFft fm ∈ FM, Ff ∈ UPF, f ∈ I. (9.38)
m F
The costs for an inventory facility include those for holding and transferring the
material. The inventory holding costs include a fixed and a variable cost, similar to
production costs.

cstft = βcft fcapft + mfmf t βmmf t fm ∈ FM, f ∈ I. (9.39)
m
The income from sale of material at its true price (Eq. (9.40) and at a discounted
price (Eq. (9.41)) are given below.

csadt = φmdt smdt dm ∈ FM. (9.40)
m
The discounted price for the excess of the material produced is written in
Eq. (9.37a).

cwft = WDCmf t wmf t fm ∈ FM, f ∈ IO. (9.41)
m
The sales for a product cannot exceed its demand, which is different for upgraded
and nonupgraded products.
smdt Dmdt dm ∈ FM. (9.42)
Startups and shutdowns of a facility involve costs.

csupt σ supt (EZpt − EZp(t−1) ) t 2, (9.43a)
csdpt σ sdpt (EZp(t−1) − EZpt ) t 2, (9.43b)
cswpt = csupt + csdpt t 2. (9.43c)
The cost of depreciation depends on the rate of depreciation and asset value. The
rate of depreciation for an existing facility may be different from that of a new
facility or capacity addition. However, if there are several simultaneous additions,
the same rate applies.
cdft = fcapf (1) DPoft DRoft + (fcapft − fcapf (1) )DPnft DRnft . (9.44)
Just as the minimum capacity addition at a new location is different from that at
existing locations, the fixed costs for these two investments are also different.
ciN N
ft ft (EYft − EEYf (t−1) ) f ∈ PIN, t > CLTf , (9.45a)
ciO O
ft ft (EEYf (t−1) + EYft − 1) t > CLTf . (9.45b)
However, the variable costs are the same.
cift = TU
ft
TUft + ciN O
F t + cift + ft faft
F ∈ PIN, t > min[CLTf , TUPf ], t > CLTf , t > TUPf . (9.45c)

Equation (9.45c) also involves a fixed cost associated with technology upgrade. This
fixed cost may arise from acquiring new technology, investing in new equipment,
and so on. We assume that there is no capacity addition from a technology upgrade;
hence, there is no variable cost associated with technology upgrade.
The revenue from disinvesting a facility is proportional to the value of facility
per unit capacity and the disinvestment factor. Disinvestment factor is the ratio of
revenue from disinvestment and the value of the unsold facility. We use a disin-
vestment factor because the value of an unsold facility is usually different from the
value recoverable if the facility is sold.
rdeft fcapf (t−1) VEUf (t−1) RFf (t−1) t > TDf , (9.46)
rdeft fcapU
f VEUf (t−1) RFf (t−1) FXft t > TDf . (9.47)
The revenue from disinvestment is obtained at the time when production stops for
the first time and it is given by Eq. (9.47). The total revenue at time t is:

rdeft RDtU t > TDL , t > TDf . (9.48)
f
The total investment at time t is given by the following and is incurred at the time
when the new capacity becomes available for production/storage. In reality, the cost
of investment is spread over a period of time. However, these costs can be bundled
and can be represented as a single cost associated with the time when the facility
is available for use. The total amount of capital that can be raised is limited. The
capital for investments is raised through loans and bonds.

cift BPbt nbbt + lalt
f b l
min[CLTf , TUPf ] < t H − BLb , t > min[CLTf , TUPf ]. (9.49)

The total amount of capital that can be invested in a given interval is limited (Eq.
(9.50)). This is to limit the organization from handling high level of investments.

t
cift TCIU
t t (I vI + min[CLTL ,
t−I vI +1 f
TUPL ]), t > min[CLTf , TUPf ]. (9.50)

The tax that MNC must pay is computed based on the country that hosts the fa-
cilities. The tax is computed on income from sale of products less the cost of pur-
chases, production, storage, transportation, depreciation, shutdown/startup, and
depreciation.

ctxnt TXnt smdt φmdt + cwft − crmmst
d∈Dn m∈FM f ∈IOn s∈Sn m∈FM

− cstft − (cprpt + cprtup t )
f ∈I n p∈P n,p ∈UPn

− cswpt − ctrft − cdft . (9.51)
p∈P n f ∈I n f ∈P I n
The NPV of partial profit, not including the value of assets, loans, and bonds at
time t in a given nation is given by the following equation.

npnt = −ctxnt + smdt φmdt + cwft − crmmst
d∈Dn m∈FM f ∈IOn) s∈Sn m∈FM

− cstft − (cprpt + cprtup t )
f ∈I n p∈P n,p ∈UPn

− cswpt − ctrft + rdeft − cift . (9.52)
p∈P n f ∈I n f ∈P I n f ∈P I n
The NPV of the shareholder value includes the time value of money, values of assets
at the beginning and end of the horizon, and transactions due to loans and bonds.
npnt
shNPV =
n t
(1 + IRnt )
(Y T (t)−1)
fcapf H VEUf H
+ − fcapf,1 VEU f,1
(1 + IRnH )(Y T (H )−1)
f ∈f P I n
BPbt nbbt
+
(1 + IRnt )(Y T (t)−1)
b∈nB t (tH −BLb )
cbbt
−
(1 + irnt )(yr(t)−1)
b∈nB t (t>BLb )
lalt
+
(1 + IRnt )(Y T (t)−1)
l∈nL t

ialt
− . (9.53)
t
(1 + IRnt )(Y T (t)−1)
l∈nL
where fcapft VEUft is the value of an unsold facility and it is proportional to its ca-
pacity and unit value of the capacity. It should be noted that IRnt is the cumulative
interest rate in a nation. It can be computed from the actual interest rate in each
period.
This completes our complex and comprehensive model for asset management.
Equations (9.36), (9.37a), (9.38), (9.39), (9.40), (9.41), (9.43c), (9.44), (9.45c), and
(9.52) can be directly incorporated in Eq. (9.53) in the implementation of the model.
In the above presentation, these equations are written separately so as to improve
the readability. The model is clearly large with many variables and constraints and
is not easy to solve.
9.6.7
Remarks
A more important issue concerns the quality and speed of solution for such a com-
plex model. In a survey [12], it was found that the average hurdle rate or the min-
imum rate of return on investment (RORI) for project decisions is about 22.3%
with 62.5% of respondents specifying a value greater than 15%. This hurdle rate
is used as a measure against any risks that an investment decision may carry. In
this respect, researchers have presented what they called the “optimizer’s curse”
[61]. It says that an optimization analyst should, in general, be prepared for disap-
pointment; the management would take a skeptic’s view of the analysis because no
analysis is perfect. Since a strategic asset management model involves several thou-
sand parameters and nearly all of them are uncertain, a single optimal solution is
not always preferable. Even if we get an optimal solution from our model, it may
not be the scenario that is actually realized in future because of these uncertainties.
However, analyzing various scenarios would be of interest to the senior manage-
ment. In this respect, a solution obtained in reasonable computation time and one
that gives considerable profit would suffice. Thus, it is our belief that the integrality
gap for this complex strategic asset management problem is not as important as it
would be for a tactical planning/scheduling problem. Consequently, we allow one
day of computing time for each run in this work and present our results.
9.7
Example
We consider an MNC with four production facilities (three existing and one po-
tential, P1 to P4) in two nations with different tax rates and import–export restric-
tions, five raw materials (including one recycle) and one intermediate (1 to 6), six
9.7 Example 277
products (including the upgraded product, 7 to 12), two suppliers (S1 and S2) who
supply all the raw materials and intermediates, eight input inventory facilities (six
existing and two potential, I1 to I8), eight output inventory facilities (six existing
and two potential, O1 to O8), and five inventory facilities at distribution centers
(three existing and two potential, D1 to D5). We consider a planning horizon of 40
quarters or 10 years. Production facilities P2 (existing) and P4 (potential produc-
tion facility) manufacture the same products; hence, this example could illustrate
investment at P4 or relocation of P2 to P4. Among the products, the materials “8”
and “9” represent the same product before and after technology upgrade.
Table 9.1 gives the initial capacities, capacity expansion and construction limits,
upper limit on capacity, mass balance for production facilities, material storage for
inventory facilities, fixed and variable production costs, fixed and variable costs for
storage. It should be noted that while input/output inventory facilities can hold only
certain materials, the distribution centers can store all. Table 9.2 shows the raw ma-
terial availability and its purchase costs both within and outside a contract, demand
for products, its selling price and discounted sales price, and transportation costs
for all materials. The raw material availability and purchase costs are given accord-
ing to the quantity flexibility. The excess material is sold at a discounted price and it
can be sold at all the inventory facilities. Logically, all the excess material would be
sold at these inventory facilities and will not be shipped to the distribution facilities,
since the distribution involves additional costs. Hence, the selling of overproduced
material at the distribution centers does not occur. In this example, we did not write
these equations pertaining to sale of waste at the distribution centers in order to
reduce the number of variables. In addition, we have limited the sale of products to
the distribution centers. Transportation costs involve insurance and tariffs also. It
should be noted that transfers between input/output inventory facilities and their
associated production facility are not considered as transportation, and hence, they
do not incur any transportation costs. The material handling costs for this case are
included in the production costs.
Table 9.3 gives the cost of investing in a facility. While the fixed costs for con-
struction and capacity additions are different, the variable cost is same for both and
proportional to the capacity increase. A technology upgrade also carries a fixed cost.
The value of an unsold facility is also given. Table 9.2 gives the quantity flexibility
for material contracts. Each contract offers two prices with corresponding ranges.
There is an upper limit on the amount of material purchased without a contract.
Other parameters used in the model are listed below (Note: f = facility; p =
production facility; (f ) or (p) = 14 + i for input inventory, = 22 + p for production
facility, = 26 + o for output inventory, = 34 + d for distribution center.):
1. The inflation rates are 9% and 5%, and tax rates are 5% and
8%, for nations 1 and 2, respectively.
2. The maximum throughput (TFU ft = 1) for each inventory
facility equals its capacity, that for each distribution center
equals twice its capacity (TFU ft = 2), and that for a production
pt = 1).
facility equals its capacity (TFU
278
Table 9.1 Initial capacities (fcapf,1 ton/quarter), lower limits

(FAN O
f and FAf ton/quarter) on capacity additions, maximum
capacity fcapU f (ton/quarter) for all, material balance data for
production and inventory facilities, production (γ cft and
γ tft $/ton), and storage costs (βcft and βmmf t $/ton).
9 Chemical Supply Chain Redesign
Facility (nation) f cap f ,1 F AN

f F AO
f f cap U
f Material (inventory) or mass balance γ cf t or βcf t γ t f t or βmmf t
for production facility ($/ton) ($/ton)
(1)PF1 before upgrade 6000 – 2000 20,000 m1 + 0.5m2 = 0.7m7 + 0.8m8 2000 + 50t 1500 + 50t
(1)PF1 after upgrade 6000 – 2000 20,000 m1 + 0.5m2 = 0.7m7 + 0.8m9 2000 + 50t 500 + 50t
(1)PF2 6000 – 2000 20,000 m3 + 0.4m4 = 0.5m6 + 0.8m10 + 0.1m3 5000 + 100t 5000 + 200t
(2)PF3 6000 – 2000 20,000 m5 + 0.7m6 = 0.9m11 + 0.8m12 2000 + 50t 1500 + 50t
(2)PF4 0 5000 2000 20,000 m3 + 0.4m4 = 0.5m6 + 0.8m10 + 0.1m3 2000 + 50t 1500 + 50t
(1)I1–I4(1)O1–O4(2)I5–I8(2)O5–O8 6000 – 1000 20,000 1(I1); 2(I2); 3(I3, I7); 4(I4, I8); 5(I5); 200 + 10f + 10t 200 + 10f + 10t
6(I6, O3, O7); 7(O1); 8–9(O2); for 15–22 and for 15–22 and
10(O4, O8); 11(O5); 12(O6) 27–34 27–34
(1)D1–D2 (2)D3–D5 0 3000 1000 20,000 7–12 in D1–D5 200 + 15t for 200 + 15t for
35–37 35–37
250 + 10t for 250 + 10t for
38–39 38–39
9.7 Example 279
Table 9.2 Raw material availability limits (QCrmst or QSmst

ton/quarter), demands (Dmdt ton/quarter), transport costs
(αmf F t $/ton), products sales price (φmdt $/ton), products
discounted sales price (WDCmf t $/ton), and raw material
purchase price (δcrmdt or δsmdt $/ton).
M t (s) QC rmst , or QS mst , α mf F t φ mdt or W DC mot or δcrmst

or D mst or δs mst
1,3,5 [1, 34](1) r = 1: 0; 300 + 10t −10eF + 20ef r = 1: 1000 + 100t for t 15,
else 2500 + 50t;
r = 2: 2500; r = 2: 900 + 100t for t 15,
r = 3: 75,000 else 2400 + 50t;
No contract: 500 + 150t
1,3,5 [35,36](1) r = 1: 0; 300 + 10t −10F + 20f Same as above
r = 2: 1000;
r = 3: 3000
1,3,5 [37,40](1) r = 1 – 3: 0 300 + 10t −10F + 20f Same as above
1,3,5 [1,5](2); 50,000; 300 + 10t −10F + 20f 500 + 150t
[6,30](2); 1000;
[31,40](2) 50,000
2,4 [1,34](1,2); 50,000; 290 + 11t −10F + 20f 1500 + 100t
[35,36](1,2); 1000;
[37,40](1,2) 50,000
6 [1,34](1); 50,000; 250 + 15t −10F + 20f 11000 + 100t
[35,36](1); 1000;
[37,40](1) 50,000
6 [1,34](2); 50,000; 250 + 15t −10F + 20f 10000 + 100t
[35,36](2); 1000;
[37,40](2); 50,000

7 [1,H] 3 200 + 700t + 20 255 + 16t −10F + 20f φmdt : 700d + 450t −400m
t + 30
WDCmot : 200o + 100t −50m

8 [1,H] 100t +
3 250 − t + 25 20 265 + 17t −10F + 20f Same as above

9 [1,H] 3 200 + 300t + 20 275 + 18t −10F + 20f Same as above

t + 15
10 [1,H] 3(200 + 10t) 285 + 19t −10F + 20f Same as above

11 [1,20] 500t +
3 200 + 100 − 3t50 295 + 20t −10F + 20f Same as above


50 + t

12 [1,20] 3 200 + 350t + 100 295 + 20t −10F + 20f Same as above
25 + t

100 − t
f – Sending facility, F – receiving facility; (f ) or (F ) = 14 + I for input inventory, = 26 + O for

output inventory, = 34 + D for distribution center.
Table 9.3 Investment costs.
Set f t O
ft N
ft V EU f t
I or O 2000 −25f + 50t 100000 + 100t 150000 + 100t 100

P 5000 + 100t 200000 + 200t 250000 + 200t 200
UP 15,000,000 – 100,000t – – –
O
ft
fixed costs for expansion ($), N
ft
fixed costs for construction ($), ft variable costs ($/ton),
TU
ft
fixed technology upgrade costs ($).
Table 9.4 Model statistics and results.
SCR NR N
Continuous variables 18,134 15,418 18,134

Binary variables 2220 1320 2220
Constraints 26,390 20,915 26,390
Nonzeros 103,302 87,008 103,302
CPU time (h) 24.00 3 24.00
Relative gap (%) 9.43 2 3.42
NPV ($ billion) 10.20 6.08 9.22
3. The minimum throughout (TFL ft = 0) is zero for an

inventory facility (hence, no lower limits on the holdups in
inventory facilities), and that for a production facility is 60%
of its capacity (TFLpt = 0.6). A technology upgrade results in
a 10% decrease in throughput limits
(TFUU pt = 0.9TFpt & TFUpt = 0.9TFpt ).
U L L
4. The most number of times that a facility may expand is three

(NEf = 3).
5. The construction lead time for all production facilities is two
years (CLTp = 8 periods) and that for all inventory facilities
is one year or four periods. The time before which a facility
cannot be upgraded is one year (TUPf = 4 periods). The
minimum period during which no disinvestment is allowed
is one year (TDft = 4 periods).
6. The minimum and maximum lengths of a contract are 20
and 30 periods (CLL s = 20 and CLs = 30), respectively.
U
7. The revenue from the disinvestment of a facility does not

exceed its current value. In this example, it is only half the
current value (RFft = 0.5).
9.7 Example 281
8. σ supt = 1000000 − 500(p) + 1000t; σ sdpt =

1000000 − 500(p) + 1000t; RDPoft = 0.08 + 0.0005t; RDPnft =
0.06 + 0.0003t; DPoft = 100 + 10t; DPnft = 100 + 10t; RLlt t =
0.25 + (t − t )/LLl fort < H ; and
RLlt t = 1.15(t − LLl )/(H − LLl )fort = H .
9. The principal value of a bond b is $1 (BPbt = 1). Only single
types of loans and bonds are possible. Furthermore, the loan
exists only in nation 2 and the bonds in nation 1. Inflation is
based on the nations in which the loan and bonds are taken.
However, the capital raised in one nation can be used to fund
investments in other nations.
9.7.1
We chose this case study to illustrate the relocation of one production facility pro-
ducing the intermediate from one nation to another nation where the intermediate
is required and also the redesign of supply chain arising from the shutdown of in-
ventory capacities in some locations and investments at other locations. We have
chosen the parameters with a view to highlight all features of our formulation.
Also, some parameters are considered as constant over the entire planning hori-
zon, although the model does allow period-dependent values. This is done in order
to minimize the data and be able to present them completely in this chapter.
In order to consider risk, we analyze three scenarios.
1. Supply-chain redesign (SCR) scenario – we consider the
supply chain redesign plan for relocating a production
facility to a different country and investments in associated
inventory facilities and distribution centers.
2. No relocation (NR) scenario – we allow investments, but not
disinvestments.
3. Nationalization (N) scenario – we allow the possibility that
the newly relocated facility (P4) is nationalized or has to be
shutdown due to some regulatory factors related to
production and hence, these is loss of production. However,
the market can be catered to because the distribution center
is still operational. We assume that the maximum time until
which P4 can operate is quarter 28 and it would be
nationalized/disinvested from quarter 29 onward.
The SCR and N scenarios use the same model. The model for the NR scenario is
obtained by eliminating those constraints and variables related to disinvestments.
The models were implemented and solved using Cplex 9.0 in GAMS 21.7 on a
Windows XP-based HP workstation with Intel Xeon 3.6 GHz dual processor.
Figure 9.3 shows the capacity profiles of the inventory holding facilities at the
production facilities for the three scenarios. In the SCR scenario, I3 and I4 are dis-
invested as they are associated with the production facility P2. At the time when P4
is invested in, I7 and I8 are also invested as they are the associated input inventory
holding facilities at P4. Other facilities add capacity while I6 does not change its
capacity. In the NR scenario, all the facilities either expand or do not add capac-
ity. No facility is disinvested as the model does not allow disinvestments. In the N
scenario, the inventory facilities I3 and I4 are disinvested in during the beginning
of the horizon along with P2. I7 and I8 are invested during the beginning of the
horizon and they are nationalized toward the end of the horizon.
Figure 9.4 shows the capacity profiles of the production facilities. It can be seen
from the SCR scenario that the production facility P2 is disinvested/stops produc-
tion at one period after P4 starts its production. This kind of operation may reduce
the risk associated with the difficulties that the new production facility P4 may
face in the production process. If P4 faces any production problems, then P2 can
be used to meet the demand for the products. The other two facilities P1 and P3
add capacities twice during the planning horizon. In the NR scenario, all facilities
other than P2 add capacity. This is because the production costs at P2 are high and
whatever additional capacity is required is added at P4 instead at P2. The N sce-
nario shows that the production facility P2 is disinvested and P4 is invested early
on during the planning horizon. P4 is disinvested toward the later part of the hori-
zon and there is reduction in profit due to loss of sales for the products that P2/P4
produces. Even when there is a possibility of nationalization, the scenario shows
that P4 should be invested in so as to capture the demand for the products until
the time when nationalization would happen.
Figure 9.5 shows the capacity profiles for the output inventory facilities at the
production facilities. It can be seen that just as I3 and I4 are disinvested along with
P2, O3 and O4 are also disinvested and O7 and O8 are invested in along with I7, I8,
and P4 in the SCR scenario. The NR scenario shows that all facilities other than O1,
O3, and O4 expand and no facility is disinvested. In the N scenario, the facilities O7
and O8 are invested in the beginning and later nationalized along with P4. It can
be seen from Figs. 9.3 to 9.5 that the input inventory holding facilities are disin-
vested along with the production facilities; however, the output inventory holding
facilities are disinvested at a later stage. This is because the output inventory can
store products that can be sold at a later date.
Figure 9.6 shows the capacity profiles of distribution centers. It can be seen that
in all the cases, the capacity of the existing distribution centers does not change
and additions are made at all new locations. Even when the production facility is
nationalized in the N scenario, the distribution centers are not nationalized.
Figure 9.7 shows the capacity profile of production facility P1 with respect to the
capacity required for meeting the demand for the products. It also shows the uti-
lization of the facility and the time at which technology upgrade needs to be done.
Production facility P1 produces products 7, 8, and 9, where product 8 on technology
upgrade becomes product 9. The curves 7, 8, and 9 are the capacity of production
facility required to meet 100% demand for products 7, 8, and 9, respectively. The
plot P1 is the capacity of production facility that is planned. For the SCR scenario,
it can be seen that there are two capacity additions. The early addition is small be-
9.7 Example 283
I1 I2 I3 I4 I5 I6 I7 I8
12
a) 10
capacity (kton/quarter)
0
1 5 9 13 17 21 25 29 33 37
tim e (quarters)
10
b) 9
7
0
1 5 9 13 17 21 25 29 33 37
time (quarters)
c) 18
16
14
12
10
0
1 5 9 13 17 21 25 29 33 37
time (quarters)
Fig. 9.3 (a) SCR scenario: Capacities of input “I6” remain unchanged, while others expand.
inventory facilities. Inventories “I3” and “I4” (c) N scenario: Capacities of input inventory
are disinvested at quarter 10, while the others facilities. Inventories “I7” and “I8” are invested
expand. “I7” and “I8” are new facilities added at quarter 5 and 9, respectively and they are
at quarters 9 and 7, respectively. Capacity of nationalized at quarter 29. Inventories “I3” and
inventory “I6” remains unchanged. (b) NR “I4” are disinvested at quarters 11 and I8,
scenario: Capacities of input inventory respectively.
facilities. Capacity of inventories “I3,” “I4,” and
P1 P2 P3 P4 P2+P4
a) 16
14
12
10
0
1 5 9 13 17 21 25 29 33 37
time (quarters)
b) 18
16
14
12
10
0
1 5 9 13 17 21 25 29 33 37
tim e (quarters)
c) 18
16
14
capacity (kton/qua rter)
12
10
0
1 5 9 13 17 21 25 29 33 37
time (quarters)
Fig. 9.4 (a) SCR scenario: Capacity profiles of production

facilities. Facility “P2” is disinvested at quarter 10 and relocated
to “P4” at quarter 9. Other facilities expand. (b) NR scenario:
Capacity profiles of production facilities. Capacity of facility
“P2” remains unchanged, but all other facilities expand. (c) N
scenario: Capacity profiles of production facilities. Facility “P4”
is invested at quarter 9 and is nationalized at quarter 29.
9.7 Example 285
O1 O2 O3 O4 O5 O6 O7 O8
a) 9
7
0
1 5 9 13 17 21 25 29 33 37
tim e (quarters)
b) 9
7
capacity (kton/qua rter)
0
1 5 9 13 17 21 25 29 33 37
tim e (quarters)
c) 9
7
0
1 5 9 13 17 21 25 29 33 37
tim e (quarters)
Fig. 9.5 (a) SCR scenario: Capacities of output expand, while others expand. (c) N scenario:
inventory facilities. Inventories “O3” and “O4” Capacities of output inventory facilities.
are disinvested at quarter 10, while some of the Inventory “O1” does not expand. Inventories
others expand. “O7” and “O8” are invested at “O7” and “O8” are invested at quarter 9 and
quarters 6 and 9, respectively. (b) NR scenario: are nationalized at quarters 31 and 29,
Capacities of output inventory facilities. respectively.
Inventories “O1,” “O3,” and “O4” do not
D1 D2 D3 D4 D5
a) 7
5
4
0
1 5 9 13 17 21 25 29 33 37
tim e (quarters)
b) 7
5
0
1 5 9 13 17 21 25 29 33 37
tim e ( quarters)
c) 7
5
0
1 5 9 13 17 21 25 29 33 37
tim e (quarters)
Fig. 9.6 (a) SCR scenario: Holding capacities at distribution

centers. Inventories “D1,” “D2,” and “D3” do not expand.
(b) NR scenario: Holding capacities at distribution centers.
Inventories “D1,” “D2,” and “D3” do not expand. (c) N
scenario: Holding capacities at distribution centers. Inventories
“D1,” “D2,” and “D3” do not expand.
9.7 Example 287
P1 7 8 9 P1-UT
a) 20 4
18
16
3
14
12
10 2
8
6
1
4
UT
2 Technology Upgrade
0 0
1 5 9 13 17 21 25 29 37 1
time (quarters)
b) 20 4
18
16
3
14
12
10 2
6
1
4
2 Technology Upgrade
UT
0 0
1 5 9 13 17 21 25 29 33 37 41
tim e (quarters)
c) 20 4
18
16
3
14
12
10 2
8
6
1
4
Technology Upgrade
UT
0 0
1 5 9 13 17 21 25 29 33 37 41
tim e (quarters)
Fig. 9.7 (a) SCR scenario: Capacity and the materials in inventories) of products “7,”
utilization (UT) of production facility “P1.” “8,” and “9,” respectively. Plot P1-UT shows
Plots 7, 8, and 9 show the desired capacities of the capacity utilization for “P1.” Technology
“P1” to meet 100% demands (not considering upgrade occurs at quarter 32. (c) N scenario:
the materials in inventories) of products “7,” Capacity and UT of production facility “P1.”
“8,” and “9,” respectively. Plot P1-UT shows Plots 7, 8, and 9 show the desired capacities of
the capacity utilization for “P1.” Technology “P1” to meet 100% demands (not considering
upgrade occurs at quarter 35. (b) NR scenario: the materials in inventories) of products “7,”
Capacity and UT of production facility “P1.” “8,” and “9,” respectively. Plot P1-UT shows
Plots 7, 8, and 9 show the desired capacities of the capacity utilization for “P1.” Technology
“P1” to meet 100% demands (not considering upgrade occurs at quarter 26.
P2+P4 6-11 6-12 10 P2-UT P4-UT
a) 20 5
18
16 4
14
12 3
10
8 2
4 1
UT
2
0 0
1 5 9 13 17 21 25 29 33 37 41
time (quarters)
b) 20 5
18
16 4
14
12 3
10
8 2
4 1
UT
2
0 0
1 5 9 13 17 21 25 29 33 37 41
time ( quarters)
c) 20 5
18
16 4
14
12 3
10
8 2
6
4 1
UT
2
0 0
1 5 9 13 17 21 25 29 33 37 41
time (quarters)
Fig. 9.8 (legend see p. 289)

9.7 Example 289
cause the demand for the product 8 is decreasing and the increased demand for
product 7 does not create enough profit so as to offset the loss due to selling the
excess of product 8 as waste. The second addition is delayed and the amount of ad-
dition is considerable and this is because the addition is stalled until the time when
the demand for the product 9 is sufficiently high. It can be seen that the profiles for
the NR and the N scenarios are also similar, but the time at which the technology
upgrade occurs is different.
Figure 9.8 shows the capacity profile of production facility P2 and P4 combined
and its utilization. P2/P4 produces product 10 and intermediate 6. The intermedi-
ate 6 is sent to production facility P3, and the products 11 and 12 are made by P3.
In the figures, 6–11 is the capacity of P2/P4 required to meet 100% demand for
intermediate 6 at production facility P3 so that it can meet 100% demand for the
product 11 that it produces. 6–12 is the capacity of P2/P4 required to meet 100%
demand for intermediate 6 at production facility P3 so that it can meet 100% de-
mand for the product 12 that it produces. The SCR scenario has an excess capacity
to produce intermediate 6 and product 10 briefly at quarter 9. However, the pro-
duction capacity is utilized completely and the products are stored in inventory for
later use. In the SCR and NR scenario, there are two capacity additions. However,
in the N scenario, the facility P4 is nationalized toward the end of the horizon and
hence, there is no scope for the second addition. The total capacity of P2 and P4 is
less in NR and N scenarios when compared to the SCR scenario because either the
production costs at P2 are high or P4 would be nationalized in the future.
Fig. 9.8 (a) SCR scenario: Capacity and and “P4” (“P2 + P4”) combined that is
utilization (UT) of production facilities “P2” required for satisfying the 100% demand for
and “P4” (“P2 + P4”) combined. Plot 10 shows intermediate “6” from downstream production
the desired capacity of production facilities facility “P3,” if “P3” has to satisfy the 100%
“P2” and “P4” (“P2+P4”) combined that is (not considering inventory) demand of product
required for satisfying the 100% (not “11” and “’12,” respectively. Plots P2-UT and
considering inventory) of demand for the P4-UT show the capacity utilizations for “P2”
product “10.” Plots 6–11 and 6–12 show the and “P4,” respectively. (c) N scenario: Capacity
desired capacities of production facility “P2” and UT of production facilities “P2” and “PF4”
and “P4” (“P2 + P4”) combined that is (“P2 + P4”) combined. Plot 10 shows the
required for satisfying the 100% demand for desired capacity of production facility “P2” and
intermediate “6” from downstream production “P4” (“P2 + P4”) combined that is required for
facility “P3,” if “P3” has to satisfy the 100% satisfying the 100% (not considering
(not considering inventory) demand of product inventory) of demand for the product “10.”
“11” and “’12,” respectively. Plots P2-UT and Plots 6–11 and 6–12 show the desired
P4-UT show the capacity utilizations for “P2” capacities of production facility “P2” and “P4”
and “P4,” respectively. (b) NR scenario: (“P2 + P4”) combined that is required for
Capacity and UT of production facilities “P2” satisfying the 100% demand for intermediate
and “P4” (“P2 + P4”) combined. Plot 10 shows “6” from downstream production facility “P3,”
the desired capacity of production facility “P2” if “P3” has to satisfy the 100% (not
and “P4” (“P2 + P4”) combined that is considering inventory) demand of product
required for satisfying the 100% (not “11” and “’12,” respectively. Plots P2-UT and
considering inventory) of demand for the P4-UT show the capacity utilizations for “P2”
product “11.” Plots 6–11 and 6–12 show the and “P4,” respectively.
desired capacities of production facility “P2”
P3 11 12 P3-UT
a) 18 4
16
14
3
12
10
2
8
6
1
4
UT
2
0 0
1 5 9 13 17 21 25 29 33 37 41
tim e (quarters)
b) 18 4
16
14
3
12
10
2
8
1
4
UT
2
0 0
1 5 9 13 17 21 25 29 33 37 41
time (quarters)
c) 18 4
16
14
3
12
10
2
8
1
4
UT
0 0
1 5 9 13 17 21 25 29 33 37 41
tim e (quarters)
Fig. 9.9 (a) SCR scenario: Capacity and “P3” required for satisfying the 100% (not
utilization (UT) of production facility “P3.” considering inventory) of demand for the
Plots 11 and 12 are the desired capacities of products “11” and “12,” respectively. (c) N
production facility “P3” required for satisfying scenario: Capacity and UT of production facility
the 100% (not considering inventory) of “P3.” Plots 11 and 12 are the desired capacities
demand for the products “11” and “12,” of production facility “P3” required for
respectively. (b) NR scenario: Capacity and UT satisfying the 100% (not considering
of production facility “P3.” Plots 11 and 12 are inventory) of demand for the products “11”
the desired capacities of production facility and “12,” respectively.
9.7 Example 291
Figure 9.9 shows the production capacity for the facility P3 for the three cases. It
can be seen that there are two additions in each of the scenarios. When the down-
stream facility P4 is nationalized and P2 had been disinvested, the intermediate
6 is purchased from the open market and the production facility adds capacity to
meet the demand for its products.
We allow contracts only for raw materials “1,” “3,” and “5” from supplier “1.” In
the SCR scenario, the contracts for all the three materials were signed at quarter 6
and the contracts lasted until quarters 53, 34, and 34 for materials “1,” “2,” and “3,”
respectively. In the NR scenario, the contracts for materials “1,” “3,” and “5” are
signed at quarters “6,” “7,” and “6” and they last until quarters “35,” “36,” and “34,”
respectively. In the N scenario, the contracts are signed at quarter 6 for all three
raw materials and they last until quarters “34,” “28,” and “35” for raw materials
“1,” “3,” and “5,” respectively. In all the three models, loans/bonds were taken at a
time when there was a capacity addition.
The model statistics and results are reproduced in Table 9.4. The SCR scenario
gives a profit of $10.2 billion while the NR scenario gives a profit of $9.46 billion. It
can clearly be seen that our model that includes facility disinvestment gave a 7.82%
higher profit for this example. Having efficient supply chain redesign models can
thus help the organizations involved in making optimal business decisions and in-
crease their profit. The N scenario gave a profit of $9.22 billion, which is much less
than both the scenarios. If the organization has greater confidence in the govern-
ment at the new location, then it may wish to relocate and have the possibility of
achieving a 10.63% more profit on top of the $9.22 billion, else it may choose not
to invest in the new location. However, since the model was solved with nation-
alization in mind, the capacity profiles for the N scenario are different from that
for the SCR scenario. Hence, if the plan of relocation has to be implemented, then
the SCR plan has to be implemented and not the N plan as the N plan only gives
the possibility of what may happen. If the organization chooses to implement the
SCR plan and the nationalization happens, then the profit would be less than $9.22
billion.
9.7.2
Future Work
As has been described in Section 9.6.7, the management would require a solution
in reasonable computational time and hence, the relative gap is not very impor-
tant. At the same time, the management would require a set of solutions that gives
them an opportunity to analyze the effect of some of the decisions not being im-
plemented as planned. Hence, a population of solutions is required instead of a
single solution obtained by using GAMS. In addition, this reasonable sized prob-
lem with only 4 production facilities and 21 inventory holding facilities has tens
of thousands of constraints and hundreds of thousands of variables. A real-world
problem would be much larger and the tree size in the branch and bound algo-
rithm of Cplex 9.0 may become larger than the current computational resources
available. Hence, there is a need for alternative solution strategies that are popu-
lation based and are not limited by the computational resources at hand. To this
end, we are working on evolutionary algorithms, which are population based and
are not limited by memory as most of these algorithms do not store any history of
the progress of the solution.
9.8
Conclusion
In this chapter, we have presented a novel MILP model for making efficient capac-
ity management and supply chain redesign decisions for an MNC. To the best of
our knowledge, such a supply chain redesign model for chemical supply chains is
absent in the literature. Using our model, we have also demonstrated that we can
obtain the best strategy for investment, involving a variety of real decisions, such
as facility relocation, disinvestment, technology upgrade, raising capital through
loans and bonds, and hedging risk through signing contracts for material supply.
It is also possible to use this model repeatedly to analyze various probabilistic sce-
narios and thus make the choice of implementing one of the various resulting
plans. In a realistic case study in this chapter, it was observed that plans can dif-
fer by as much as 7% in shareholder value, which is a considerable amount as the
profits are usually in billions of dollars. We have only analyzed the risk at the high-
est level because tactical issues would only change the profit by a small percentage
compared to the worst-case high-level scenario.
From the computational results, it can be seen that even a relatively small exam-
ple takes about a day to obtain a reasonable plan. The problem sizes in the industry
are large and may require an excessive amount of computation time. Hence, it
would be an attractive area for future research to seek alternative methods of opti-
mization.
Acknowledgments
We would like to acknowledge the financial support from the Institute of Chemical
& Engineering Sciences under the project 06-432001.
References
1 Dixit, A., Pindyck, R. S., The op- 3 Grossmann, I. E., Challenges in the
tions approach to capital investment. new millennium: Product discovery
Harv. Bus. Rev. May–Jun (1995), pp. and design, enterprise and supply
105–115. chain optimization, global life cycle
2 Grossmann, I. E., Westerberg, A. W., assessment. Comput. Chem. Eng. 29
Research challenges in process sys- (2004), pp. 29–39.
tems engineering. AIChE J. 46 (2000), 4 Julka, N., Srinivasan, R., Karimi,
pp. 1700–1703. I. A., Agent-based supply chain man-
References 293
agement – 1: Framework. Comput. location factors: A cross national

Chem. Eng. 26 (2002), pp. 1755–1769. comparison between Singapore and
5 Julka, N., Karimi, I. A., Srinivasan, Malaysia. J. Bus. Logist. 24(1) (2003),
R., Agent-based supply chain man- pp. 147–170.
agement – 2: A refinery application. 17 Partovi, F., An analytic model for lo-
Comput. Chem. Eng. 26 (2002), pp. cating facilities strategically. Omega
1771–1781. 34(1) (2006), pp. 41–55.
6 Lasschuit, W., Thijssen, N., Sup- 18 Olhager, J., Rudberg, M., Wikner,
porting supply chain planning and J., Long-term capacity management:
scheduling decisions in the oil and Linking the perspectives from manu-
chemical industry. Comput. Chem. facturing strategy and sales and oper-
Eng. 28 (2004), pp. 863–870. ations planning. Int. J. Prod. Econ. 69
7 Holland, C. P., Shaw, D. R., Kawalek, (2001), pp. 215–225.
P., BP’s multi-enterprise asset man- 19 McDonald, C. M., Reklaitis, G. V.,
agement system. Inf. Softw. Technol. 47 Design and operation of supply chains
(2005), pp. 999–1007. in support of business and financial
8 Timpe, C. H., Kallrath, J., Optimal strategies. Foundations of Computer-
planning in large multi-site produc- Aided Process Design, FOCAPD.
tion networks. Eur. J. Oper. Res. 126 Princeton, NJ, USA, July 11–16, 2004.
(2000), pp. 422–435. 20 Amiri, A., Designing a distribution
9 Shapiro, J. F., Modeling the Supply network in a supply chain system: For-
Chain., Duxbury, Thomson Learning, mulation and efficient solution proce-
California, USA, 2001. dure. Eur. J. Oper. Res. 171 (2006), pp.
10 Stadtler, H., Supply chain manage- 567–576.
ment and advanced planning – basics, 21 Bjorkqvist, J., Roslof, J., Strategic
overview and challenges. Eur. J. Oper. investment planning in the pulp and
Res. 163 (2005), pp. 575–588. paper industry using mixed integer
11 Hill, T., Operations Management: linear programming. AIChE Annual
Strategic Context and Managerial Meeting, Cincinnati, OH, USA, Octo-
Analysis, Palgrave, New York, 2000. ber 30–November 04, 2005.
12 Slagmulder, R., Bruggeman, W., 22 Bok, J.-K., Lee, H., Park, S., Robust
van Wassenhove, L., An empirical investment model for log-range capac-
study of capital budgeting practices ity expansion of chemical processing
for strategic investments in CIM tech- networks under uncertain demand
nologies. Int. J. Prod. Econ. 40 (1995), forecast scenarios. Comput. Chem.
pp. 121–152. Eng. 22(7/8) (1998), pp. 1037–1049.
13 Akalu, M. M., The process of invest- 23 Bhutta, K. S., Huq, F., Frazier, G.,
ment appraisal: The experience of 10 Mohamed, Z., An integrated location,
large British and Dutch companies. production, distribution and invest-
Int. J. Proj. Manag. 21 (2003), pp. 355– ment model for a multinational corpo-
362. ration. Int. J. Prod. Econ. 86 (2003), pp.
14 Peel, M. J., Bridge, J., How planning 201–216.
and capital budgeting improve SME 24 Gatica, G., Papageorgiou, L. G.,
performance. Long Range Plan. 31(6) Shah, N., Capacity planning under
(2005), pp. 848–856. uncertainty for the pharmaceutical
15 Pirttila, T., Sandstrom, J., Manufac- industry. Trans. Inst. Chem. Eng. 81
turing strategy and capital budgeting (2003), pp. 665–678.
process. Int. J. Prod. Econ. 41 (1999), 25 Levis, A. A., Papageorgiou, L. G., A
pp. 335–341. hierarchical solution approach for
16 Bhatnagar, R., Jayaram, J., Phua, multi-site capacity planning under
Y. C., Relative importance of plant
uncertainty in the pharmaceutical in- 36 Cohen, M. A., Lee, H. L., Resource

dustry. Comput. Chem. Eng. 28 (2004), deployment analysis of global manu-
pp. 707–725. facturing and distribution networks.
26 Papageorgiou, L. G., Rotstein, G. E., J. Manuf. Oper. Manag. 2 (1989), pp.
Shah, N., Strategic supply chain op- 81–104.
timization for the pharmaceutical 37 Pyke, D. F., Cohen, M. A., Multiprod-
industries. Ind. Eng. Chem. Res. 40 uct integrated production-distribution
(2001), pp. 275–286. systems. Eur. J. Oper. Res. 74 (1994),
27 von Lanzenauer, C. H., Eschen, E., pp. 18–49.
Pilz-Glombik, K., Capacity planning 38 Tsiakis, P., Shah, N., Pantelides,
in a transitional setting with simula- C. C., Design of mulit-echelon supply
tion based modeling: A case study. Int. chain networks under demand uncer-
Trans. Oper. Res. 9 (2002), pp. 125–139. tainty. Ind. Eng. Chem. Res. 40 (2001),
pp. 3585–3604.
28 Lee, H.-K., Lee, I.-B., Reklaitis, G. V.,
Capacity expansion problem of multi- 39 Gjerdrum, J., Shah, N., Papageor-
site batch plants with production and giou, L. G., Transfer prices for multi-
distribution. Comput. Chem. Eng. 24 enterprise supply chain optimization.
(2000), pp. 1597–1602. Ind. Eng. Chem. Res. 40 (2001), pp.
1650–1660.
29 Rajagopalan, S., Swaminathan, J. M.,
A coordinated production planning 40 Goetschalckx, M., Vidal, C. J., Do-
model with capacity expansion and gan, K., Modeling and design of
inventory management. Manag. Sci. global logistics systems: A review
47(11) (2001), pp. 1562–1580. of integrated strategic and tactical
models and design algorithms. Eur. J.
30 Oh, H.-C., Karimi, I. A., Regulatory
Oper. Res. 143 (2002), pp. 1–18.
factors and capacity-expansion plan-
ning in global chemical supply chains.
Reklaitis, G. V., A simulation-
Ind. Eng. Chem. Res. 43 (2004), pp.
optimization framework for address-
3364–3380.
ing combinatorial and stochastic as-
31 Melo, M. T., Nickel, S., Saldanha da pects of an R&D pipeline manage-
Gama, F., Dynamic multi-commodity ment problem. Comput. Chem. Eng. 24
capacitated facility location: A math- (2000), pp. 1005–1011.
ematical modeling framework for
strategic supply chain planning. Com-
Reklaitis, G. V., A simulation-
put. Oper. Res. 33 (2006), pp. 181–208.
optimization framework for research
32 Gue, K. R., A dynamic distribution and development pipeline manage-
model for combat logistics. Comput. ment. AIChE J. 47 (2001), pp. 2226–
Oper. Res. 30 (2003), pp. 367–381. 2242.
33 Arntzen, B. C., Brown, G. G., Har- 43 Subramanian, D., Pekny, J. F.,
rison, T. P., Trafton, L. L., Global Reklaitis, G. V., Blau, G.,
supply chain management at digi- Simulation-optimization framework
tal equipment corporation. Interfaces for stochastic optimization of R&D
25(1) (1995), pp. 69–93. pipeline management. AIChE J. 49
34 Oh, H.-C., Karimi, I. A., Global multi (2003), pp. 96–112.
product production-distribution plan- 44 Perea-Lopez, E., Ydstie, B. E., Gross-
ning with duty drawbacks. AIChE J. mann, I. E., A model predictive con-
52(2) (2006), pp. 595–610. trol strategy for supply chain opti-
35 Bradley, J., Arntzen, B. C., The si- mization. Comput. Chem. Eng. 27
multaneous planning of production, (2003), pp. 1201–1218.
capacity, and inventory in seasonal de- 45 Guillen, G., Mele, F. D., Bagajewicz,
mand environments. Oper. Res. 47(6) M. J., Espuna, A., Puigjaner, L., Mul-
(1999), pp. 795–806. tiobjective supply chain design under
References 295
uncertainty. Chem. Eng. Sci. 60 (2005), 54 Tsay, A. A., Nahmias, S., Agrawal, N.,
pp. 1535–1553. in: Quantitative models for supply chain
46 Ryu, J.-H., Pistikopoulos, E. N., De- management, Tayur, S., Ganeshan,
sign and operation of an enterprise- R., Magazine, M. (eds.), Kluwer Acad-
wide process network using operation emic Publishers, Massachusetts, USA,
policies. 1. Design. Ind. Eng. Chem. 1999.
Res. 44 (2005), pp. 2174–2182. 55 Fan, J. P. H., Price uncertainty and
47 McDonald, C. M., Karimi, I. A., Plan- vertical integration: An examination
ning and scheduling of parallelmi- of petrochemical firms. J. Corp. Fi-
continuous processes. 1. Production nance 6 (2000), pp. 345–376.
planning. Ind. Eng. Chem. Res. 36 56 Miller, K. D., Waller, H. G., Sce-
(1997), pp. 2691–2700. narios, Real Options and Integrated
48 Wilkinson, S. J., Cortier, A., Shah, Risk Management. Long Range Plan
N., Pantelides, C. C., Integrated pro- 36 (2003), pp. 93–107.
duction and distribution schedul-
57 Alessandri, T. M., Ford, D. N., Lan-
ing on a Europe-wide basis. Comput.
der, D. M., Leggio, K. B., Taylor,
Chem. Eng. 20 (1996), pp. S1275–
M., Managing risk and uncertainty in
S1280.
complex capital projects. The Q. Rev.
49 Sharma, M., Ammons, J. C., Hart- Econ. Finance 44 (2004), pp. 751–767.
man, J. C., Asset management with
reverse product flows and environ- 58 Karimi, I. A., Srinivasan, R., Por,
mental considerations. Comput. Oper. L. H., Unlocking supply chain im-
Res. 34 (2007), pp. 464–486. provements through effective logis-
tics. Chem. Eng. Prog. 98 (2002), p.
50 Romero, J., Badell, M., Bagajewicz,
32.
M. J., Puigjaner, L., Integrating bud-
geting models into scheduling and 59 Naraharisetti, P. K., Karimi, I. A.,
planning models for the chemical Srinivasan, R., Location allocation
batch industry. Ind. Eng. Chem. Res. 42 production distribution in the chem-
(2003), pp. 6125–6134. ical process industries. INFORMS
51 Barbaro, A., Bagajewicz, M. J., Use Annual Meeting, 12–16 Nov, San Fran-
of inventory and options contracts to cisco, USA, 2005.
hedge financial risk in planning un- 60 Naraharisetti, P. K., Karimi, I. A.,
der uncertainty. AIChE J. 50(5) (2004), Srinivasan, R., Capacity management
pp. 990–998. in the chemical supply chain. Interna-
52 Sykuta, M. E., Futures trading and tional Symposium on Advanced Control
supply contracting in the oil refining of Chemical Processes-ADCHEM, Gra-
industry. J. Corp. Finance 2 (1996), pp. mado, Brazil, 02–05 April, 2006.
317–334. 61 Smith, J. E., Winkler, R. L., The op-
53 Annupindi, R., Bassok, Y., in: Quan- timizer’s curse: Skepticism and post-
titative models for supply chain man- decision surprise in decision analysis.
agement, Tayur, S., Ganeshan, R., Manag. Sci. 52(3) (2006), pp. 311–322.
Magazine, M. (eds.), Kluwer Acad-
emic Publishers, Massachusetts, USA,
1999.
Appendix
Nomenclature
Note: Superscript L (U ) denotes the lower (upper) limit or the lowest (highest)
value in a parameter vector, and superscript TU denotes technology upgrade.
Indices
b number of bonds
d distribution center
f facility (production facility or inventory holding facility)
i input inventory holding facility
l number of loans
m raw materials and products
p production facility
o output inventory holding facility
t period
Sets
D set of distribution centers
FM set of materials that a given facility may hold
I set of inventory facilities
IIN set of input inventory facilities in a production facility
IINRset of input inventory facilities in a production facility, which hold a recy-
cled material
IO set of input and output inventory holding facilities at production facilities
M set of materials
MF set of facilities that a given material may occupy
MFF set of arcs that a given material may follow
OIN set of output inventory facilities in a production facility
P set of production facilities
PI set of facilities – production and inventory
PIN set of facilities that do not exist at the beginning of the horizon
RCM set of recycled materials
RM set of raw materials, including intermediates
S set of raw material suppliers
UM set of materials that may undergo a technology upgrade
UMM set of material that a given material is upgraded into
UP set of production facilities that may undergo a technology upgrade
UPF set of facilities that may send materials to a given facility
Appendix 297
Parameters
BLb duration after which a bond matures
BPbt principal value of a bond b issued at t
CLs interval for which a contract may be signed
CLTf construction lead time for a facility
Dmdt demand for a material at an entity at t
DPft depreciation price or value of a plant on which depreciation is calculated
DRft rate of depreciation on a facility
FAN f minimum capacity for a new facility at a location at t
FAO f minimum capacity expansion allowed for an existing facility at t
IvI investment interval during which there is a limit on investment amount
IRt interest rate for computing time value of money
KCmst fixed cost for purchasing a raw material in the range r and (r + 1) from
a supplier at t when a contract is signed
LLl duration in which a loan of type l has to be paid back
NEf maximum number of times that a facility may expand
PTp interval in which routine maintenance must be done at least once for a
production facility
QCrmst maximum amount of material that may be purchased from a supplier
under a contract
QSmst maximum amount of material that can be purchased from a supplier
without contract
RBbt rate of interest paid for a bond
RDt total revenue from disinvestments at t
RFft revenue factor for an entity for disinvestment
RLlt t the rate of installment paid at t for a loan l taken at t
STp maximum number of periods for which a facility may operate during
PTp (STp < PTp )
TCIt cost for investments in all facilities in a given interval ending at t
TDf time before which a facility cannot be disinvested
TFft throughput factor for a facility at t
TFUft throughput factor at t for an entity that has undergone a technology
upgrade
TUPf initial interval during which no technology upgrade is allowed for a fa-
cility
TXnt tax rate in a nation at t
VEUft value of a facility per unit capacity
WDCmot unit cost for disposing waste at an entity at t
Y Tt year corresponding to t
αmfFt unit cost of transporting a material from one entity to another at t
βcft fixed unit cost of operating a storage facility at t
βmft variable unit cost of operating a storage facility at t
γ cft fixed unit cost of operating a production facility at t
γ tft variable unit cost of operating a production facility at t
δcre et unit cost of material purchased in range r and (r + 1) at time t under a

contract
δsmst unit cost of material purchased at t without a contract
σ suft cost of starting production at a facility at t
σ sdft cost of stopping production at a facility at t
φmdt unit price of a material sold at a distribution center at t
ft variable unit cost for expanding a facility
Nf fixed cost for adding capacity at a new location
Of fixed cost for adding capacity at an existing location
mf stoichiometry coefficient of a material that is processed in a production
facility
Binary Variables
CRrmst 1, if material is purchased through a contract in the range r and (r + 1)
from an entity at t
FXft 1, if a facility is sold and the capacity is no longer available at t
FYft 1, if an added capacity is available for production at t
FZpt 1, if a production facility is in operation at t
TUft 1, if a facility undergoes a technology upgrade at t
Continuous 0–1 Variables

CSmst 1, if a contract is signed for purchasing a material from a supplier at t
FEYft 1, if a facility exists at t
FSYft 1, if a facility that existed before t is sold-off at t or before
UFft 1, if a facility has undergone a technology upgrade
Continuous Variables
cbbt cost of paying back a bond b at time t
cdft cost of depreciation at t for an entity
cift cost for investing in a facility at t
ciN
ft fixed cost for investing in a new facility at t
O
cift fixed cost for investing in an existing facility at t
cprpt cost of production at a facility at t
crmmst cost of purchasing a raw material from a supplier at t
cstft cost of storing at a facility at t
cswpt cost of switching production at a facility at t
csadt income from sales for at a distribution center at t
ctsft cost of transportation for a facility at t
cwot cost for disposing waste at a facility at t
dcrmst amount of material purchased in the range r and (r + 1) from a supplier
at t under a contract
dsmst amount of material purchased from a supplier at t without a contract
iait installment for a loan at t
shNPV net present value of the shareholder value for the entire horizon
npt net profit at t
rdeft revenue from disinvesting an entity at t
Appendix 299
Positive Variables
cprtupt change in cost of production after technology upgrade at a facility at t
csupt cost of startup at a facility at t
csdpt cost of shutdown at a facility at t
ctxnt amount of taxes paid in nation n at t
faft amount of capacity added by a facility at t
fcapft capacity of an entity at t
mdmfFt amount of a material m sent from f to F at t
mfmft amount of a material present at a location at t
ibbt amount of capital raised through bonds b at t
lalt amount of capital raised through loans l at t
nbbt number of bonds issued at t
smdt amount of material sold at an entity at t
wmot amount of material disposed as waste at an entity at t
301
10
A Hybrid Tabu-Branch and Bound Approach for the Solution of
Large-Scale Supply Chain Management Models
João M. Natali, Peter Chen, Jose M. Pinto
The growing complexity of modern supply chain networks usually requires exces-
sive computational effort for their integrated management, with no guarantee of
optimality – or even feasibility – due to technological limitations. This issue is often
solved with methods that break down the problem into solvable sizes. Among the
solution strategies are decomposition techniques and metaheuristic approaches.
The former, despite some recent success with Lagrangean decomposition, rely too
heavily on the specific structure of the model, whereas the latter generally lack the
necessary structure to solve large and complex problems. To improve the tractabil-
ity of large-scale problems that arise in industry, the development of methodolo-
gies that harness the advantages of multiple solution approaches are needed. In
this work, we propose an algorithm based on a hybrid implementation of Branch
& Bound approaches and the tabu search strategy to solve models in which a subset
of the discrete variables is responsible for a large portion of the computational com-
plexity of the problem. This algorithm takes simultaneous advantage of the highly
structured tree search from B&B and the memory flexibility from tabu strategies to
reduce computational requirements without significantly compromising the qual-
ity of the solutions. The method is applied to the optimization of a supply chain
containing continuous flexible process networks. The results show that the method
is capable of finding optimal solutions to medium-scale problems and solutions
with average gap less than 0.2% for large-scale instances, when compared to full-
scale solutions.
10.1
Introduction
Chemical process industries are usually composed of various production sites with
many processes and flows of chemicals whose interconnection forms a chemical
process network. This network ensures the maximum flexibility in production by
utilizing the various synergies between the processes, which are classified as either
dedicated or flexible. While dedicated processes are used for chemicals produced
ISBN: 978-3-527-31693-9
302 10 A Hybrid Tabu-Branch and Bound Approach
in large volumes and only operate with one production scheme at all times, flexible
processes are used for small-scale chemicals and can operate with different produc-
tion schemes at different times. Multiperiod models for dedicated [1] and flexible
[2, 3] process networks have been proposed. Bok et al. [4] extended the previous
works by incorporating inventory profiles, changeover costs, intermittent supplies,
production shortfalls, and transportation costs for delivery when the network con-
sists of processes located at different sites.
The incorporation of various features in process network models, such as the one
proposed by Bok et al. [4] made them relatively difficult to solve due to their larger
size and complexity. To make the model more realistic and allow it to accommodate
larger networks, several solution strategies based on decomposition were proposed.
Bok et al. [4] implemented a bilevel decomposition method. In fact, there is an in-
creasing number of applications of decomposition methods to models in schedul-
ing [5], planning [6], and supply chain optimization [7, 8]. The above-mentioned
works conclude that the computational efficiency is improved significantly with
the use of decomposition, at the expense of a relatively high implementation effort.
An often employed alternative to decomposition-based deterministic solution
strategies to large-scale process planning problems is the use of metaheuristic pro-
cedures that are capable of finding good feasible solutions in a convenient time
frame. Among metaheuristic strategies, tabu search [9] has been widely regarded
as an effective approach that is frequently able to avoid suboptimal traps that occur
with other such methodologies. Recently, the literature has presented a number
of successful attempts to apply tabu search strategies to solve problems originated
from process planning and scheduling [10–13]). The main advantages associated
with the use of tabu search procedures lie on their ability to systematically avoid
previously exploited suboptimal regions in an attempt to obtain higher quality so-
lutions. As is the case with metaheuristic approaches, however, tabu search strate-
gies do not account for optimality criteria and, therefore, must be employed as a
means of finding solutions with acceptable quality with relatively small computa-
tional requirements and higher ease of implementation.
The objective of this work is to develop a hybrid approach within a tree enu-
meration framework that can make use of the inherent advantages of determinis-
tic and metaheuristic strategies. This approach makes use of the parametric tabu
search procedure for the solution of MILPs and of special topological characteris-
tics presented by some combinatorial problems. The hybrid approach is then im-
plemented in a continuous flexible process networks (CFPN). The results from the
full-scale solution and the results from the hybrid approach are then compared and
contrasted.
This chapter is organized as follows: Section 10.2 presents the multiperiod model
for the CFPN and discusses the computational complexity of different continuous
relaxations of the model. In Section 10.3, the proposed hybrid approach is derived
and a reinterpretation of the parametric tabu search algorithm proposed by Glover
[14] is offered. Section 10.4 contains the results of the implementation of the pro-
posed approach for the solution of instances of the CFPN problem with different
planning horizons, and a discussion about the efficacy of this approach with its ad-
vantages and limitations. Finally, Section 10.5 reviews the main accomplishments
of the presented work and offers concluding remarks. Notation for the method is
provided at the end of the chapter.
10.2
Problem Statement
The CFPN model of Bok et al. [4] is used as a case study for the present work. The
CFPN model is a production/distribution network consisting of a set J of chem-
icals and a set I of processes that interconnect in a finite number of ways. Each
process i has a set Ki of production schemes and can be either dedicated or flexi-
ble. Dedicated processes have only one production scheme and flexible processes
have two or more production schemes. A set Jik of chemicals is involved for each
process i with a specific scheme k. The chemicals can be raw materials, interme-
diates, products, or byproducts. The chemicals are bought or sold from a set L of
markets. The raw materials are purchased from a set of LP markets and the end
products are sold in a set of LS markets. All the processes are located in a set of
C production sites. Each site has its own set of processes and production schemes
and they need not to be identical. Figure 10.1 shows the general structure of the
CFPN model.
Fig. 10.1 General continuous flow process network diagram.

10.2.1
Mathematical Structure of the CFPN Model
The objective of the CFPN model is to maximize the operating profit of the network
over a planning horizon. The planning horizon is expressed as a set of time periods
of equal length during which prices and demands of chemicals and costs of oper-
ation and inventory can vary. For all processes and schemes, the material balances
for raw materials and byproducts are assumed to be linear ratios to the main prod-
uct of that scheme. All the processes are assumed to have fixed costs. The operating
costs for each process and production scheme are assumed to be proportional to
the amount of main product produced for that scheme. Finally, changeovers are
assumed to only imply costs and the overall time spent in changeover is assumed
negligible. The detailed description of the objective function (OF) and the list of
constraints of the CFPN model is given in the appendix to this chapter. The set of
decision variables in the model is composed of two sets of binary variables, Y and
YP. As is evident from the complete model formulation in the appendix, the CFPN
model can be written in the following generalized form:
(CFPN) max z = α T x + β T Y P ,
s.t.: A.x + B.Y b11 ,
C.x + D.YP b12 ,
(10.1)
E.x + F.Y = b21 ,
x L x x U , x ∈ Rm ,
Y ∈ {0, 1}n1 YP ∈ {0, 1}n2 ,
where the vector x contains all the continuous operating variables. The definition
of model CFPN presented in (Section 10.2.2) shows that the sets of constraints that
depend on variable Y and on variable YP are mutually exclusive. This particular
characteristic implies that the structure of these sets of variables may be analyzed
for they impact on the solution for the problem. A further analysis of the structure
and complexity of the model is presented in the following section.
10.2.2
Preliminary Analysis of the CFPN Model Solution Approaches
The CFPN model proposed by Bok et al. [4] incorporates various features, such as
intermittent delivery and intersite transportation, to give a more practical represen-
tation of the real world supply chain. However, it also makes the model relatively
difficult to solve as the planning horizon considered expands and as the number of
chemicals and/or processes involved increases. This is shown in Chen and Pinto
[8], where the rise in complexity makes the problem scale nonpolynomially with the
number of time periods. In dealing with the rise in complexity, Chen and Pinto [8]
proposed several decomposition strategies, relaxing the binary variables and link-
ing constraints, and solved the model in a period-by-period fashion by relaxing the
constraints connecting different time periods. It is evident from their results that
most of the complexity comes from the two sets of decision variables, namely, the
(a) Computational time (b) Solution value

Fig. 10.2 Comparison between full scale and relaxation:
(a) computational time and (b) solution value.
assignment variable Y , and the delivery variable YP. To examine the level of com-
plexity each exerts on the model, variables Y and YP are relaxed, either individually
or together, between 0 and 1. Figure 10.2 shows the comparison in computational
times and solution values between the full scale and the relaxations.
From the comparison, the relaxation of variable YP resulted in a greater reduc-
tion in computational time than that of variable Y . An examination of the model
shows that variable YP affects the purchase of raw materials, which in turn affects
most other major decisions in the supply chain, while variable Y only affects the
decisions on changeover. Another observation is that when YP is relaxed between
0 and 1, it removes the limitation of the number of deliveries allowed within a set
of time periods. This relaxed the model dramatically since the delivery of raw ma-
terials is now allowed for every time period instead of limiting it to just once over a
set of consecutive time periods.
It is evident from the above analysis that the majority of complexity comes from
the decision variable YP. If the definition of the optimal values of YP can be rele-
gated to a procedure that is less sensitive to the complexity this variable imposes
to structured methods, then the computational efficiency of the solution proce-
dure can be improved significantly. With this observation in mind, an alternative
approach is devised using the traditional Branch & Cut (B&C) algorithm and the
parametric tabu search procedure proposed by Glover [14]. The following section
gives a detailed description of this hybrid approach.
10.3
Hybrid Tabu Search Procedure
The metaheuristic approach devised for the solution of the studied supply chain
management model is based on the parametric tabu search procedure introduced
by Glover [14], who employed a tabu guidance framework for the solution of mixed
integer programs, relying solely on continuous linear relaxations to obtain feasible

solutions.
In this chapter, we propose the use of a tabu search framework to deal with only
a subspace of the set of discontinuous variables, present in a given mixed integer
programming model, which accounted for the vast majority of the computational
complexity of the original problem. This approach enables us to exploit the advan-
tages of a structured Branch & Cut (B&C) algorithm to deterministically find inte-
ger solutions for only the subset of binary variables that does not present prohibit-
ing computational challenges, while it simultaneously makes use of the flexibility
presented by metaheuristic frameworks to handle the binary variables for which a
rigorous solution algorithm incurs in an explosion of computational effort.
The hybrid tabu search approach derived in this work makes use of successive
deterministic solutions of a relaxed mixed-integer version of the original problem,
in which the set of complicating variables is allowed to be assigned continuous
values between their integer limits, and the integrality of the noncomplicating vari-
ables is still enforced in the solution. The metaheuristic approach, then, consists
of enforcing the integrality of the relaxed variables by adding to the objective func-
tion the minimization of a Manhattan distance between the vector of complicating
variables and integer values recursively defined. A tabu strategy further defines
“forbidden” areas of the search space in order to avoid cycling and local trapping.
The following is a thorough description of the computational procedure.
10.3.1
Mixed Integer Linear Relaxation
The algorithm proposed in this work is, in theory, applicable to a wide range of
MILP problems and, indeed, will coalesce into the original framework from Glover
[14], in the extreme case in which all discrete variables are considered complicating
and are let to be treated by the metaheuristic implementation. However, such an
approach would present the entire spectrum of drawbacks inherent to any heuristic
strategy without necessarily adding advantages to well established structured solu-
tion strategies, such as B&C. Therefore, it is assumed that the general algorithmic
description presented is applied to a combinatorial problem in which the subset of
the integer variables that are responsible for most of the solution effort is known a
priori. Furthermore, we offer a reinterpretation of the general procedure by Glover
[14] for dealing with this subset of variables. Additionally, it is assumed, without
loss of generality, that the complicating integer variables are binaries. From this
perspective, the MILP problem can be posed in general form as

(MILP): min z = ci wi + dk vk ,
i∈N k∈Q
s.t.: (wi , vk ) ∈ W ∀i ∈ N, ∀k ∈ Q, (10.2)
wi ∈ {0, 1} ∀i ∈ N,
W = {(w, v): A1 w + A2 v b1 , A3 w + A4 v = b2 , 0 wi 1 ∀i ∈ N} ,
where w is the set of complicating binary variables, v is a set that contains the
remaining model variables, which can be continuous, integer or binary, N =
{1, 2, . . . , n} is the index set for w, and Q = {1, 2, . . . , q} is the index set for x.
The tabu search procedure is based on the solution, at each iteration, of a re-
laxation of Eq. (10.2) in which each component of w assumes continuous val-
ues between 0 and 1. Let N + and N − be any two subsets of N such that N =
N + ∪ N − , N + ∩ N − = ∅. The method, then, proceeds by attempting to enforce the
following conditions:
wi 1 ∀i ∈ N + , (10.3a)
wi 0 ∀i ∈ N − . (10.3b)
Instead of being enforced as constraints, as is usually the case of branch and bound
methods, conditions (10.3a) are added as penalties to the objective function, which
results in the following relaxation for the original problem:.

(MILP ): min z = ci wi + dk vk
i∈N k∈Q

+ Mj 1 − wj + Mj wj , (10.4a)
j ∈N + j ∈N −
s.t.: (wi , vk ) ∈ W ∀i ∈ N, ∀k ∈ Q, (10.4b)

ci wi + dk vk z0 − ε, (10.4c)
i∈N k∈Q
W = {(w, v): A1 w + A2 v b1 , A3 w + A4 v = b2 , 0 wi 1 ∀i ∈ N} ,
where Mj is a large positive constant, z0 is an upper bound to the optimal solu-
tion of the original problem, and ε is a small positive constant. Constraint (10.4c)
defines a hyperplane that cuts the feasible region of (MILP ) and excludes all solu-
tions with objective values greater than or equal to the current best integer solution.
If no integer solution is assumed to be known, z0 is set to a large positive value (say,
∞) and problem (MILP ) is feasible for any (w, x) feasible to the original problem.
Moreover, a feasible solution to the relaxed problem in which w is integer is also
feasible to the original problem. It is important to note that if w∗ corresponds to the
values of the complicating variables in the optimal solution of the original problem,
and if
N = N with N + = {i|wi∗ = 1} and N − = {i|wi∗ = 0}, (10.5)
then the optimal solution to (MILP ) results in w = w∗ , provided that the constant
M is appropriately large. Therefore, the problem reduces to finding the compo-
nents of the sets N + and N − so that the optimal solution to (MILP ) corresponds
to the optimal solution to the original problem. However, defining the components
of N + and N − in Eq. (10.5) constitutes precisely the same problem as the one
of finding the values of w∗ and thus, it incurs the same complexity of the original
problem. Nevertheless, it must be noted that conditions (10.5) are sufficient but not
necessary for obtaining w = w∗ . Moreover, the goal of the proposed metaheuristic
approach is primarily to generate, in a convenient time frame, feasible solutions to

the original problem that are as close as possible to the optimum, since no optimal-
ity criteria can be used in first place. Note that the procedure employs a relaxation
of the original model in which a subset of the binary variables is not relaxed to
continuity and, therefore, the solution of this relaxation is able to provide a tighter
lower bound to the solution of the full-scale problem than is the one obtained by a
rigorous LP relaxation. However, metrics of the quality of feasible integer solutions
based on the OF value of the proposed partial relaxation are highly dependent on
the topology of the problems to be solved and, therefore, fail to provide comparable
measurements of distance to optimality.
In Section 10.3.2, a description of the recursive procedure that iteratively updates
the components of N + and N − in an attempt to continuously obtain better feasible
solutions to problem (MILP) is provided.
10.3.2
Algorithmic Description
The hybrid approach derived in the present work is based on the use of a tabu
search guidance procedure to find integer – and therefore feasible – solutions for
a set of complicating variables in a mixed integer linear problem. As mentioned
in the previous section, such method is based on the recursive solution of sub-
problems in which the integrality constraints of the set of complicating variables
are relaxed, resulting in a problem that requires substantially less computational
effort to be solved by a deterministic B&C procedure. Hence, the metaheuristic
problem is to define the composition of sets N + and N − in a manner that the
solution to the relaxed problem is feasible and ultimately optimal to the original
problem. The assignment of components to these sets is based on the results of
the solutions of (MILP ) from previous iterations, and the procedures underlying
these mechanisms are described in this section.
We start composing the algorithm by analyzing the possible outcomes of the so-
lution of problem (MILP ). Let N + and N − have arbitrary components (respecting
their definition) in some iteration of the algorithm. The optimal solution for this
instance of (MILP ) may fall into three distinct cases:
Case 1: All values of w are integer. In this case, the optimal solution to (MILP ) is
also feasible to (MILP) and, therefore, the upper bound to the optimal solution of

the relaxed problem is updated to: z0 = i∈N ci wi + k∈Q dk vk . In this case, no
changes in the composition of N + and N − are required since the goal of defining
these sets in order to obtain solutions that are feasible to (MILP) have been attained.
Case 2: One or more components of w are not binary but none of these compo-
nents belongs to N . We name this case an unrestricted fractional solution, in agree-
ment with the terminology proposed by Glover [14]. In such case, the objective
function is able to zero the distance between wi |i ∈ N and the parameters defined
in constraints (10.3a) and (10.3b), but no integer solution is obtained. This is an
indication that the current components of N are not sufficient to ensure feasibility
to the original problem and the algorithm proceeds to systematically add compo-
nents to N + and N − . This situation necessarily arises in the first iteration of the
algorithm, when N is initialized as an empty set, unless the proposed relaxation
generates an integer, and therefore, optimal solution.
Case 3: One or more components of w are not binary and at least one of these
components belongs to N . This situation is referred to as restricted fractional solu-
tion. In this case, the objective function of the relaxed problem is unable to enforce
constraints (10.3a) and (10.3b). If the values of M are sufficiently large, this re-
sult indicates that the candidate integer solution which the components of N are
attempting to enforce is not feasible to the original problem, and, therefore, the
components of N + and N − are overly stringent and the algorithm proceeds to sys-
tematically delete elements from these sets.
The algorithmic steps required when solutions of (MILP ) fall into Case 1 are
simple and have been described. In the next sections, the iterative steps for the
cases of unrestricted fractional solutions (Case 2) or of restricted fractional solu-
tions (Case 3) are characterized.
10.3.2.1 Unrestricted Fractional Solutions

Let F = {i ∈ N|wi ∈ / {0, 1}} be the set of indexes from variables that are frac-
tional in a given optimal solution to (MILP ). Let G = {i ∈ F |i ∈ N } and
H = {i ∈ F |i ∈ / N }. Case 2, as defined, corresponds to the situation in which
H = F and G = ∅ for an instance of the relaxed model. Hence, the response
to this case consists of the definition of a set H0 ⊂ H whose components must be
added to N . The criterion for selection of the components of H0 follows the simple
rule of the smallest distance to the bounds or, similarly, the largest distance to the
centroid of the domain of each component, although more sophisticated criteria
may be employed, at the expense of simplicity of implementation. This distance
can be defined by:

RiH = wi − wiU − wiU /2 = |2wi − 1|/2. (10.6)
Therefore, set H0 is defined by:
CH0

H0 = i1 , . . . , iCH0 ik = arg max RjH , (10.7)
j ∈H
k=1 / {ik−1 ,...,i1 }
j∈
where CH0 , can be defined by the appropriate choice of a parameter h0 ∈ [0, 1],
such that CH0 =
h0 .|H |, which may be tuned in an attempt to improve perfor-
mance.
The new composition of N can be subsequently defined by imposing condition
(10.3a) in the case that the fractional variable is found closer to its upper bound and
condition (10.3b) otherwise. This procedure can be summarized as
if (1 − wi ) (wi ) for i ∈ H0 , then set: N + ← N + + {i} ,
(10.8)
if (1 − wi ) > (wi ) for i ∈ H0 , then set: N − ← N − + {i} .
10.3.2.2 Restricted Fractional Solutions

Restricted fractional solutions are defined by the case in which the set G is non-
empty. As previously discussed, if sufficiently large values of M are used, this cor-
responds to a situation in which the conditions corresponding to Eqs. (10.3a) and
(10.3b) attempt to enforce an integer solution that is infeasible to problem (MILP).
Thus, the responses to this case follow the opposite direction of those of the case of
an unrestricted fractional solution, and components must be subtracted from the
set N . Similarly to the procedure that handles unrestricted fractional variables, a
set G0 ⊂ G is defined and the following condition is enforced:
if i ∈ G0 , then set: N ← N − {i} . (10.9)
The components of G0 are chosen from the components of G as the ones that
present the strongest resistance to the conditions enforced by Eqs. (10.3a) and
(10.3b). A simple measure of this resistance can be defined by the absolute dis-
tance from the value of the restricted variable in the optimal solution of the relaxed
problem to the imposed bound. Thus we define this resistance as

RiG = (1 − wi ) ∀i ∈ G ∪ N + ,
(10.10)
RiG = (wi ) ∀i ∈ G ∪ N − .
The set G0 is then defined by
CG0

G0 = i1 , . . . , iCG0 ik = arg max RjG , (10.11)
j ∈G
k=1 / {ik−1 ,...,i1 }
j∈
where CG0 is defined as CG0 =

g0 .|G|, where g0 ∈ [0, 1] is a tunable parameter
of the algorithm.
In addition to the update of the components of N , an exponential update to the
values of Mi for the indexes that were chosen to compose G0 is also implemented,
such that:
Mi ← (1 + δ) · Mi ∀i ∈ G0 ,
where δ is a positive constant. This updating strategy to M is implemented in an
effort to minimize the possibility of cycling within specific sequences of solutions.
It is important to note that although the tabu search guidance described in the next
section helps avoid cycling, it does not necessarily prevents it, since this implemen-
tation is itself cyclic.
10.3.2.3 Tabu Search Guidance

Metaheuristic algorithms based on tabu search use imposed restrictions on possi-
ble solutions as a way of avoiding the convergence to nonoptimal regions of the
search space, which are classified as tabu for a certain number of iterations of
the algorithmic procedure. In the implementation of the present hybrid approach,
the responses to unrestricted fractional solutions are dynamically associated with
a tabu status during a defined number of iterations. We define the set of indexes
of variables that are in tabu status as TB ⊂ N, whose composition is updated in
each iteration. Thus, for any given iteration, the composition of H0 is modified to
exclude tabu indexes in the following manner:
CH0

H0 = i1 , . . . , iCH0 ik = arg max H
Rj . (10.12)
j ∈H
k=1 j ∈T
/ B
/ {ik−1 ,...,i1 }
j∈
We further define the criteria for indexes to enter and exit set TB, thus allowing
and restricting responses. Following the procedure defined by Glover [14], we will
add to set TB the indexes of all unrestricted fractional variables selected to compose
H0 at any iteration. Furthermore, it is imposed that indexes remain in set TB for
a predefined number of iterations, Ttenurei , ∀i ∈ N, whose values can be either
fixed or dynamic, before being removed from it. To keep track of the tenure of
a specific index as in “tabu” status, we define a parameter, Tstarti , that is set to
the iteration number in which the corresponding variable was defined as tabu. We
chose to implement the tabu search procedure in the following way:
(1) During the initialization phase of the algorithm, all values of
Ttenurei are set to the value Tmin iterations.
(2) In any iteration, the values of Ttenurei for all indexes that
compose H0 are increased by TB unit and the
corresponding Tstarti is set to the current iteration, after the
composition of the set TB is updated.
(3) After a predefined number of TIter iterations, the values of
Ttenurei are reset to their initial values.
Tabu search strategies are often implemented in a nondeterministic way, in an at-

tempt to avoid the exclusion of high quality solutions from the search space. In the
present implementation, we associate a probability γ that each of the indexes in
TB may be able to be selected for H0 , at each iteration. Thus, there is a probability
γ that any given index will be subtracted from the set TB for exactly one iteration,
after which the set TB is recomposed to its previous state. It is important to stress
that this manipulation of the set TB occurs apart from the tabu status update proce-
dure and, therefore, does not interfere with it. The associated probability γ allows
for the stochastic tuning of tabu search guidance procedure. When this probability
is in its lower bound (γ = 0), the tabu search guidance is implemented determin-
istically, whereas when it is on its upper bound (γ = 1), the tabu search guidance
is ignored.
10.3.3
General Algorithm
The general algorithm for the tabu search-based hybrid approach to solving the de-
fined continuous flexible process network problem is stated below, and a flowchart
representation can be found in Fig. 10.3.
Initialization
Set N
Iter 0
Step 3 z0
Update Ttenure, Step 1

Solve (MILP’)
Tstart, M
(MILP’) no
Stop
feasible? Current Solution
is optimal
Update yes
N+, N-, TB
yes
w Integer? Update z0
no
Step 2
Create H0
yes Generate
G=Ø?
F, H, G
no
Create G0
Fig. 10.3 Flowchart of the hybrid tabu search algorithm.
Initialization:
Set N = ∅,
F, G, H = ∅,
Iter ← 1,
z0 ← +∞,
wi∗ ← 0 ∀i ∈ N,
Ttenurei ← Tmin ∀i ∈ N,
Tstarti ← 0 ∀i ∈ N.
Step 1 – Solve (MILP ) to find an optimal solution to the relaxed problem. If
(MILP ) is infeasible, stop and set the current solution as the optimal so-
lution. Otherwise, if wi is integer for all i ∈ N, set z0 = z , wi∗ = wi ∀i ∈ N
and restart Step 1. If at least one value of wi is not integer, build the sets
F, G, and H and go to Step 2.
Step 2 – If G = ∅, generate the set H0 , according to Eq. (10.12) and update the sets
N + and N − according to Eq. (10.8). Go to Step 3.
If G = ∅, generate the set G0 , according to Eq. (10.11) and update set N
according to Eq. (10.9). Set Iter ← Iter + 1 and go to Step 1.
Step 3 – Update set TB according to the results of Step 2 by setting T B ←
T B +H0 , and T B ← T B −{i ∈ T B|Tstarti + Ttenurei = Iter.}. Set Tstarti ←
Iter ∀i ∈ H0 and Ttenurei ← Ttenurei + T B ∀i ∈ H0 . Set Iter ← Iter + 1
and go to Step 1.
Since no optimality criteria is employed, apart from the termination due to infea-
sibility of the relaxed problem, we specify a maximum number of iterations after
which the procedure will be halted and the current best solution is returned as the
final solution from the algorithm. Note, however, that the termination after a pre-
defined number of iterations is the most frequent outcome of the algorithm, since,
assuming that the value of ε is small enough, the only instance in which the defi-
nition of the optimal solution of the original problem would result in an infeasible
relaxed problem would be in the uncommon situation that (MILP ) has zero gap
with respect to (MILP) – i.e., both problems have the same OF value in their op-
timal solution, although the solution to (MILP ) is not required to be integer with
respect to the relaxed variables.
10.4
In Section 10.2, it was shown that the set of binary variables YP, which accounts
for a relatively small fraction of the total number of discontinuous variables in the
(CFPN) model, in fact contributes to most of the computational effort required
to deterministically solve the problem. In Section 10.3, an algorithmic procedure
based on tabu search that relies on the successive solution of problems in which a
set of complicating binary variables is relaxed was introduced.
The discussion in Section 10.2.2 shows that there is a strong dependence of com-
putational time requirement to the number of time periods used in the model, and
that this dependence is much weaker for the case in which YP is relaxed to a set
of continuous variables ranging from 0 to 1, even though the number of binary
variables still increases linearly with the number of time periods in the relaxed
problem. This is the behavior on which the hybrid tabu search approach relies for
the solution of the (CFPN) problem.
The hybrid algorithm was applied to the (CFPN) problem assuming YP as the
complicating set of variables. Five different instances of the (CFPN) model, with
planning horizons of 7, 21, 42, 63, and 84 days were used (all with τc = 2). In
the remainder of the text, we will add the number of time periods to the name
of the model instance in order to make a clear distinction between problems of
different scales, whenever this distinction is necessary – e.g., model (CFPN) with
63 time periods is named (CFPN63). It is also noted that the algorithm was defined
assuming a minimization problem is at hand, whereas the (CFPN) model involves
profit maximization. Modifying the algorithm to deal with maximization problems
is straightforward.
The algorithmic procedure was implemented to all five different size models
using parameter values fine tuned for the intermediate sized problem (CFPN42).
All parameters were kept constant for the solution of problems of different sizes
to increase the significance of the comparison of results and run statistics. The
parameter values employed for the tabu guidance were Tmin = 3 and TB = 1,
and the M update value chosen was δ = 0.07. Moreover, it was verified that the
algorithm is not significantly sensitive to values of γ within the range of 0–0.3, but
values of γ above 0.5 tend hinder the tabu search guidance implementation and
decrease the efficiency of the procedure. A value of γ = 0.1 was employed. The
dependence of the algorithm performance with respect to parameters h0 and g0
was found to be much more significant than the algorithm sensitivity to the tabu
guidance parameters. The choice of small ( 0.1) values for h0 and g0 naturally
resulted in an increased time necessary to find integer solutions to the original
problem, since, in such cases, only a small fraction of the number of restricted and
unrestricted fractional variables are dealt with in each iteration. In contrast, high
values of these parameters ( 0.7) obstructed the progression of the algorithm
because a large fraction of the relaxed variables were kept in set TB, preventing the
generation of a significant set H0 . The values of h0 and g0 used for all the results
presented were 0.25 and 0.25, respectively.
The algorithm was implemented in the GAMS modeling environment [15] and
both the full scale and the relaxed MILP problems were solved with the CPLEX
10.1 library [16].
10.4.1
Algorithmic Analysis
This section presents a detailed analysis of the solution procedure applied to the
(CFPN) model with different planning horizons. The results are then compared to
the solutions of the full-scale problems.
The analysis of the behavior of the hybrid tabu search algorithm for the solu-
tion of the continuous flexible process network problems begins with a thorough
examination of the solution path of the procedure. Figure 10.4(A) shows the run
history for the solution of problem (CFPN42), which was solved to optimality in
64 iterations of the algorithm. The thin line represents the objective function value
for the relaxed problem obtained in each iteration. The thick line represents the
value of the best integer solution found in each iteration and, thus, represents an
lower bound for the OF of the relaxed problem. The dashed line corresponds to the
optimal solution to the problem, obtained by solving the full-scale problem. Fig-
ure 10.4(B) shows the cardinality of the sets N , H0 , and G0 in each iteration of the
algorithm.
It is observed that the OF value for the first iteration corresponds to the low-
est bound on the solution to the relaxed problem, since the initial solution of this
problem is the only one in which all conditions from Eqs. (10.3a) and (10.3b) are
guaranteed not to be enforced. Therefore, this initial relaxed solution may be inte-
ger, in which case the optimal solution to the original problem is obtained initially,
or, as is the present case, unrestricted fractional, and the set H0 is then assembled
and its components added to N + or N − (note that N = N + ∪ N − ). Figure 10.4(B)
shows that the cardinality of H0 has a high value after the first iteration and imme-
diately starts decreasing as the set N is being populated. The first few iterations
commonly display similar behavior, with components being added to or subtracted
from N as unrestricted or restricted fractional solutions are found. In Fig. 10.4(A),
4
x 10
1.8
1.7
OF value
1.6 Relaxed objective

Current best solution
Optimal solution
1.5
0 10 20 30 40 50 60 70
100
80
Cardinality
60 |N |
|H0|
40
|G0|
20
0
0 10 20 30 40 50 60 70
Iteration
Fig. 10.4 Solution history for the model (CFPN42).

(Top) Objective function values for the relaxed problem, best
integer solution, and optimal solution from the full-scale
problem. (Bottom) Cardinality of sets N , H0 , and G0 .
this situation is represented by the steep descending thin line, followed by a plateau
in the initial part of the plot. Subsequently, after the first integer solution is found,
the algorithm updates the upper bound for the objective of the relaxed problem,
thus deeming any addition to N that would result in further increase in the OF as
unbinding. The algorithm proceeds to obtain unrestricted and restricted fractional
solutions and by updating the lower bound to the relaxed objective as improved
integer solutions are found, which is represented by the increasing thick line. Note
that peaks in the cardinality of H0 and G0 are perfectly associated with a respec-
tive steep increase or decrease of the number of components in N , which clearly
exemplifies the functioning of the algorithm.
It is further worth to note that during the algorithmic procedure applied to the
model (CFPN42) most of the improvement in the solution values to the original
problem occurs rapidly during the initial 30–35 iterations, whereas further im-
provement takes place in a much slower pace. It is observed in Fig. 10.5 that the
same behavior is found during the solution of problems of different magnitudes,
with exception of the case with 7 days as planning horizon, for which the first in-
teger solution found coincided with the optimal solution of the original problem.
These results demonstrate the usefulness of the procedure as a means of finding
Horizon: 7 days Horizon: 21 days

2998 9000
Gap = 0 8900
2997.5
OF value 8800
2997 Gap = 0
8700
2996.5
8600
2996 8500
0 1 2 3 4 0 20 40 60
4 Horizon: 63 days 4 Horizon: 84 days

x 10 x 10
2.7 3.6
3.5
2.6
OF value
3.4
Gap = 0.000235 Gap = 0.000282
3.3
2.5
3.2
2.4 3.1
0 100 200 300 0 100 200 300
Iteration Iteration
Fig. 10.5 Solution paths for models with planning horizons of 7, 21, 63, and 84 days.
Table 10.1 Comparison of tabu search solutions to full-scale solutions.
Time horizon Tabu solution Optimal solution Gap (%) No. of Iterations
7 days 2996.7 2996.7 0.00 3

21 days 8891.1 8891.1 0.00 54
42 days 17709 17709 0.00 64
63 days 26518 26524 0.024 300
84 days 35335 35344 0.028 300
generally suboptimal but good feasible solutions for the complex problems studied
in this work, with reduced computational effort.
10.4.2
Comparison to Full-scale Results
The original full-scale models (CFPN) for different planning horizons were solved
to optimality with the use of the solver package CPLEX 10.1. Table 10.1 contains
a summary of the solutions obtained by the metaheuristic approach and by the
commercial solver.
The tabu search procedure was able to find the optimal solutions for problems
with planning horizons of 7, 21, and 42 days, whereas for the problems with 63

5 15
4
10
Gap (%)
2
5
1
0 0
0 20 40 60 0 20 40 60 80

10 15
8
10
Gap (%)
4
5
2
0 0
0 100 200 300 0 100 200 300
Iteration Iteration
Fig. 10.6 Gap history for problems with time horizons greater than or equal to 21 days.
and 84 time periods, only suboptimal solutions with gaps of 0.024% and 0.028%,
respectively, could be provided, given the limit of 300 iterations specified. How-
ever, it is important to note that, as stressed in the previous section, the hybrid
procedure implemented does not utilize any optimality criterion and, therefore, it
does not have the definition of a proven optimal solution as a final goal. Rather,
as is the case with metaheuristic procedures, the purpose of the algorithm is to be
able to provide solutions acceptably close to the optima using little CPU resources.
The results presented in Fig. 10.6 show that solutions with gaps smaller than 1%
were obtained within a relatively small number of iterations (100–150) for the prob-
lems not solved to optimality. Additionally, each iteration of the proposed algorithm
is computationally inexpensive, and the actual solution of the relaxed problem is
its most resource intensive part. Furthermore, since it is an unstructured solving
procedure that does not recall information about past solutions with exception of
scalar bounds, the composition of set N and the tabu guidance implementation,
the memory requirements for the algorithm are minimal.
Despite the ease of implementation when compared to structured algorithms,
it must be noted that the method suffers from the drawback of being somewhat
sensitive to parameters, such as h0 and g0 , which are likely to require specific tun-
ing approaches for solving different models. As discussed in the introduction of
this section, a single set of parameter values were used for the solution of prob-
(a) Full scale
(b) Proposed approach
Fig. 10.7 Solution value of delivery variable YP for the 63 time

period horizon: (a) full scale and (b) proposed approach.
lems with different planning horizons, for the purpose of coherence regarding the
comparison of different results. These parameters were initially adjusted for prob-
lem (CFPN42), which, as a result, required relatively less computational effort than
the remaining models. The solution quality and performance of each of the other
problems would benefit significantly from individualized parameter fine tuning.
In addition, it is important to consider an analysis of the effect of sub-optimality
in the solutions obtained by the proposed algorithm. It is known that suboptimal
solutions of combinatorial models even with reduced optimality gaps often present
considerable distances to the optimal one. To analyze such behavior in the studied
model, the solution of variable YP obtained from the full scale and the proposed
approach for the 63 time period horizon is presented in Fig. 10.7.
These results show a significant agreement in the total deliveries made across
the 63 time periods – 83 times in both the full-scale optimal solution and the sub-
optimal solution obtained by the proposed approach. This concurrence can be asso-
ciated with the fact that all feasible solutions need to conform to the same demand
profile, differing only with respect to the profitability of the strategy found. How-
ever, in contrast with the total number of deliveries, the delivery profile between
markets and sites at each time period is different. The delivery profiles across all
time periods appear to be deviated and uncorrelated between the optimal solution
and the suboptimal feasible points obtained with the proposed approach. This re-
sult exemplifies the typical behavior of combinatorial problems and demonstrates

the model’s flexibility to generate discrepant operation profiles with very small dif-
ferences in resulting profitability.
10.5
Conclusions
A hybrid algorithm for the solution of mixed-integer linear programming problems

in which a small subset of the discrete variables accounts for most of the compu-
tational complexity of the model has been proposed in this work. The algorithm
builds upon the parametric tabu search procedure proposed by Glover [14] and the
use of a structured B&C solver to harness the advantages of both methods.
The proposed method relies on successive solutions by B&C of a relaxed prob-
lem in which only the complicating variables are allowed to assume continuous
values and feasible solutions with improving objective function values are recur-
sively found by the metaheuristic approach.
The procedure was then applied to a continuous flexible process network model.
This mixed integer linear problem was verified in a previous work to require non-
polynomial computational resources for its solution, with the increase of the prob-
lem size, which justifies the search for alternative suboptimal solution methods
[8].
The obtained results show that the proposed methodology was able to find the
optimal solution of small to large-scale instances of the model, whereas suboptimal
solutions to larger scale models presented optimality gaps smaller than 0.03%. It
was also verified that the algorithm was capable of obtaining solutions with gaps
smaller than 1% in a relatively small number of iteration (100–150), which makes
it a suitable procedure for the fast identification of good feasible solutions for the
studied problem.
Acknowledgment
JMP would like to acknowledge financial support from the Wechsler Award.
References
1 Sahinidis, N. V., Grossmann, I. E., processing networks with dedicated

Fornari, R. E., Chathrathi, M., and flexible plants. Ind. Eng. Chem.
Optimization model for long range Res. 29 (1991), pp. 1165–1171.
planning in the chemical industry. 3 Norton, L. C., Grossmann, I. E.,
Comput. Chem. Eng. 13 (1989), pp. Strategic planning model for complete
1049–1063. process flexibility. Ind. Eng. Chem. Res.
2 Sahinidis, N. V., Grossmann, I. E., 33 (1994), pp. 68–76.
Multiperiod investment model for
4 Bok, J. K., Grossmann, I. E., Park, S., 10 Shakib, S., Logendran, R., A mathe-
Supply chain optimization in contin- matical programming-based schedul-
uous flexible process networks. Ind. ing framework for multitasking en-
Eng. Chem. Res. 39 (2000), pp. 1279– vironments. Eur. J. Operat. Res. 176
1290. (2007), pp. 193–209.
5 Maravelias, C. T., Grossmann, I. E., 11 Lukac, Z., Soric, K., Rosenzweig,
A hybrid MILP/CP decomposition V. V., Production planning problem
approach for the continuous time with sequence dependent setups as a
scheduling of multipurpose batch bilevel programming problem. Eur. J.
plants. Comput. Chem. Eng. 28(10) Oper. Res., (2006), doi: 10.1016/j.ejor.
(2004), pp. 1921–1949. 2006.09.029.
6 Maravelias, C. T., Grossmann, I. E., 12 Bock, S., Rosenberg, O., A new paral-
Simultaneous planning for new prod- lel breadth first tabu search technique
uct development and batch manufac- for solving production planning prob-
turing facilities. Ind. Eng. Chem. Res. lems. Int. Trans. Oper. Res. 7 (2000),
40 (2001), pp. 6147–6164.
pp. 625–635.
7 Jackson, J. R., Grossmann, I. E.,
13 Al-Fawzan, M. A., An algorithm for
Temporal decomposition scheme
for nonlinear multisite production production planning in a flexible pro-
duction system. Comput. Ind. Eng. 48
planning and distribution models.
Ind. Eng. Chem. Res. 42 (2003), pp. (2005), pp. 681–691.
3045–3055. 14 Glover, F., Parametric tabu search
8 Chen, P., Pinto, J. M., Lagrangean- for mixed integer programs. Comput.
based techniques for the supply chain Oper. Res. 33 (2006), pp. 2449–2494.
management of flexible process net- 15 Brooke, A., Kendrick, D., Meeraus,
works. Proceedings of the 16th European A., Raman, R., GAMS: A User’s Guide,
Symposium on Computer-aided Process GAMS Development Corporation,
Engineering, 2006, vol. 21, pp. 2003– Washington, 1998.
2008.
16 ILOG, ILOG CPLEX 10.0: Users Man-
9 Glover, F., Laguna, M., Tabu Search, ual, 2006.
Springer, Berlin, 1997.
Appendix A.
Nomenclature
The hybrid tabu search methodology relies in the following notation:
Sets
N indexes of complicating variables
N+ indexes of variables associated with condition (10.3a)
N− indexes of variables associated with condition (10.3b)
N intersection of sets N + and N −
Q indexes of noncomplicating variables
F subset of N containing fractional variables at any iteration
G set of indexes from restricted fractional variables
G0 indexes chosen for response to restricted fractional variables
Appendix A. 321
H set of indexes from unrestricted fractional variables

H0 indexes chosen for response to unrestricted fractional variables
TB indexes of variables associated with a tabu status
Parameters
Mi penalties associated with complicating variable of index i ∈ N
h0 ratio between the desired cardinality of H0 and the cardinality of H
g0 ratio between the desired cardinality of G0 and the cardinality of G
CH0 desired cardinality of H0
CG0 desired cardinality of G0
RiH absolute distance from unrestricted fractional variable i to its closest
bound
RiG absolute resistance from restricted fractional variable i
Ttenurei number of iterations variable i must be kept as tabu
Tstarti iteration number in which variable i was set as tabu
TB increment in Ttenure when a tabu status is set
Tmin minimum allowed value for Ttenure
Variables
wi , i ∈ N complicating variables
vi , i ∈
/N noncomplicating variables
Continuous Flow Process Network Model
Objective function:

Max ZCFPN = γj lt Sj lct − ϕj lt Pj lct
j ∈J l∈L c∈C t∈T j ∈J l∈L c∈C t∈T

− δikct Wij kct − ξj c Vj ct
i∈I j ∈JMik k∈Ki c∈C t∈T j ∈J c∈C t∈T

− ζikk c Zikk ct − θj lt SFj lt
i∈I k∈Ki k ∈Ki c∈C t∈T j ∈J c∈C t∈T

− ωdlct YPdlct − φj ct Fj ct . (10.A.1)
d∈D l∈L c∈C t∈T j ∈J l∈L t∈T
Ratio of input chemicals to the main product
Wij kct = µij kc Wij kct

i ∈ Ij , j ∈ JIikc , j ∈ JMik , k ∈ Ki , c ∈ C, t ∈ T , (10.A.2)
Wij kct = µij kc Wij kct

i ∈ Ij , j ∈ JOikc , j ∈ JMik , k ∈ Ki , c ∈ C, t ∈ T . (10.A.3)
Limits production under available capacity
Wij kct ρij ck Qic i ∈ Ij , j ∈ JMik , k ∈ Ki , c ∈ C, t ∈ T . (10.A.4)
Indicates when changeover occurred (Zikk ct = 1)
Yikct + Yik c,t+1 − 1 Zikk ct

i ∈ Ij , k ∈ Ki , k ∈ Ki , c ∈ C, t ∈ T , (10.A.5)
Limits 1 production scheme per time period

Yikct = 1 i ∈ Ij , k ∈ Ki , c ∈ C, t ∈ T . (10.A.6)
k∈Ki
Mass balance of chemicals in the network

Vj c,t−1 + Wij kct + Pj lct + Fj ct = Vj ct
i∈Oj k∈Ki l∈L

+ Sj lct + Wij kct + Fj c t
l∈L i∈Ij k∈Ki c ∈C−{c}
j ∈ J, c ∈ C, t ∈ T . (10.A.7)
Prohibits production when no raw materials are delivered

Pj lct YPdlct PjUlct j ∈ J, l ∈ L, c ∈ C, t ∈ T . (10.A.8)
d∈D
Limits 1 delivery of chemicals over τc time interval

c
t+τ
YPdlct 1 d ∈ D, l ∈ L, c ∈ C, t ∈ T . (10.A.9)
t =t
Prevents purchasing from exceeding availability

Pj lct ajUlt j ∈ J, l ∈ L, t ∈ T . (10.A.10)
c∈C
Limits the product sales from exceeding maximum demand

Sj lct djUlt j ∈ J, l ∈ L, t ∈ T . (10.A.11)
c∈C
Shortfall penalty when the minimum demand is no met

SFj lt SFj l,t−1 + djLlt − Sj lct j ∈ J, l ∈ L, t ∈ T . (10.A.12)
c∈C
Bounds
Vj ct VjUct j ∈ J, c ∈ C, t ∈ T , (10.A.13)
Zikk ct 1 i ∈ Ij , k ∈ Ki , k ∈ Ki , c ∈ C, t ∈ T , (10.A.14)
Fj ct , Pj lct , Sj lct , SFj lt , Vj ct , Wij kct , Zikk ct 0,

Yikct , YPdlct ∈ {0, 1}. (10.A.15)
Appendix A. 323
Nomenclature for the CFPN Model
Sets
C production sites
D delivery methods
I production processes
Ij processes that consume chemical j
J chemicals involved in the process network
JIikc input chemicals involved in production scheme k of process i of site c
JMik main products for production scheme k of process i
JOikc output chemicals involved in production scheme k of process i of site c
K production schemes
Ki production schemes in process i
L markets
LP markets where chemicals are purchased
LS Markets where chemicals are sold
Oj processes that produce chemical j
T time periods
Parameters
ajUlt upper bounds for the amount of chemical j purchased from market l dur-
ing time period t
djLlt lower bounds for the amount of chemical j sold to market l during time
period t
djUlt upper bounds for the amount of chemical j sold to market l during time
period t
PjUlct upper bounds for the amount of chemical j purchased from market l for
site c during time period t
Qic capacity for process i at site c
VjUct upper bounds for the inventory amount of chemical j at site c during time
period t
WijUkc upper bound for the amount of chemical j produced from production
scheme k of process i at site c
δikct unit operating cost for production scheme k for process i at site c during
time period t
φj ct cost for transportation of chemical j to site c during time period t
τc minimum time interval that an intermittent delivery can be made for site c
ϕj lt price of purchase of chemical j in market l during time period t
γj lt price of sales of chemical j in market l during time period t
ρij kc relative maximum production rate of main product j , for production
scheme k in continuous flexible process i at site c
µij kc material balance coefficient for chemical j for production scheme k of
process i at site c
θj lt penalty cost for shortfall of chemical j in market l during time period t
ωdlct transportation cost of chemical j from market l to site c during time pe-
riod t
ξj c inventory cost for chemical j at site c
ζikk c changeover cost for changing from production scheme k to k’ in process I
at site c
Variables
Fj ct amount of chemical j shipped to site c during time period t
Pj lct amount of chemical j purchased from market l for site c during time
period t
Sj lct amount of chemical j made form site c and sold at market l during time
period t
SFj lt amount of shortfall of chemical j in market l during time period t
Vj ct amount of inventory of chemical j at site c during time period t
Wij kct amount of chemical j produced from scheme k of process i at site c during
time period t
Zikk ct relaxed binary variable: 1 if process i at site c operates with scheme k
during the time period t and operates with scheme k during the following
period t + 1; 0 otherwise.
Binary Variables
Yikct 1 if process i at site c operates with scheme k during time period t; 0
otherwise.
YPdlct 1 if delivery type d is available in delivering from market l to site c during
time period t; 0 otherwise.
325
11
Enhancement of Oil and Gas Production via Network
Optimization
Nort Thijssen, Kuei-Jung Li, Ilse Mittendorff
The oil and gas upstream supply chain consists of reservoirs, wells, and flowlines
to gathering stations (e.g., wellhead platforms) from which the multiphase stream
(gas–oil–water) is sent via a pipeline to a production platform. In a production
platform, the streams from several wellhead platforms are combined and separated
into gas and liquid (oil–water) streams. These products are delivered through a
network consisting of pipelines and processing platforms to the crude oil and gas
export facilities.
Analyzing the network in an integrated manner, compared to looking at the in-
dividual elements, is becoming increasingly important in the current business be-
cause of the change in liquid–gas composition due to aging of fields and the im-
plementation of new projects using existing facilities.
The aim of the integrated network analysis is to determine optimal timing of de-
velopments of new fields, identifying future bottlenecks in the network and evalu-
ating options for removal of these bottlenecks. This activity is also known as strate-
gic or investment planning. The nature of the problem is of the type mixed integer,
and by using piecewise linear modeling it is prevented that the problem becomes
nonlinear.
Integrated network analysis will help to identify the impact that all changes to-
gether have on the performance of the network. Such network can consist of only
the offshore facilities and pipelines, and can also include onshore processing facil-
ities.
Recent studies confirm that integrated network analysis helps to bring produc-
tion closer to the technical potential of the wellhead platforms. An example showed
10% acceleration of 15 years oil production and 5% percent increased recovery of
15 years gas production.
ISBN: 978-3-527-31693-9
326 11 Enhancement of Oil and Gas Production via Network Optimization
11.1
Introduction
The oil and gas upstream supply chain consists of reservoirs, wells, and flowlines
to gathering stations (e.g., wellhead platforms) from which the multiphase flow
stream (gas–oil–water) is sent via a pipeline to a production platform. To reduce
the complexity of the program, this study starts from wellhead platforms to the gas
and oil export facilities.
In a production platform, the flows from several wellhead platforms are first
combined and subsequently separated into gas and liquid (oil–water) streams. They
are delivered through a network consisting of pipelines and processing platforms
to the gas or crude oil export facilities. For all these individual elements in the
supply chain, processing facilities are modeled with a process simulator and flow
behavior in the pipelines is modeled with a pipeline simulation program. Analyz-
ing the network in an integrated manner, compared to looking at the individual
elements, is becoming increasingly important in the current business to identify
future bottlenecks in the entire network.
These bottlenecks can be a result of the following:
1. Field composition (gas–liquid) changing over time and
increased water content in the oil–water liquid stream due to
aging of fields. They lead to more restrictions in the
capacities of pipelines and water handling- or gas
compression facilities.
2. Increasing use of existing infrastructure for the
implementation of new projects to compensate the
production of aging fields or to increase overall production.
New projects that come on stream typically will use part of
the existing facilities, which may have sufficient capacity to
handle increase of production from one individual project,
but will not be adequate for other projects. Since there are
many projects and initiatives ongoing simultaneously, it
would be difficult, if not impossible, to assess the
interdependency of projects and initiatives without a
computerized planning tool.
3. The drive toward no venting and/or reduced flaring policy,
which is becoming increasingly important. This introduces a
higher gas compression requirement but often also an
increased gas transmission requirement.
The aim of the integrated network analysis is to determine the optimal timing of
developments of new fields to identify bottlenecks in the network and evaluate op-
tions for removal of these bottlenecks. This analysis can also be used to underpin
investment decisions, to optimize the product slate and to analyze trade-offs be-
tween, for example, energy efficiency, production, and overall recovery.
11.2 Integrated Network Analysis 327
Overall it supports the activity known as strategic or investment/business plan-

ning.
11.2
Integrated Network Analysis
To support integrated network analysis Shell Global Solutions International

has developed the strategic planning tool: Global Manufacturing and logistic
Optimization System/Network Analysis and Supply Chain Optimization System
[1].
The GMOS/NetSim model can be split into four key modules. These are the De-
mand module, which covers any demand point (Oil and Gas demand, fuel gas per
production platform, reinjection gas, and so on); the Supply module, which include
production profiles for each existing and future wellhead platform (the so-called
technical potential); the Production module describing production/processing fa-
cilities and constraints; and the Pipeline module, which ties together supply, pro-
duction, and demand to account for pipeline constraints and capacities.
The focus of the study is on identifying the main (future) bottlenecks of the net-
work and evaluating investment options to resolve the bottlenecks. Therefore, for
these types of studies, the planning horizon typically covers 15–30 years and the
size of the smallest time bucket is normally set at a yearly interval.
This means that the production profiles and facilities and pipeline capabilities
are taken as yearly averages.
The model, as illustrated in Fig. 11.1, requires general data input including sup-
ply sources (wellhead platforms), production and processing facilities infrastruc-
ture and capacities (processing platforms, compressors, pumps and so on), trans-
portation capacity (pipelines), costs of processing (operating cost and fuel gas con-
sumption), and other business constraints in production, processing, and distrib-
ution (e.g., minimum and maximum demands, no venting after a specific year).
As described above, GMOS/NetSim will optimize the entire network over the
entire time horizon. The objective function is based on the variables in the system
and is typically the net present value of the profit margin over the time horizon.
The profit margin is, in general, the revenues from oil and gas sales minus fixed
and variable operating costs (including cost of venting) minus investment costs.
This means that the model will only drive toward producing oil and gas if the
value of the oil and gas exported is greater than all associated costs (operating,
fixed, and capital expenditures). Because the production from each wellhead plat-
form may have different oil, gas, and water contents, the model will maximize
production from certain wellhead platforms. For example, when the bottleneck oc-
curs only in the oil pipelines and the driver is the oil price, the model will try to
maximize oil production from those wellhead platforms that have a relative low
water content. If there are also gas export restrictions in the network, the model
will try to maximize the wellhead platforms that have a relative low water content
Feed Transport Process Product Transport Market
Optimal Gas Utilisation

Wellhead Gas & Oil
Processing Unit /
Platforms Gathering Nodes Trunk lines
Flowlines
Products
Max Crude production
Environmental considerations
Gaslift Lines
Fig. 11.1 GMOS/NetSim schematic overview.
and a low gas–oil ratio (GOR), because it does not want to produce the gas with the
oil as the gas evacuation system might constrain the oil production.
11.3
High-Level Network Model
As the main purpose of the model is to identify bottlenecks (capacity) at produc-

tion/processing level, it is important to model the facilities in sufficient detail, but
not necessary the detailed individual facilities itself. To simulate and define the op-
erating conditions of production facilities and pipelines (pressures, temperatures,
and so on), GMOS/NetSim can use the output of flowsheeting tools like HYSYS
and pipeline simulators like PIPESIM as input to the model.
Section 11.3.1 discusses the level of detail per facility type and the interaction
with the more detailed processing and pipeline models.
11.3.1
Supply Module
Oil and gas production from reservoirs forms the supply part of the network model.
As an example the supply point can be represented by the wellhead platforms
where multiple multiphase streams combine (compared to streams from individ-
ual wells).
For wellhead platforms with multiple headers that operate at different pressure
levels (high high pressure (HHP), high pressure (HP), or low pressure (LP)), the
wellhead platform will be modeled as an individual wellhead platform per pres-
sure level as shown in Fig. 11.2(a) for a single header and Fig. 11.2(b) for multiple
headers at different pressure levels.
a) Separator Drum
HP TOTAL HP EX SEPARATOR
LP TOTAL LP EX SEPARATOR
VENT TOTAL VENT EX SURGE DRUM
WHP_HP
OIL/H2O TOTAL OIL/H2O EX SEPARATOR
b) Separator Drum
HHP TOTAL HHP EX SEPARATOR

HP
TOTAL HP EX SEPARATOR
LP
VENT
HHP OIL/H2O
TOTAL LP EX SEPARATOR
WHP_HHP
HP
HP LP
WHP_HP VENT
WHP
OIL/H2O
LP
TOTAL VENT EX SURGE DRUM
WHP_LP
LP
TOTAL OIL/H2O EX SEPARATOR
VENT
OIL/H2O
Fig. 11.2 (a) Wellhead platform – single header. (b) Wellhead

platform – multiple headers. * A yellow triangle represents the
“wellhead platform” for which a production profile needs to be
provided.
11.3.2
Production Module
The production module consists of the production/processing facilities. These can

be separator drums, where the multiphase flow is separated into individual gas
and liquid streams, and compressors and pumps, which allow the flow of gas and
liquid through the pipeline system. This can also be dehydration installations or
conversion units, these latter two production units are not described in this chapter
but can be added to the model in a similar way as the separator.
11.3.2.1 Separators
To determine the optimal production profile of the wellhead platforms, GMOS/
NetSim will fluctuate the production from the different wellhead platforms inde-
pendently. This means that the composition of the multiphase flow entering the
separators is different for each year and every run.
GMOS/NetSim cannot simulate the separator operating under these actual op-
erating conditions and compositions, therefore HYSYS is required to model the
separation of the multiphase flow. HYSYS and GMOS/NetSim are not integrated
applications; this means that the split in HHP gas, HP gas, LP gas, surge drum gas,
HHP gas
HHP Vessel HP gas

2 Phase flow
HP Vessel LP gas
LP vessel Vent gas
Surge vessel
Oil
Water
Fig. 11.3 Separator vessels.
oil, and water, the so-called “yield”, for each wellhead platform needs to be provided
as input into GMOS/NetSim upfront, such that the integrated network model can
calculate the exact amount of each grade1) of products through adding the yields of
the individual wellhead platforms during the optimization.
The above information implies that in GMOS/NetSim it is modeled as if every
single wellhead platform has its own separator train. In this way, it is possible to
run the integrated network model independently for all different combinations of
wellhead platform.
Since the HYSYS model provides the gas and oil/water flow split in the HHP
vessel, HP vessel, LP vessel, and surge vessel, it is not required to model the in-
dividual separator vessels/drums (SD) as an independent piece of equipment in
GMOS/NetSim. The HHP, HP, LP, and surge vessels are modeled as one single
vessel with several flow streams, such as HHP gas, HP gas, LP gas, surge drum
gas, oil, and water streams leaving the separator (Fig. 11.3). This flow split is cal-
culated in HYSYS, based on (actual) fixed operating conditions for the separators
(pressures and temperatures) for each individual wellhead platform, and converted
into the required input format for GMOS/NetSim.
The error caused by assuming the separation performance to be independent
from throughput is not relevant for this study considering that it is a strategic study,
and fluctuations that cause the loss of separation performance are more as a result
of daily/hourly fluctuations.
Interaction Between GMOS/NetSim and HYSYS If the optimized result of the

GMOS/NetSim model indicates that there is room for adjusting the assumed
“fixed” operating conditions of the separator drums, a new run in HYSYS needs
to be done, reflecting these new operating conditions, to provide new input data
(“yield” tables) for the GMOS/NetSim model to verify the improvement.
1) A “grade” in GMOS/NetSim is a product that flows through the network.

LP
LPGas
Gasexex
SDSD
HP Gas ex SD
Logical Block LP Logical Block HP

EXAMPLE
EXAMPLE11 compressor
comp compressor
LP
LPGas
GasAAex
exSD
SD
HP Gas ex SD
LP
LPGas
GasBB
exex
SDSD
2 Logical Blocks LP Logical Block

Bl HP
EXAMPLE
EXAMPLE22 compressor
com compressor
compress
Fig. 11.4 Logical compressor blocks.
A total throughput capacity of the separators is modeled; the individual con-

straints on the separators (HHP, HP, LP, or surge) are not. These constraints will
need to be verified in the full HYSYS model that is available for the separation sys-
tem. This so-called “feedback” loop is critical to ensure the quality for the overall
study result.
11.3.2.2 Compressors
The individual stages of the compressors are simulated in detail in HYSYS with the
actual operating conditions and average compositions. The output from HYSYS is
used as input data to GMOS/NetSim. However, not every single stage or compres-
sor needs to be modeled in GMOS/NetSim. For example, if there are five compres-
sors running in parallel without an intermediate stage, one compressor block will
be modeled, representing the total capacity of the five.
If, however, HP gas can bypass the first stage and only “uses” the capacity of the
second stage, the first and second stage will need to be modeled as two separate
compressors in GMOS/NetSim. Also, if the compressors are not really in parallel
and a particular compressor can only process a specific flow, the logical block will
need to be modeled accordingly. This is illustrated in Fig. 11.4.
11.3.2.3 Pumps
Similar to the compressors, the pumps are also modeled as logical blocks. If, for
example, there are two pumps running and one pump is on standby, but in practice
it is possible to run all three pumps in parallel, in GMOS/NetSim one pump with
the capacity of the three pumps will be modeled.
The capacities for the pumps are simulated in HYSYS with the actual operating
conditions and the actual pump curves are used as input data to GMOS/NetSim, to
ensure that the correct pressure profile can be calculated.
11.3.3
Pipelines Module
Different types of pipelines are modeled: flowlines (multiphase flow), oil and gas
trunk lines, and gas lift lines. Only pipelines between platforms and from platform
to export nodes are modeled; piping on production platforms does not need to be
modeled in GMOS/NetSim, since the pressure drop is small compared to the pres-
sure drop in the pipelines between platforms. However, when they are constraints
in practice they can be modeled.
GMOS/NetSim requires the maximum capacity of pipelines as input to the
model. These capacities are dependent on, among others, diameter, length, in-
let and landing pressure, gas–liquid ratio (GLR), and water content. A number
of these items are fixed parameters but others depend in a nonlinear way on the
amount and composition of the flow. Therefore, there is a need for a correlation
in GMOS/NetSim that can be used to calculate the capacity of a specific pipeline
over time under different operating conditions. This has been implemented by the
use of piecewise linear functions (via lookup tables) for the different pipeline types.
GMOS/NetSim will have these piecewise linear functions and the inlet and landing
pressures for the pipelines as input. By using the piecewise linear modeling, the
resulting problem can stay of the type mixed integer instead of becoming nonlin-
ear. Other mixed integer aspects are the 0-1 decisions on whether a hardware piece
is used or not (fixed costs) or whether an investment is beneficial or not (capital
expenditure) [2–4].
Another integer aspect arises from the physical phenomena that a capacity of a
pipeline can have a step-change depending on the oil–water ratio. GMOS/NetSim
will calculate the actual gas–liquid ratio depending on the actual production from
the wellhead platforms in that specific year and thus derive the maximum capacity
of that pipeline in that year (depending on the combination of producing wellhead
platforms, the water content of the pipeline will fluctuate).
Interaction Between GMOS/NetSim and PIPESIM As the maximum capacities of

pipelines are also a function of inlet and landing pressures, GMOS/NetSim will
track pressures throughout the network to determine the actual pressure at points
in the network based on the actual (optimized) flow. Pressure tracking is based
on PIPESIM lookup tables and the optimized GMOS/NetSim pipeline utilization.
The calculated landing pressures, assuming a known operating pressure of a well-
head platform and/or separator, will be compared with the assumed minimum
required landing pressure. In case the calculated pressure is outside these limits,
further study is required to understand if the landing pressure can be lowered at
11.4 Workflow 333
Vent
WHP DA
Gas lift
Re-injection gas
Re-injection
Compressor
Vent
DA
HHP Exp Compressor
HP Gas Export
WHP Separator Drum LP

DA
Fuel gas
Water Oil LP
Compressor
Oil
To Oil export terminal
Fig. 11.5 Model overview – production platform example.
that platform or if additional pump/compressor capacity is required to prevent the

low landing pressure. In the case the lower landing pressure is acceptable, a re-
run of PIPESIM can be required to update the lookup tables to reflect the lower
pressure, which is relevant for the connecting pipeline.
Figure 11.5 is a schematic overview of a production platform in a GMOS/NetSim
model, with the items as described in the previous sections.
11.3.4
Demand Module
In the demand module, the product demand can be specified. Product can be oil
and gas but can also be end consumer demand (e.g., power or derived oil prod-
uct demand). This demand can be specified as a fixed, minimum, or maximum
demand (or a combination). The pricing of demand can be fixed or consist of dif-
ferent tiers to indicate different markets and market behaviors.
In the demand module also, environmental constraints can be added by intro-
ducing a maximum demand on the gas vented or CO2 produced. Fuel to operate
the production facilities can also be introduced as a fixed or variable demand by
throughput.
11.4
Workflow
An integrated network analysis study normally consists of the following steps:

• functional specification
• data gathering and validation
• network model development (GMOS/NetSim)
• process (HYSYS) and fluid transportation modeling and

analysis (PIPESIM)
• model validation based on reference year (single period)
• integrated supply chain analysis using multi-period network
model
• scenario runs and idea generation
• sensitivity analysis for selected scenario(s)
• final investment analysis and recommendation
The first step – the functional specification – is to define the level of detail to be
modeled. It needs to be a good balance between the complexity of the model to
represent the actual network and the simplicity to identify main drivers and con-
straints in the model over the time horizon defined and will be determined by the
overall objective of the study (e.g., time frame, number of projects).
The next step is the data gathering process. In the functional specifications, the
boundaries of the study and the resulting data requirements are defined, and the
objective of the data gathering is to get, besides the design data, the actual op-
erating data of equipment and pipelines as well as the technical capability of the
reservoirs. Also information on deterioration of pipelines or unavailability of equip-
ment/pipelines is crucial to acquire. In some cases, performance testing will need
to be performed, as accurate data is not readily available.
At the start of the data gathering validation years will need to be chosen, which
are used to validate the GMOS/NetSim model and the HYSYS and PIPESIM corre-
lations. The validation of the GMOS/NetSim model will be done for a limited num-
ber of historical years and has a purpose to validate the configuration of the model,
to validate the operating conditions of facilities and pipelines (are constraints seen
in the model also a constraint in reality), and to get the initial understanding of the
key drivers in the model.
After approval of the validation model, the building of the multi-period model
starts and scenarios and sensitivities will be defined.
Scenarios are used to test different strategic directions of the company (e.g., dif-
ferent projects included, focus on gas production versus oil production, no vent
/flare policy). Sensitivities (e.g., prices, start-up dates) are typically evaluated to test
the severity of a certain bottleneck.
GMOS/NetSim results will show the objective function, the production of the in-
dividual wellhead platforms, the utilization of the different facilities, and the con-
straints in the model. GMOS/NetSim has several output pages showing the results.
One of these pages is the constraints analyzer. This constraints analyzer indicates
which items are limiting the objective function of the solution. The analyzer is
split up between, for example, constraints in processing facilities (“process lines”
constraints) and constraints in pipelines (“mode of transport” constraints).
The constraints analyzer will show a marginal value for each constraint for each
year. This marginal value is an indication of the value gained if the constraint is
removed (e.g., pipeline capacity increase). To identify the actual value of removing
this constraint, the model has to be rerun with the constraint removed. It could
11.5 Results 335
be well that by removing this constraint another constraint becomes active and
prevents the objective function/margin from increasing further.
The constraints analyzer is used to help the team to quickly identify the major
constraints for each year. A comparison of the constraints is made between the
different cases and scenarios to identify the commonalities and thus deriving a
list of constraints that impact the solution the most. The next step is to find a
way to remove the identified constraint, which is done through expert reviews and
brainstorming sessions. The ideas generated in these sessions will be incorporated
in the model and a new GMOS/NetSim run will be performed to verify if removing
this constraint in that particular way actually improves the objective function and
if so to which extend. The output of this run will show whether an idea solves the
bottleneck and also if it is economic to invest in eliminating the bottleneck. From
the results of the new run, new bottlenecks will appear which can be solved further.
Through analysis of the constraints that are highlighted by GMOS/NetSim in the
different runs, the project team can identify and resolve bottlenecks to further im-
prove the margin and to bring the actual optimal production closer to the technical
potential of the wellhead platforms.
Since it is a multi-period model it will indicate the maximum achievable produc-
tion profile per wellhead platform considering the constraints over the entire time
period. It will also indicate the level of utilization of all equipment and pipelines
modeled such that debottlenecking of the network can be done more easily.
It also has the possibility to make decisions on investment options that are pre-
defined by the team. If such option is economic to invest, the model will select that
option.
The GMOS/NetSim model will support the study by showing the impact of in-
dividual decisions and/or projects on the entire network and by calculating how
much oil and gas can be produced in an economic way.
In all steps in the process, it is mandatory to have intense customer involvement
as inputs should reflect actual operation and a good interaction between the dif-
ferent team members is important to define the possibilities for debottlenecking
based on the output constraints as indicated by the model.
11.5
Results
As indicated before the resulting problem is of the type mixed integer (MIP). These
type of problems can be solved via branch and bound techniques to a proven opti-
mum. For a network of some hundred and forty pipelines, fifty wellhead platforms,
and twenty production platforms over a time horizon of, for example, fifteen years,
this results in a MIP problem of a size of hundred and forty thousand rows, hun-
dred and fifteen thousand variables, of which some fifteen thousand are of the type
integer, and three hundred and forty thousand nonzero matrix elements.
The potential or the results of the study are typically expressed in oil and/or gas
that can be accelerated through the debottlenecking exercise.
Base Case Tech. Potential Optimised Case

140
130
120
110
100
kbbls/day
90
80
70
60
50
40
30
20
2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Fig. 11.6 Example of oil production acceleration.
Figure 11.6 shows an example of the benefits that are typically achieved. Approx-
imately 10% of the 15 years oil production can be accelerated if the debottlenecking
options will be implemented. These options vary from extra processing facilities,
via small piping changes to bypassing compressors and swapping lines in direc-
tion or by type of flow going through the pipeline. The thick line indicates the
maximum of oil that can be produced if all wellhead platforms are producing at
their technical potential. The bars indicate the oil production before the debottle-
necking exercise and the thin line indicates the oil acceleration that can be achieved
having implemented all viable debottlenecking options.
Maximizing oil production is not always the only objective; increasing gas pro-
duction (or recovering gas from the vent and flare) can also be economic. This
is shown in Fig. 11.7, where the gas export and vent gas is shown for the base
case (the situation before the debottlenecking exercise) and the optimized case (all
viable debottlenecking options implemented). The overall production of gas is in-
creased through additional compression capacity and also the amount of venting
is reduced through the implementation of debottlenecking options, both leading
to additional gas for export. Overall some 5% of the total of 15 years gas export
volume is recovered through implementation of the debottlenecking options.
11.6
Conclusions
Recent studies confirm that integrated network analysis helps to bring production
closer to the technical potential of the wellhead platforms. It will help to identify the
impact that all changes together have on the performance of the network. The net-
work can consist of only the offshore facilities and pipelines, but can also include
onshore processing facilities.
Optimised Case Base Case Vent-Base Vent-Optimised

MMSCFD
2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Fig. 11.7 Example of acceleration of gas production.
The order of magnitudes for benefits, which is achieved in an example, is as

follows:
>10% of 15 years oil production accelerated
>5% of 15 years gas production recovered
As these studies focus on identifying actual bottlenecks and future bottlenecks,

accurate representation of the network is crucial. This means that good data is key
to success and that design data of equipment and pipelines alone is not sufficient. If
actual performance data is not available on site, performance testing prior/during
these types of studies will be required. It is crucial that the client is a member of
the study team.
Integrated network analysis will look at the network in an integrated matter and
will evaluate the impact of all of the individual changes and projects on the en-
tire network. It is a combination of an economic optimization within the physical
boundaries determined by facilities and pipelines and also by strategic directions
or operating policies.
This approach will help visualizing the expected bottlenecks in the system and
identifying solutions to these bottlenecks. Especially with emissions becoming
more stringent, the study will demonstrate the impact on the production capability
and helps identifying the best solution.
References
1 Lasschuit, W., Thijssen, N., Sup- 3 Grossmann, I. E., Review of non-

porting supply chain planning and linear mixed-integer and disjunctive
scheduling decisions in the oil and programming techniques for process
chemical industry. Computers and systems, engineering. Optimization
Chemical Engineering, 28 (2004), Else- and Engineering, 3 (2002), Springer,
vier, pp. 863–870. Netherlands, pp. 227–252.
2 Floudas, C. A., Nonlinear and Mixed- 4 Kallrath, J., Mixed integer optimiza-
Integer Optimization: Fundamentals tion in the chemical process indus-
and Applications. Oxford University try: Experience, potential and future.
Press, 1995. Trans. I. Chem. E. 78(Part A) (2000),
pp. 809–822.
339
Index
a assembled-to-order (ATO) 257

advanced planning systems (APS) 219ff. asset
agent – management 276
– central 70 – value 274
– emulation 52
– factory 53f. b
– inventory 56ff. back orders 262
– production 58ff. batch chemical processes 2
– purchasing 59f. biodiesel 93f.
– real 51f. biofuels 93
– real central 51 bonds 263, 272, 275, 281, 291
– sales 53ff. Boolean variables 92, 97
– simulated 51f., 71 borrows 6
– site 53 Branch&Bound (B&B) 301
– storage 53f. Branch&Cut (B&C) 304f., 308
– subagent 52ff. budgeting 4
– transportation 61 bullwhip effect 108
approach business solvency 2
– bound 301
– Branch & Bound (B&B) 301 c
– combined 228 campaigns 165, 175, 180
– combined strategic and operative capacity 12f.
(CSOP) 222, 230f., 233, 236f., 239 – addition 266f., 274, 278
– distribution-based 29 – adjustment 233
– hierarchical 221 – buffer 253
– holistic 38 – change 254f., 265
– hybrid 302, 304, 308, 310 – coefficients 16
– ideal infinate optimization 33 – consumption 17
– impact assessment 78f. – equipment 15
– integrated tailor-made 230 – expansion 254f., 258, 267
– metaheuristic 301f., 304, 306ff. – increment 22ff.
– multiobjective 70 – location 12, 31
– primal-dual heuristic 159 – long-term 160, 252f.
– product centric 89 – management 245ff.
– production system centric 89 – distribution centers 165f.
– scenario-based 29 – plan 253
– transportation system centric 89 – production 161, 164, 169, 175, 188,
– two-stage stochastic optimization 229, 259
29, 32 – profile 282, 284, 289
ISBN: 978-3-527-31693-9
340 Index
– short-term 259 – network structure 163

– storage 72, 228 – nonnegative 125, 140
– strategy 253 – operational planning 117, 228
– timing 253ff. – production 117, 120, 164, 203, 228
– total 253 – returns 202
– utilization 264, 267 – sale 117
capital – sequence 119
– cost 8f. – single source 162f.
– investment 25ff. – transportation flows 205
– structure ratios 8 – upper bound 117, 121, 164f.
carbon capture and sequestering systems – warehouse storage 204
(CCS) 96ff. consumption
– oxygen-combustion 96 – energy 3
– postcombustion 96 – material 3
– precombustion 96 continuous flexible process network (CFPN)
cash 302ff.
– balance 5 contracts 247, 259, 261, 263
– exogeneous 5f. – supply 271f.
– flow 1f., 4, 8, 221 contribution margin 229
– inflow 5 corporate value (CV) 8, 21ff.
– minimum 6, 8 cost 272
– outflow 5, 20 – analysis 182
– short-term management 67 – bottom line 157
changeovers 165f., 304 – break-down 150, 180
channel – capital 166
– coordination 187 – collection 141
– decision 187 – duty 167f.
– reverse 190 – establishment 170f.
chemical flow 122f. – fixed 10, 18f., 141, 145, 159, 166,
chemical process industry (CPI) 1, 3f., 45, 220, 273
50, 62, 71 – global 182
CIF (cost, insurance, freight) 263 – handling 167f., 171, 182
cleanup 3 – infrastructural 162, 166, 168, 170f.,
collection plan 212f. 182
communication – inventory 15, 72, 119, 121, 273
– horizontal 250 – investment 16f., 220, 238, 280
– vertical 250 – labor 170f.
compressors 331 – minimizing 98, 135, 141, 162
constraints – one-off 166
– allocation 162 – operating 96, 162, 166, 252
– assignment 119, 121 – processing 141
– binary logical 140 – production 10, 21, 72, 157, 166ff.
– capacity 140 – rental 170f.
– cost 121 – shipping 141
– cyclic scheduling 118 – startup 96
– demand 202 – storage 141, 167, 206, 263
– disassembly center storage 204 – transition 119, 121
– factory-capacity 203 – transport service 10
– integer restriction 140 – transportation 19f., 72, 144, 148ff.
– logical 164 – variable 145, 159, 222, 273
– loxer bounds 165 credit line 6
– mass balance 118 creditor 5
– material flow balance 139, 205 customers 88f., 115f.
Index 341
– allocation 178 – strategic 219, 226

– assignment 180 – supply-chain 223
– cluster 207 disassembly 198
– multisite 228 – center 198ff.
– portfolio 223, 228f. discrete event simulator 46, 49f.
– returns 189 disinvestment 262, 264ff.
– zone 162ff. – factor 274
cycle time 118ff. – generic 46
cyclic scheduling 119f. – intervals 14
– limits 14, 264
d – stochastic 46
daily production capability 165 disjunctions 92
data gathering disposal
– process 334 – fraction 202, 208
– validation 334 – options 189ff.
decisions – structures 197
– continuous 92 distribution
– discrete 92 – batch sizes 158
– financial 2, 4, 8 – centers (DCs) 11f., 15ff.
– maximum profit 2 – facilities 158, 161
– minimum cost 2 – flow 198
– planning 2 – network 161, 168
– process operation 2 – problem 159
– production-distribution 2 – resource 160
– scheduling 2 – supply chain network 174
– strategic 3, 28, 38, 102 – system 159f.
– support tools 4, 34 duty 167
– tactical 4 – coefficient 178
– variables 2 – drawbacks 259
– wait-and-see 30 – structure 177
decomposition
– bilevel 302 e
– Lagrangean 301 eco-indicator 3
– method 302, 304 economic
– technique 301 – life 8f.
demand 171 – performance 78ff.
– allocation 220 economies of scale 247
– aspects 219 electricity purchase 96
– customer 172, 180 emission 62, 73ff.
– efficiently seasonality 159 – CO2 87f., 97, 99ff.
– market 164 – SOx 91, 93ff.
– product 18, 176f. energy consumption 91
– profiles 264 Energy Information Administration (EIA)
depletion 164 91
depreciation 5, 10, 20, 264, 274 energy production system 92f., 97f.
design – flowsheet 92
– financial 219, 223 engineered-to-order (ETO) 256
– functional 223 enterprise-wide optimization (EWO) 108
– modules 238 environmental
– multisite 219, 222, 233 – burdens 76f., 78
– mutual 219 – categories 64
– optimization 224 – effects 87ff.
– software 233 – impact 63, 78f., 89, 159, 198, 213
342 Index
– indexes 74 incomes 272

– performance 62, 79, 88f., 100 industry
environmental supply chain management – chemical 219, 221f., 247, 249
(ESCM) 62 – food 219, 222
environmentally – metals 219
– conscious production 89, 91 – pharmaceutical 219, 222
– harmful materials 88f. – process 219, 221
equipment 18 inflation rate 280
equity 9, 20 initial debt 17
equivalent deterministic formulation (EDF) in-sourcing 219
226 integrated network analysis 325ff.
Euclidean distance 99 intermediates 245, 262, 281
execution 21 internal rate of return (IRR) 8
extended producer responsibility (EPR) inventories 17, 77
189f., 194 – control strategies 49
– facility 248, 250, 261, 268ff.
f – holding facilities 245, 265
facility 12ff. – level 72, 80, 121
– location 12, 31, 159, 223 – negative, see back orders
finances 1, 4f. – parameters 72, 79
financial – profiles 82
– flows 88, 250 – table 76
– formulations 4 investment 8, 18f., 251, 262, 264, 267ff.
– variables 4 – limits 264
flow-shop 248 – net 10
forecasting technique 168
fossil fuels 91f., 96f. j
free on board (FOB) 263 job-shop 248
functional specification 334
k
g key performance indicators (KPIs) 2, 4, 24,
Gantt chart 27 68, 136, 182
– financial 27f., 36f. – myopic 2, 24
gas/liquid ratio (GLR) 332
gas production acceleration 337 l
Global Manufacturing and logistic Optimiza- labor
tion System (GMOS) 327ff. – tasks 20
global warming potential (GWP) 78 – utilities 19
globalization 249, 259 lead time 122ff.
greenhouse effect (GHE) 79ff. legislation 5
greenhouse gases (GHG) 89, 91ff. leverage 8
lifecycle assessment (LCA) 62ff.
h – assessment 67
holdup 269 – framework 63, 64, 65
hybrid-tabu – inventory analysis phase 66
– branch 301 – methodology 63f., 65, 78
– search procedure 305 linear programming computation methods
4
i liquidity ratios 8
import-export loan 205, 272, 275, 281, 291
– limits 264 – installment 205, 272
– tariffs 264 location
inbound flow 203f. – allocation 195
Index 343
– analysis 160 models

– distribution center 168, 173 – agent 53
– geographical 173 – bicriterion optimization 107
logistic – boiler 93f.
– global 159 – budgeting 4f., 11, 24
– infrastructure 89 – capacity expansion 258
– network design 195 – capital asset pricing (CAPM) 9
– point-to-point 148 – cash management 4
– recycle 89 – close-loop supply chain 195
– reuse 89 – continuous flexible process network
– reverse 89 (CFPN) 303f., 313f.
long-term credit line 5, 20 – demand 49
long-term debt 8f., 20 – design-planning 23, 201
– deterministic analytical 192
m – disjunctions 92
make-to-order (MTO) 256f. – dynamic facility location 258
make-to-stock (MTS) 256f. – economic 192
manufacturing – facility location 159
– discrete 161 – favor individual tailor-made 219
– facilities 136, 249 – financial 2, 5
– plants 73f. – fuel tank 94
– process 76, 161 – global supply chain 159
– sites 11f., 16, 22ff. – GMOS/NetSim
marketable securities 5f., 11, 17, 20, 78 – holistic 1, 4, 38
market-to-book value 8 – inhouse 238f.
market value 8f. – integrated 4f., 28, 38, 330
mass – inventory 49
– balance 117, 121 – location-allocation 195, 199, 205
– conservation 96 – mathematical programming 157,
– fractions 17 160, 169
material 164 – multiperiod mixed-integer nonlinear
– availability 246ff. programming (MINLP) 20, 107ff.
– balance 76, 96, 117, 264, 269, 278 – Nash-type 160
– flow 180f., 222 – nonmathematical 252
– handling capacity 179 – operational planning 117, 219, 222
– movement 245 – optimization 2
– recycle 245, 261 – planning-scheduling 238
– supply 246f., 259 – process network 301f.
mathematical formulations 29, 163, 172 – recourse 29
maturing period 20f. – reverse logistic 137
maturity 6, 9, 272 – simulation 46ff.
maximum profit network design 22 – strategic-operative planning (SOP)
merger 247 222, 230, 233
method – strategic planning 161
– e-constraint 21, 35, 98 – stochastic programming 29ff.
– discounted-free-cash-flow (DFCF) – stochastical analytical 192
8f. – turbine 94
– mass allocation 76 – two-stage stochastic 35, 38
– myopic 45 – uncertain 29
– probabilistic 29 – warehouse location 193
– weighted average 9 modularity 196
minimum costs 2 modules 52
mixed integer programming 159 – demand 327, 333
344 Index
– environmental 62, 65, 74, 78f. – inhouse 239

– financial 68, 79 – portfolio 229
– pipeline 327, 332 – robust 226ff.
– production 327, 329 – stochastic 226
– supply 327f. Order-Point 49
Monte Carlo sampling 30, 36, 68f., 78, 83 Order-Quantity 49
multiagent Order-Up-To-Level 49
– framework 45, 50ff. Original Equipment Manufacturer (OEM)
– system 50, 52, 58ff. 193
multinational corporation (MNC) 260f., outbound flow 203f., 213
292 outsourcing 164, 179f., 219, 228
multiobjective (MO) problem 33, 35, 78,
91f., 98 p
– mixed integer linear 35 packaging material 88f.
multiphase flow stream 326 Pareto
multiproduct 159, 161, 195 – curves 21ff.
multistage stochastic programming 29 – front 80ff.
– optimal 126f.
n – solutions 35ff.
nationalization 281ff. participation rate 144, 148
net present value (NPV) 8, 20ff. payback time 8
network performance
– Analysis and Supply Chain Optimiza- – indicator 38, 196
tion System (NetSim) 327ff. – measures 3
– balance 205 – superior economic 38
– configuration 24f., 28, 30, 179ff. pipeline 326ff.
– design 23f., 37, 196 – simulation program 326, 328
– distribution 158 planning
– flow 175 – budget 221
– four-echelon 194 – capacity 223, 253
– global 161, 220 – design 220f., 228f.
– multisite 222 – facility 253
– optimization 178f., 325 – financial 221
– performance 183, 194 – global 220
– redesign 168 – long-range chemical process 221
– regional 161 – long-term 246
– reverse logistic 136, 194f. – master 219ff.
– sensitivity 182 – mid-term 219
– structure 114f., 122, 127ff., 182, – multisite design 220
196, 200, 220 – operative 220ff.
– superstructure 174 – producing plants 219
– two-echelon 193f. – production 219ff.
nodes 72, 87, 90, 92, 100, 164 – sales 229
– short-term 259
o – strategic 220ff.
objective function 31, 38, 139, 147f., 166f., – supply chain 220, 222, 234
196, 206, 229, 304, 307f., 313ff. – systems 219, 250
– overall 167 – tactical 222, 226, 246, 252, 256, 258
oil production acceleration 336 – tax 258
open database connectivity (ODBC) 232 plant 219
operations management 250 – capacity 117, 160
optimization – electroplating 72f.
– approaches 226f. – equipment 15, 17, 21
Index 345
– input 115f. – upgraded 274

– location 161 production 87f.
– manufactoring 164 – allocation 174f., 178, 180, 259, 262
– multiproduct 110, 114, 118, 121ff., – centers 102
259 – design of 161
– multipurpose 161, 182 – facilities 245f., 252f., 261f., 265,
– multisite batch 220 268ff.
– output 115f. – gas 325
– production 162ff. – just-in-time 203
– site 114f., 127 – mass 135
pledging 5f., 10, 24 – multiechelon 161
– costs 11 – multisite 165, 219, 223
Poisson process 49 – network 111
potential locations 16 – oil 325
price – platform 333
– discounted 273f. – pre-allocation 180
– nonlinear 160 – problem 159
probabilistic optimization problem (POP) – process 75
226 – pull method of 256
probability – push method of 256
– curves 37 – rate-off 164ff.
– functions 36 – shut-down 166, 262
process – site 165, 168, 228
– collection 191 – starting 262
– inspection 190 – strategy 256
– operation 1, 4f., 221 – systems
– reverse 191 – unit 72
– sorting 190f. profit 1ff.
– supply chain network (PSCN) – after-tax 223, 229
107ff. – histogram 32
process systems engineering (PSE) 1, 8, 45, – margins 1, 327
62, 68 – maximum 21f., 25, 37, 89
processing – Pareto curves 22f.
– rate 120 – potential 223
– time 120f., 124 – requirements 160
product – total 22ff.
– allocation 159 – worst-case 35f.
– complex 150 profitability 107
– design 196 pumps 331
– disassembly 90 purchases 13
– end-of-life 191, 194 – external 13ff.
– end-of-use 188ff. – prices 14
– flow 136, 190, 248
– life cycle 90, 190f. q
– nonupgraded 274 quality function deployment 252
– packaging 191 quantitative tools 4
– portfolio 160, 165, 223, 229 quantity discounts 14
– prices 72, 90
– quality 191 r
– recovery 135f., 190 rate of return on investment (RORI) 276
– remanufacture ratio of throughput 262
– storage 169 raw materials 2, 6f., 11ff.
– take-back 190 – inventory 92
346 Index
– storage 169 – investment 9

receivables 5f. – management 31f.
recovery – metric 35
– energy 189f. – minimization 32f.
– facilities 136 – premium 9
– large-scale 150 robustness 32
– processes 137
– system 136 s
recyclability 196 sales
recycling 136, 150, 189f., 248 – location 8, 263
– carpet 137f. scheduling 220ff.
– curb-side 138 – medium-term 230
– facility 138 – short-time 223, 230
– material 189 sector
– network 137 – commercial 1f.
– process 143 – industrial 1f.
– product 194 – residential 1f.
– system 150 – transportation 1f.
refurbishing 90 separator vessels 330
relocation 281 shareholder value (SHV) 2, 8, 68, 249ff.
reparability 196 short-term credit line 5, 8, 17
repayments 6 short-term debt 9
replenishment quantity 71, 79 short-term financing source 6
request-for-material (RFM) 47, 56f. simulation-based optimization 69, 83
request-for-quote (RFQ) 56ff. software
resource – commercial 234, 236ff.
– allocation 159 – individual 236f.
– recovery 136 – planning 235
responsiveness 107ff. – provider 239
retail 90 – refinery scheduling 235
– outlets 90 – scheduling 235
– site 138 solution
retailers 73f., 88ff. – full-scale 316
retrofitting problem 15 – integer 315
return – paths 316
– customer 189 – relaxed 314
– distribution 189 – restricted fractional 308f., 314
– manufacturing 189 – tabu-search 316
– on equity 8 – unrestricted fractional 308f., 314
reuse 136, 188, 194 source, see distribution centers
reverse specific applicable rules 5, 10
– distribution 191 State-Task Network (STN) 193
– flow 188 steam
– logistic systems 187ff. – high pressure 93ff.
– production system 135, 137 – low pressure 93f.
Reverse Logistics Executive Council (RLEC) – middle pressure 93f.
191 stock
risk 1, 160 – accumulation 164
– averse decision maker 32 – adjustment 189
– curve 32f. – security ratios 8
– downside 32ff. strategy
– financial 32ff. – lag 253f.
– free rate 9 – lead 253f.
Index 347
– metaheuristic 302 throughput 262f., 268, 330f.

– tabu search 302 – maximum 280
– track 253ff. – minimum 280
subassemblies 203f. time
suppliers 73f., 88, 90 – delay 124
supply-chain (SC) 1ff. – horizon 199f., 209f., 230, 240
– chemical 1, 29, 247ff. – macro 200ff.
– close-loop (CLSC) 89f., 137, 187ff. – micro 200ff.
– configurations 22, 27, 37 – modelling 200
– end-to-end 260 – period 120f.
– energy 91 – resolution 230
– forward 188, 190ff. – scale 199f., 207f.
– gas upstream 326 top-down 221
– generic unit 47, 68 total rate of flow 164
– infrastructure 139 trade-off 118f., 128, 150, 326
– inventory 157 traditional flow 188
– large-scale 301 transfer prices 159
– management (SCM) 1ff. – intercompany 160
– multienterprise 160 transportation
– network 11, 21, 24f., 45, 50f., 71f., – air 100f.
87, 108, 110ff. – cost 102
– oil upstream 325 – flow 164, 169, 205
– operations 11, 173 – intermodal 101ff.
– optimization 1, 45, 110, 194, 249 – intersite 115
– partnership 160 – pipeline 100f.
– performance 158 – rail 100f.
– pharmaceutical 160 – road 100f.
– real-world 51f., 69, 110, 160 – sea 100f.
– redesign 245, 247, 250, 260, 265, – single-mode 101f.
281ff. – sustainable 89, 100
– representation 162 – system 100ff.
– reverse 136, 190ff. – time 160
– strategy 109 turnover ratios 8
– structure 11, 15f., 117, 190, 223
– system 87, 89 u
– traditional 190 uncertainty 28ff.
– analysis 198
t – demand 28, 30, 260
tabu search – industry 260
– algorithm 302, 312 – SCM 28
– based hybrid approach 311 unified modeling language (UML) 70
– framework 305 unitary
– guidance 310f., 313 – production 15
– procedure 308 – transport 14
tailor-made methods utilization 287, 289f.
– individual 236 – per site 165, 182
taxes 5, 9, 125 utopia distance 99
– net operating profit after (NOPAT)
9, 162 v
technical value accumulation 25f.
– coefficient 6, 20
– potential 327
technology upgrade 266, 268f.
348 Index
w waste
warehouse – collection 90
– capacity 168ff. – disposal facilities 89f., 188
– design 158 – generation 3, 88f., 91
– location 158, 161, 195 – industrial 137
– major 169 – packaging 190
– mission 158 – prevention 189
– multiproduct 195 workflow 333
– secondary 169 worst case 35f.
– storage 204 – maximum 37f.
– utilization 158

Process Systems Engineering Vol. 3 - Supply Chain Optimization, Part I (Wiley-VCH, 2008) PDF

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Process Systems Engineering Vol. 3 - Supply Chain Optimization, Part I (Wiley-VCH, 2008) PDF

Transféré par

Droits d'auteur :

Formats disponibles

Process Systems Engineering

Engell, S. (ed.) Ingham, J., Dunn, I. J., Heinzle, E.,

Chemical Production Chemical Engineering Dynamics

Volume 3: Supply Chain Optimization, Part I

Typesetting VTEX, Litauen

Printed in the Federal Republic of Germany

Preface – Volume 3: Supply Chain Optimization XIII

1 Bridging the Gap Between Production, Finances, and Risk in

2 Multiagent Framework for Modeling and Design of Supply Chain

3 Environmentally Conscious Supply Chain Management 87

3.4.1 Intermodal Transportation 101

4 Optimal Design and Operational Planning of Responsive Process

5 Reverse Production Systems – Optimization Modeling to Support

5.5 Case Study Data and Scenarios 143

6 Optimal Design of Supply Chain Networks using Mathematical

7 Design and Planning of Closed-Loop Supply Chains 187

7.3.1 Forward Supply Chain Models 192

8 Combining Strategic Design and Operative Planning in the Process

9 Chemical Supply Chain Redesign 245

10 A Hybrid Tabu-Branch and Bound Approach for the Solution of

10.3.2.2 Restricted Fractional Solutions 310

11 Enhancement of Oil and Gas Production via Network

Preface – Volume 3: Supply Chain Optimization

Modern industrial enterprises are typically multiproduct, multipurpose, and multi-

called advanced planning systems to academic achievements (for example, linear

London, June 2007

Jane C. Ammons Dr. Michael C. Georgiadis

Professor Josef Kallrath João M. Natali

Professor Rajagopalan Srinivasan Dr. Panagiotis Tsiakis

Table 1.1 Supply chain models for the CPI.

To make optimal ﬁnancial decisions in the context of SCM we need to properly

1.3.1.1 Budgeting Model

beginning of a period is received during that period through pledge, as stated by

1.3.1.2 Objective Functions

The Discounted-Free-Cash-Flow Method (DFCF) Our strategy applies the discounted-

CV = DFCF − NetDebtT , (1.18)

NetDebtT = CLineT + LDebtT − CashT . (1.19)

WACCt = λt · E(ROE) + irt · (1 − λt ) · (1 − trate) ∀t, (1.21)

E(ROE) = r 0 + rp. (1.22)

To compute the expected return on equity, which is denoted by E(ROE), we ap-

FCFt = Proﬁtt (1 − trate) − NetInvestt − NWCt ∀t. (1.23)

ARect = ARect − ARect−1 ∀t. (1.28)

APayt = APayt − APayt−1 ∀t. (1.30)

transactions of marketable securities.

As mentioned before, integrated models for SCM can be constructed by merging ﬁ-

1.3.2.1 Mass Balance Constraints

RAW SITES & DISTRIBUTION

Fig. 1.1 Supply chain structure taken as reference.

1.3.2.2 Capacity and Facilities Location Constraints

The external purchases of raw materials to supplier e at every period t (EPurchet )

The capabilities of the proposed approach are illustrated by solving a retroﬁtting

Fig. 1.2 Supply chain structure of the case study.

markets located in different European countries. A set of potential technologies are

Table 1.2 Speciﬁc volume of products i (υi (10−5 m3 /kg)).

Product Raw material

P1 8.00 7.20 8.40 7.60

Table 1.6 Cost (m.u./kg) and maximum availability

Availability 150.0 125.0 335.0 60.0 60.0 40.0

Fixed cost 25.0 37.0 50.0 45.0

Table 1.8 Distribution centers ﬁxed cost (FCFWwt (m.u./m3 ))

Fixed cost 17.5 15.0 15.0 13.75

m = M1, M3, M4, M5

Table 1.10 Transportation cost of product i from site s to

FCFt = Proﬁtt (1 − trate) − NetInvestt − NWCt ∀t. (1.23)

ARect = ARect − ARect−1 ∀t. (1.28)

APayt = APayt − APayt−1 ∀t. (1.30)

FR = P (Proﬁt < ), (1.63)

minimizing FR for a continuous range of proﬁt targets results in an inﬁnite

Proﬁts i − Us · Zsi ∀i, s, (1.65)

Proﬁts i + Us · (1 − Zsi ) ∀i, s, (1.66)

Zsi ∈ {0, 1} ∀i, s, (1.67)

maximise {E[Proﬁt], −FR }.

maximise {E[Proﬁt], −DR }