Académique Documents
Professionnel Documents
Culture Documents
A R T I C L E I N F O A B S T R A C T
Article history: The optimal design of water-using systems involves necessarily the exploitation of all possible water
Received 10 July 2008 reuse and recycling alternatives. The general problem can be formulated as a non-convex nonlinear
Received in revised form 3 April 2009 program (NLP), but due to the presence of bilinear terms, it may be difficult for local optimization solvers
Accepted 29 April 2009
to attain global optimal solutions. To overcome this difficulty, this paper presents two mixed integer
Available online 21 May 2009
linear programming (MILP)-based procedures to generate a few structurally different starting points for
Keywords:
the NLP. In both, the problem is decomposed into calculation stages by assuming that the water streams
Water minimization progress in series through the water-using units, with the binary variables selecting which unit belongs to
Water re-use a certain stage. Their main difference concerns the way fixed flowrate units are handled, either separately
Design or in conjunction with a fixed load operation, since the former comprise a linear subsystem. The two
Mathematical modelling algorithms are compared to a closely related LP-based method taken from the literature and to the one
Optimisation employed by the global optimization solver BARON. The results from a large set of example problems
confirm their effectiveness in avoiding local solutions despite the small number of starting points. In
contrast to the previous method they are easily scalable and, for some of the larger problems, could
find better solutions than BARON with significantly fewer computational resources. The results have also
shown that the option of tackling one unit at a time is the most favorable.
© 2009 Elsevier Ltd. All rights reserved.
Water scarcity and escalating costs of wastewater treatment (in Wang and Smith (1994) were the firsts to apply pinch technol-
order to meet stricter environmental regulations on industrial efflu- ogy to the targeting and design of WUNs with minimum freshwa-
ents) have led to a growing emphasis on both freshwater minimiza- ter consumption and wastewater generation. These targets can be
tion and wastewater generation in industry. As a consequence, the easily found following the representation of the water-using opera-
identification of water reuse and/or recycling opportunities has been tions data on a concentration vs. cumulative mass load diagram and
regarded as an innovative step in the design and optimization of wa- generation of the limiting composite curve (LCC). The pinch point is
ter systems within the two most commonly used scientific solution then obtained by increasing the slope of the water supply line until it
approaches: the insight-based pinch technology, largely supported touches the LCC. After the targeting step, a few sets of rules are used
by graphical methods, and the more systematic mathematical pro- to design networks that verify those targets. In fact, the design pro-
gramming approaches. Other design methods use genetic algorithms cedure itself can ensure that the minimum targets are met as shown
(Prakotpol and Srinophakun, 2004; Lavric et al., 2005). Besides the by Castro et al. (1999). Similar design procedures have been used by
type of method being used, targeting/design methods for water sys- Feng and Seider (2001) and, for multicontaminant systems, by Wang
tems can be classified based on the problem being handled. Ideally, a et al. (2003). The network structure features internal water mains
fully integrated water-using and treatment network (WUTN) should to which the water-using operations are connected. Internal water
be sought but due to the inherent complexity of the problem, many mains mix outlet streams from some operations and distribute them
researchers have focused on either the water-using (WUN) or the to other units. The aim is to simplify the design, operation and con-
distributed wastewater treatment (WTN) designs. trol of the network, since they serve as buffers, i.e. the whole water
system can be operated smoothly if the water quality of the internal
water mains can be controlled.
While the above methods focus on mass transfer-based water-
using operations, important work has appeared in the literature
∗ Corresponding author. Tel.: +351 210924643; fax: +351 217167016. also dealing with fixed flowrate operations. A targeting method for
E-mail address: pedro.castro@ineti.pt (P.M. Castro). single contaminant systems that is based on composite curves and
0009-2509/$ - see front matter © 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ces.2009.04.051
J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752 3737
on the concept of water surplus is given in Hallale (2002) together nonlinear constraints involve bilinear terms due to the product of
with a design method that meets the targets on freshwater use and flows by concentrations or of mass flows by split fractions, depend-
wastewater generation. Manan et al. (2004) described such graphical ing on how the problem is formulated.
procedure as tedious and presented the Water Cascade Analysis as Takama et al. (1980) were the firsts to address the simultaneous
its numerical equivalent, which is able to generate more quickly design of water usage and treatment networks and few additional
the exact utility targets and pinch location. Similar work has been features have been introduced since then (Huang et al., 1999). The
performed by Aly et al. (2005), who use the Load Problem Table. shortcoming of non-convex NLPs is that they are very difficult to ini-
A very comprehensive non-iterative targeting procedure that is tialize. This is particularly important if local optimization solvers are
nearly identical to that of El-Halwagi et al. (2003) has been pro- used, since the solution returned is very dependent on the starting
posed by Prakash and Shenoy (2005). When compared to the orig- point. To overcome this limitation, Gunaratnam et al. (2005) pro-
inal work of Wang and Smith (1994), the graphical representation posed a mixed integer nonlinear (MINLP) formulation together with
uses cumulative contaminant load vs. cumulative flowrate, instead a decomposition strategy that divides the problem into mixed inte-
of concentration vs. cumulative mass load. Furthermore, the limit- ger linear (MILP) and linear (LP) subproblems. These are then solved
ing composite curve and water supply line are replaced by source in an iterative manner to provide an initial starting point that is
and demand composite curves that are generated from the opera- then refined through the solution of the general MINLP. The basis of
tions inlet and outlet data, respectively. Feasible networks are those the main design stage comes from physical insights gained through
where the source composite is located below the demand compos- conceptual approaches and aspects of mathematical programming.
ite, with the optimal ones being those where the two meet in at least Overall, their approach is able to explore the synergies between the
one point, the pinch. This can be easily found by horizontal shift- two subproblems, is capable of solving them separately and pro-
ing of the source composite. In the process, the origin of the source vides a robust technique, although it does not necessarily lead to the
composite moves away from that of the demand composite, with global optimum.
the distance giving the amount of freshwater that is required. At the Building on this work, Alva-Argaéz et al. (2006) proposed a novel
other end of the chart, we get the total wastewater generation. Af- decomposition approach that simplifies the difficulties of the opti-
ter targeting, network design is achieved through an algorithm that mization problem, making systematic use of water-pinch insights
uses the principle of nearest neighbors, i.e. the water source streams to define successive projections in the solution space. Their generic
that are chosen to satisfy a particular water demand are the nearest methodology is applied for total water system design in petroleum
available neighbors in terms of contaminant concentration. refineries and has the potential to address retrofit problems.
Targeting methods can be applied for re-use, regeneration re-use While in the conventional superstructure the operations are sim-
and regeneration recycle. The consideration of regeneration units is ply connected to each other, it may be convenient, according to Feng
a way of incorporating some elements of wastewater treatment into and Seider (2001), to use internal water mains to simplify design
the design of the WUN and thus achieving some level of integration and control. Such hybrid structure has been used by Liu et al. (2008),
between the two problems. Agrawal and Shenoy (2006) extended who have recently proposed a MINLP that aims to reach a trade-off
the approach of Wang and Smith (1994) for fixed flowrate prob- between freshwater consumption and the controlling complexity of
lems. Targets for minimum regeneration, freshwater and wastewa- the water-using network. For a measure of the latter a control num-
ter flowrates can also be determined by the procedures of Ng et al. ber is used, which is defined as the summation of the number of op-
(2007) and Relvas et al. (2008). The ultimate goal is to achieve a wa- erations that receive water, either directly from other units or from
ter system with zero wastewater discharge, a specific case of regen- the internal water mains.
eration recycle systems with no wastewater outlet. In this respect,
Sustainability concerns are becoming increasingly more impor-
Deng et al. (2008) were the firsts to present a graphical method
tant in process design, with water networks being no exception. Lim
for single contaminant systems with fixed mass load that is able
and Park (2008) have proposed a NLP formulation to design an en-
to find the necessary conditions for zero discharge. Cormos and
vironmentally friendly WUN. The concept of life cycle assessment
Bandyopadhyay (2008) also used the concept of regeneration recy-
(LCA) is integrated into the objective function with a view to opti-
cle to significantly reduce freshwater requirements, while satisfy-
mize trade-offs among effect scores of the main environmental pol-
ing environmental discharge regulations. These authors propose a
lutants, which is in itself new and contrary to the traditional ap-
graphical representation, which together with analytical algorithms,
proach of minimizing solely the economic costs.
addresses the integrated water management in processes involving
Karuppiah and Grossman (2006) solved the design problem of an
a single treatment unit.
integrated WUTN to global optimality through a new spatial branch
Overall, the main merit of pinch analysis is that the various tar-
and contract algorithm. Piecewise linear under- and over-estimators
gets are identified ahead of the detailed network design. Neverthe-
were used to approximate the non-convex terms in the original NLP
less, the graphical technique suffers from scaling problems when the
in order to obtain a MILP relaxation whose solution provides a tight
compositions or flows of the various units and streams are vastly
lower bound at every node of the tree. Their algorithm is computa-
different, which can distort the representation. Its major disadvan-
tionally faster than the global optimization solver BARON, but is still
tage is, however, the inability to address multicontaminant problems
computationally expensive for medium sized problems. In addition,
systematically.
water-using units were assumed to have fixed flowrate demands.
While it is straightforward to extend the approach for variable wa-
1.2. Mathematical programming approaches ter intakes, which is the normal assumption, the added complexity
of potentially loose bounds for such variables might compromise the
Mathematical-based optimization techniques are the best option efficiency of the algorithm. The work has been extended (Karuppiah
for more complex water systems with multiple contaminants. A com- and Grossman, 2008) to deal with uncertainty in the model param-
mon feature is the consideration of a superstructure that includes eters, such as contaminant loads for the water-using units and re-
all network design possibilities, namely water-using and/or treat- moval ratios for the treatment units. A multi-scenario non-convex
ment units as well as splitters and mixers for the water streams. MINLP model is formulated to globally optimize the integrated wa-
The problem is generally formulated as a non-convex nonlinear ter network operating under uncertainty. The novelty lies in com-
program (NLP) or a mixed integer non-linear program (MINLP), if bining the concepts of Lagrangean and convex relaxations in order
the connections between the units need to be accounted for. The to generate strong bounds.
3738 J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752
There are also mathematical programming-based methods in the streams down to the environmental discharge limits. The two
literature that attempt to find global optimal solutions following broadly classified categories (Prakash and Shenoy, 2005) of water-
the generation of multiple starting points for the general NLP. They using units (set O) are considered, i.e. fixed contaminant load and
can be considered global search heuristics even though they can- fixed flowrate. Fixed contaminant load operations (set Ofl ) are qual-
not guarantee global optimality. The methods of Hernández-Suárez ity controlled and may be modelled as mass transfer units (e.g.
et al. (2004) and Castro et al. (2007, 2009) for WTNs, and of Teles washing, scrubbing, and extraction). Their data is often expressed
et al. (2008) for WUNs, have in common the use of a superstructure by a limiting flowrate (folim ) together with maximum inlet (co,c
in max )
decomposition approach and the solution of several LPs and NLPs. out max ) concentrations, which are related to the mass
and outlet (co,c
Hernández-Suárez et al. (2004) employ a parametric optimization exchange (mo,c ) through Eq. (1). It is assumed that inlet and outlet
approach that fixes the split fractions after the treatment units, in flowrates from a fixed load operation are equal. Set C includes the
order to overcome the bilinear terms resulting from the product of pollutants that need to be taken into account.
split fractions by flows. Although very thorough, the search strategy
may involve a prohibitively large number of problems to solve. In mo,c = folim (co,c
out max in max
− co,c ) ∀o ∈ Ofl , c∈C (1)
Teles et al. (2008), an alternative initialization strategy is proposed
On the other hand, fixed flowrate operations (Off ) are quantity con-
where the different substructures are generated from all possible se-
trolled and may include water-using units like boilers, cooling tow-
quences of operations. Each operation is then tackled one at a time,
ers and reactors that do not involve any mass transfer. The flowrates
from the first to the last element in the sequence. By doing this, when
entering (foin ) and leaving (foout ) a particular unit need not be equal
addressing a particular operation, the concentration of all possible in max . Outlet con-
while their inlet concentrations can vary up to co,c
inlet streams is known beforehand, so only the flowrate variables out max and are thus independent on the
centrations are fixed to co,c
are unknown and the problem can be formulated as a LP. Thus, for
inlet concentrations.
a certain substructure, this procedure replaces a nonlinear program
To meet the demands of the water-using units, a set W of fresh-
by a succession of LPs. Although part of the feasible region is lost in wat ) and value v are
water sources with known concentrations (cw,c w
the process, the results for a collection of test cases have shown that
given. The objective is to minimize the total freshwater flowrate into
the best solution from the initialization stage is often the global op-
the system. As an alternative, given the cost parameters for the dif-
timal solution, which was always obtained following the NLP stage.
ferent water sources, it is straightforward to implement the mini-
The major drawback is that for a problem involving |Ofl | fixed load
units and at least one fixed flowrate unit, a total of |Ofl |! × |Ofl |2 LPs mization of the total cost of freshwater.
and |Ofl |! × |Ofl | NLPs need to be solved, a number that grows rapidly
with the number of units. Nevertheless, the proposed algorithm was 2.1. Remarks
generally much faster than BARON.
With the aim of reducing the overall computational burden of the While fixed flowrate operations may seem an odd concept, due
search procedure, while keeping the same success rate, this paper to the fact that they assume outlet concentrations to be indepen-
proposes two novel MILP-based algorithms that have the advantage dent of inlet concentrations, they can be very helpful when dealing
of not requiring a scan over all possible sequences of operations. The with industrial cases. The limiting case of foout = 0 allows modeling
novelty is to employ binary variables to select the best operation for a water sinks, while that of foin = 0 models water sources with known
particular position (stage) in the sequence. Again the initial network flowrates and properties. In between, we can handle a subsystem of
is generated sequentially, with only the previously unselected units units whose internal operation can be modified to keep the outlet
being considered in subsequent calculation stages. The conceptual conditions unchanged for slightly different inputs. A good example
differences between the three methodologies will be highlighted can be seen in El-Halwagi et al. (2004) and El-Halwagi (2006), where
after the formal problem statement is presented. the optimization of a pulp and paper process is considered, see Fig. 1.
in = 100
The inlet flowrates to the paper machines are fixed to fPMI
in
and fPMII = 40 t/h. Other given data states that the properties of the
2. Problem statement
in max = 0.02, cin max = 0.00125,
inlet stream must be lower than: cPMI,OM PMI,k
in max = 0.00, cin max = 0.00125. The aqueous outlet streams
This paper focus on the design of multicontaminant industrial cPMII,OM PMII,k
water-using networks without considering the treatment units that from the paper machines go to a hydro-pulper and a hydro-sieve,
are required to lower the concentration of the network's wastewater which generates a single output stream, containing broke, with the
Fig. 1. Pulp and paper process before and after optimization (adapted from El-Halwagi, 2006).
J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752 3739
out
following characteristics: fBroke out max = 0.08, cout max =
= 30, cBroke,OM Splitters located immediately after the units (SPo ), generate the input
Broke,k
0.00130. In the current solution (on the left), this valuable stream is streams to the mixers or to the wastewater treatment system located
discarded. A better network design is achieved if most of the broke downstream.
is recycled to PMI, so as to meet the upper bound on OM (shown The general non-convex NLP formulation is given next. The model
on the right). Note that despite the change in the inlet conditions uses total flows and concentrations as variables, more specifically:
to PMI and PM2 (in all three properties), the broke stream remains wat , represents the flowrate of fresh water source w needed to sat-
Fw,o
unchanged, which clearly highlights the usefulness of fixed flowrate isfy the demand for operation unit o; Co,c in and C out are, respectively,
o,c
operations. the inlet and outlet concentrations for the fixed contaminant load
oper
In this particular example, the paradigm of property integration operations; Fotot is the total flowrate into fixed load operation o; Fj,o
(El-Halwagi et al., 2004; El-Halwagi, 2006) was the approach used. It gives the total flowrate from operation j to operation o; and Fo rep-
tsys
considered three distinct properties: (i) objectionable material, OM; resents the outlet flowrate from operation o heading for the treat-
(ii) absorption coefficient, k; (iii) reflectivity, R ∞ . One of the advan- ment system. Since the constraints have been described in previous
tages of properties over components is that they are more manage- work (Teles et al., 2008), in the interest of space, no explanation is
able to track in cases with numerous chemical pollutants (e.g. com- provided.
plex hydrocarbons and ligno-cellulosic materials). The property of The objective function is the minimization of the total freshwa-
the mixture is evaluated as a function of both the flowrate and prop- ter cost, represented by Eq. (2). If one makes vw = 1 ∀ w, the ob-
erty of each stream. For objectionable material and absorption coef- jective becomes the minimization of the total freshwater flowrate.
ficient, the mixing rules are linear so the proposed solution method We will be making this assumption throughout this work, which
can handle them in the same way as concentrations. does not compromise in any way the applicability of the algorithms
to the more general case. Since the water sources will be speci-
3. General NLP formulation fied with random concentrations for the contaminants, there will be
no cleanest water source being preferred over the others, so multi-
There are likely many possible designs for a WUN, since the inlet
ple sources will indeed be selected. Nevertheless, we recognize that
freshwater streams and the several water-using units can be inter-
the water sources defined in the example problems do not exist in
connected in a wide variety of ways. With the purpose of perform-
practice.
ing a systematic search to determine the optimal flowsheet config-
The balances on the mixers, splitters and fixed mass load units
uration, we make use of a general network superstructure (Fig. 2),
are given by Eqs. (3)–(6). Eqs. (7)–(9) place upper bounds on the
similar to the one proposed by Wang and Smith (1994) that forms
contaminant concentrations that are allowed in the unit's inlet and
the basis for the NLP mathematical formulation. This structure in-
outlet streams by Eqs. (7)–(9). In order to make it easier to identify
corporates the full set of freshwater streams (i.e. more than one
bilinear terms (in Eqs. (5), (6) and (9)), variables are represented in
quality of freshwater can be handled) and water-using units. The
capital letters to differentiate them from model parameters.
rectangles representing the fixed load operations are filled in white
(O1 , . . . , O|Ofl | ), while those corresponding to the fixed flowrate units Min wat
vw · Fw,o (2)
are shadowed (O|Ofl |+1 , . . . , O|Ofl |+|Off | ). Each unit is preceded by a mixer w∈W o∈O
(MXo ), which merges freshwater and/or the reuse streams coming oper
out of the other units. A particular freshwater source is split (SPw ) Fotot |o∈Ofl + foin |o∈Off = wat
Fw,o + Fj,o ∀o ∈ O (3)
into several streams and sent to the various mixers in the network. w∈W j∈O
tsys
oper
Fotot |o∈Ofl + foout |o∈Off = Fo + Fo,j ∀o ∈ O (4) section presents three alternative efficient methods to generate fea-
j∈O sible water-using networks for the initialization of the general NLP.
We first give an overview of the LP-based method of Teles et al.
oper
Fotot · Co,c
in
= wat
Fw,o wat
· cw,c + Fj,o out
· (Cj,c out
|j∈Ofl + cj,c max
|j∈Off ) (2008) and then present the new MILP-based strategies.
w∈W j∈O
∀o ∈ Ofl , c ∈ C (5) 4.1. Review of LP-based method of Teles et al. (2008) (TCN)
Table 1
Problem data for example 13.
c in max
c out max
c in max
c out max
c in max
c out max
c in max
c out max
Table 2 lected units are known. Eq. (14) replaces the mass balance over the
Results obtained for algorithm M1 in Ex13.
fixed load units (Eq. (6)). Since the outlet concentrations need not
Initial solution (t/h) NLP solution (t/h) Obtained sequence to be determined, it is enough to ensure that the maximum outlet
concentration is not exceeded. Note that the mass exchange term is
S1 181.820 178.263 O1 –O6 –O4 –O5 –O4 –O3
S2 209.623 178.263 O2 –O4 –O6 –O5 –O3 –O1 only added for the chosen unit, while for the others, Eq. (13) ensures
in max ⱕ cout max . Eq. (15) assures
that Eq. (14) is always met since co,c
S3 201.483 189.849 O3 –O4 –O5 –O6 –O2 –O1 o,c
S4 216.764 178.263 O4 –O5 –O6 –O2 –O3 –O1 that the outlet flowrate from previously selected unit j to the chosen
S5 216.764 178.263 O5 –O4 –O6 –O2 –O3 –O1 unit o, does not exceed the amount still available. Eq. (16) ensures
S6 213.227 178.263 O6 –O4 –O5 –O2 –O3 –O1
a zero inlet flowrate for non-selected fixed mass load units, where
the limiting flowrate folim acts as the upper bound. Finally, Eq. (17)
guarantees that a single unit is selected.
The MILP formulation is given next and its constraints are very
wat
similar to those of the general NLP formulation given in Section 3. Min Fw,o (10)
The objective function in Eq. (10) minimizes the flowrate entering w∈W o∈AcO
the set of active operations (AcO). Eqs. (11) and (12) give the flowrate oper
balances over the MXo mixers and SPo splitters, respectively. No- Fotot |o∈Ofl + foin |o∈Off · Yo = wat
Fw,o + Fj,o ,
tice that the inlet/outlet flowrates to the fixed flowrate units, sec- w∈W j∈PvO∨(j∈AcO∧j∈Off ∧j=o)
ond term on the left-hand side (LHS) are only accounted for if the ∀o ∈ AcO (11)
unit is selected, i.e. if the binary assignment variable Yo is equal to
tsys oper
1. In Eq. (11), only previous selected units or the unit itself, if of the Fotot |o∈Ofl + foout |o∈Off · Yo = Fo + Fo,o |o∈Off , ∀o ∈ AcO (12)
fixed flowrate type, can feed operation o, second term on the right-
hand side (RHS). In Eq. (12), the unit outlet stream can either go to wat
Fw,o wat
· cw,c + [ Fj,o
oper out
· cj,o
|j∈Ofl + Fj,o
oper out
· cj,c max
|j∈Off ]
treatment (in subsequent stages this stream may actually be used w∈W j∈PvO
totally or partially in other units) or be recycled, if o ∈ Off . The max- operout max
+ (Fo,o · co,c )|o∈Off
imum inlet concentration constraint over the MXo mixers is given
by Eq. (13), where all the outlet concentrations from previous se- ⱕ (Fotot |o∈Ofl + foin |o∈Off ) · co,c
in max
, ∀o ∈ AcO, c ∈ C (13)
J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752 3743
oper oper
⎛ ⎞
wat
Fw,o wat
· cw,c + [ Fj,o out
· cj,o
|j∈Ofl + Fj,o out
· cj,c max
|j∈Off ]
oper oper
wat wat out out max
(in this case, to avoid a bilinear term, one assumes that the outlet Fw,o · cw,c + [ Fj,o · cj,o
|j∈Ofl + Fj,o · cj,c |j∈Off ]
concentrations are at the maximum values, Doyle and Smith, 1997). w∈W j∈PvO
oper
The flowchart in Fig. 4 can also be used for M2, and will now be out max
+ (Fj,o · cj,c )
illustrated with the example introduced in Section 4.2, assuming j∈AcO∧(j∈Off ∨(j∈Ofl ∧o∈Off ))
that the fixed flowrate system is fixed to position two (s = 2), like in ⱕ (Fotot |o∈Ofl + foin |o∈Off ) · co,c
in max
, ∀o ∈ AcO, c ∈ C (23)
Fig. 6.
Starting with p2 = 1, the MILP selects unit O4 , leading to P2,o4 =1. oper oper
wat wat out out max
Fw,o · cw,c + Fj,o · cj,o
|j∈Ofl + Fj,o · cj,c |j∈Off
Afterwards, the dynamic sets are updated and the active operations
w∈W j∈PvO
for the second calculation stage, which includes the fixed flowrate oper out max
units, are AcO = {O1 , O2 , O3 , O5 , O6 }, while the unit just selected goes + (Fj,o · cj,c )
into the set of previous operations, PvO = {O4 }, see Fig. 7. Continuing j∈AcO∧j∈Off
with the second iteration (p2 = 2), the optimizer selects fixed load + mo,c · Yo ⱕ Fotot · co,c
out max
, ∀o ∈ Ofl ∩ AcO, c ∈ C (24)
unit O2 and so the next assignments are: P2,o2 = 2, AcO = {O1 , O3 }
and PvO = {O2 , O4 , O5 , O6 }. In the third stage, unit O3 is selected Yo = 1 (25)
leaving O1 to be the last element in the sequence. o∈AcO
o∈Ofl
The MILP formulation is given next. M2 shares with M1 the objec-
tive function (Eq. (10)), as well as Eqs. (15) and (16). The remaining The results obtained for all starting points can be seen in Table 3.
constraints, Eqs. (21)–(25), have some differences when compared
to Eqs. (11)–(14) and (17). Firstly, in M2, only fixed load units can
4.4. Remarks
be selected and so the binary variables are restricted to the Ofl sub-
set, as it can be seen in Eq. (25). Secondly, in the calculation stage
In design problems it is often required to enforce/forbid con-
incorporating the fixed flowrate system, there are more possibilities
nections between units. In addition, it is highly desirable to avoid
for water reuse, as mentioned before. Taking Eq. (23) as an example,
streams with very low flowrates due to economical or controllability
unit j can feed unit o if it has been assigned to previous stages or if
reasons. Such constraints can be easily incorporated in the proposed
j ∈ Off ∧ j = o, like in M1 (see Eq. (13)). Now, however, more than a wat
solution strategies by using two new sets of binary variables: Yw,o
single unit will normally be involved, so units j, either fixed load or
for identifying the match between freshwater source w and opera-
fixed flowrate, can feed fixed flowrate units o. oper
tion o; Yj,o for the match between units j and o. Assuming that f min
oper represents the lower bound on the flowrate, we get the following
Fotot |o∈Ofl + foin |o∈Off = wat
Fw,o + Fj,o ,
sets of constraints for M1 (the domain of the constraints would be
w∈W j∈PvO∨(j∈AcO∧(j∈Off ∨(j∈Ofl ∧o∈Off )))
slightly different for TCN and M2).
∀o ∈ AcO (21)
wat
Fw,o ⱖ f min · Yw,o
wat
, ∀w ∈ W, o ∈ AcO (26)
tsys oper
Fotot |o∈Ofl + foout |o∈Off = Fo + Fo,j , ∀o ∈ AcO (22)
oper oper
j∈Off ∨(j∈Ofl ∧o∈Off ) Fj,o ⱖ f min · Yj,o , ∀j ∈ PvO, o ∈ AcO (27)
J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752 3745
Table 4
Physical characteristics of the example problems.
|Ofl | |Off | |W| |C| |Ofl | |Off | |W| |C| |Ofl | |Off | |W| |C| |Ofl | |Off | |W| |C|
Table 5
Lowest total freshwater consumption (t/h) after initialization.
Solver CPLEX
Table 6
Best found solution in terms of total freshwater consumption.
optimal, whereas M1 and M2 were successful in two out of four different networks with just a few starting points. A good illustra-
cases. tive example is Ex31, where every starting point leads to a different
The most interesting result from Table 7 is that the number of solution. In terms of the success rate, one can argue that it gives
different solutions is very similar for the three methods, TCN, M1 a measure of complexity, with a lower rate being associated with
and M2, which is to some extent surprising since for TCN the num- a greater tendency to get trapped in local solutions. Naturally, the
ber of NLPs solved is typically one order of magnitude larger. This larger models in terms of the number of operations, like Exs 28, 31,
confirms the effectiveness of M1 and M2 in generating structurally 32 and 34 (see Table 4), were found to be the most complex. In this
J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752 3747
Table 7
Number of local optima found and algorithm's success rate.
Method # NLPs Local Optima Success rate (%) # NLPs Local Optima Success rate (%)
Table 8
Total computational effort (in CPUs).
respect, units of the fixed load type seem to add more complexity tractable since they lead to the solution of a prohibitively high num-
than their fixed flowrate counterparts, which again was not totally ber of problems. The other alternative is to solve the general NLP
unexpected since as mentioned before, the latter form a linear sub- with the global optimization solver BARON, but even more compu-
system. tational resources may be involved, where CPU values of the order of
days can be noted, which are not even a guarantee that the optimal
solution will be found. Thus, it is reasonable to conclude that M1 is
5.4. Total computational effort the most efficient approach since it gives the best trade-off between
solution quality and speed.
The last performance indicator is the total computational effort.
The results in Table 8 show that the computational time is directly
6. Illustrative examples
related to the number of mathematical problems to solve since it in-
creases in the order M2–M1–TCN. It is particularly relevant to note
In this section, the proposed algorithms M1 and M2 are illustrated
that the average time was 92 CPUs, with 7 min being the maximum
in terms of the best water-using networks that result from the ini-
computational time for algorithm M1 (in Ex30). In contrast, the hard-
tialization and NLP phases. Four test cases, Exs 21, 22, 33 and 11 are
est problem to tackle by M2 was Ex28, which needed about 3 min,
used for this purpose. The first will be used to guide us through the
while the average was equal to 28 CPUs. Given the higher success
network representation diagrams, which feature the optimal values
of M1 in finding the optimal solution of 97% (i.e. in 35 out of 36
of the variables together with the problem data.
examples) as against 78% for M2 (28 out of 36) and the amount of
computational resources involved, the added effort can only be con-
sidered a small penalty to pay. While TCN was fully successful at 6.1. Ex21
obtaining the optimal solution, the total effort was larger by about
one order of magnitude. Furthermore, examples comprising more Ex21 involves four freshwater sources, four fixed load and two
than seven fixed load operations (Exs 26–36) can be considered in- fixed flowrate units, and six contaminants. The best solution from
3748 J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752
Fig. 9. Best network from the initialization process of TCN for Ex21.
the initialization stage of M2 is given in Fig. 8. The chosen sequence as TCN, since the underlying superstructure is the same, so this is a
of operations is O4 –{O1 ,O5 , O6 }–O3 –O2 resulting in a total fresh- surprising result. If one looks at the first calculation stage, involving
water consumption of 217.259 t/h. This value can be found above O4 , one realizes that the solution from both methods is the same,
the left-most rectangle, which represents the source of the different 9.286 t/h of W3, but for the second stage the solutions degenerate for
freshwater streams. The right-most rectangle represents the sink for a common partial freshwater consumption of 125.466 t/h. While in
the wastewater streams heading for the treatment system down- Fig. 8 (M2), the freshwater source for O1 is W3, in Fig. 9 (TCN) is W1.
stream, which total 167.259 t/h. Above the arrows representing each The main consequence is that the outlet concentrations are not the
freshwater stream, the required flowrate is indicated, and below, same, which will affect the freshwater requirements in the follow-
the contaminants' concentration, one per row. In this case, three ing two calculation stages. In fact, for M2, O3 and O2 require slightly
freshwater streams were selected (W1, W3 and W4) with flowrates higher flowrates (34.943 and 47.565 vs. 34.917 and 47.437 t/h, re-
equal to 106.481, 28.271 and 82.508 t/h, respectively. The same logic spectively) than in the resulting network from TCN, leading to an
is used to present the contaminant values for the inlet and outlet overall higher water consumption and wastewater generation. This
streams to/from a particular unit but now two columns of data are is a good illustration of the major disadvantage of decomposition
required: the optimal and the maximum allowed concentrations. methods over the solution of the full problem.
Whenever the maximum values are reached, the values are shown in The optimal network has a total freshwater consumption equal
bold. Notice that fixed flowrate units O5 and O6 are characterized by to 216.730 t/h and is a global optimal solution within a 0.02% toler-
fixed maximum outlet concentrations and this is why all values are ance (see Table 6). Besides BARON, all three algorithms were able to
in bold. Units of these type also have fixed inlet and outlet flowrates find networks with such consumption, which does not necessarily
(e.g. 125 and 40 t/h in O6 ). In contrast, fixed load units have equal mean that the networks were exactly the same. In fact, the result-
inlet and outlet flowrates and, besides the maximum concentration ing networks are often different, indicating some degree of degen-
values, are characterized by a limiting flowrate, which is given above eracy. The solution from M2 is given in Table 9, which also features
the corresponding block. the initial values prior to the solution of the NLP. It is obvious that
A very similar solution but with a lower freshwater consumption the structures are significantly different. First, no stream kept the
(e.g., 217.105 t/h), was obtained by TCN for the same sequence of op- same flowrate. Second, of the 15 initial interconnections, only eight
erations (see Fig. 9). In theory, M2 is able to obtain the same solution were selected in the optimal solution. Third, on some of these, the
J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752 3749
Table 9
Flowrates in best initialization network and corresponding NLP solution from M2, Ex21.
(t/h) O1 O2 O3 O4 O5 O6
Init. NLP Init. NLP Init. NLP Init. NLP Init. NLP Init. NLP
flowrate values changed considerably (e.g. from 11.284 to 17.314 in 323.505 t/h. In contrast, M2 generated a network with a consump-
O1 –O2 ). Similar results were observed in other examples. tion of 327.237 t/h (1.2% higher). Of the five freshwater sources, W2
was discarded by the optimization procedure, while W4 was the
6.2. Ex22 preferred one.
In the optimal network shown in Fig. 10, five units have inlet
Ex22 consists of four fixed load and three fixed flowrate oper- concentrations equal to their maximum values for at least one con-
ations involving six contaminants and five water sources. The best taminant and the limiting contaminants are C1 for O3 and O6 , C2 for
solution found for Ex22 can also be considered a global one since, at O5 , C3 for O7 , C5 for O1 and O3 , and C6 for O3 . Notice that whenever
the time of termination, the lower bound from BARON was within the maximum inlet value is reached for a contaminant, the fixed load
0.03% (see Table 6). Besides BARON, methods TCN and M1 could unit's inlet flowrate is equal to the limiting flowrate and the outlet
also find networks with a total freshwater consumption equal to concentration is set at its maximum. In fixed flowrate units it is as-
3750 J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752
Table 10
Units inlet and outlet concentrations for the network shown in Fig. 11.
Conc. (ppm) C1 C2 C3 C4
C in
C out
C in
C out
C in
C out
C in C out
sumed that the outlet concentrations are not affected by the quality 6.3. Ex33
of the inlet stream. This is the reason why their recycling streams
have normally a higher flowrate than those of fixed load units (e.g., Ex33 is one of the examples were the new algorithms were able
30.270 t/h for O6 ). to beat BARON in both solution quality and speed so it deserves some
One important feature of the optimal network shown in Fig. 10 attention. The best solution obtained features a total freshwater con-
is that the outlet stream from O3 has outlet concentrations below sumption of 565.343 t/h and may not be the global optimum. The
the maximum values, which contradicts one of the necessary condi- corresponding network is given in Fig. 11, while the concentration
tions of optimality for multicontaminant systems given in Bagajewicz data can be found in Table 10. It shows a complex mesh of inter-
(2000) that states that at least one component must reach its max- connections, in a total of 34 amid reutilizations and recycles. Some
imum outlet concentration. Networks with similar characteristics of these involve very low flowrates (e.g. 0.499 t/h in the recycling
had already been reported in Teles et al. (2008). The fact that the stream of O7 ), which can be avoided by using the sets of constraints
outlet concentration constraints are not active in the solution is described in Section 4.4, at the likely penalty of a small increase in
an indication that O3 is not determining the solution. This can be freshwater consumption. More specifically, for fmin = 1 t/h, the to-
easily verified by performing a sensitivity analysis on the influ- tal freshwater flowrate increases by less than 0.1% to 565.715 t/h.
ence of the limiting flowrate on the value of the objective func- Finally, and as noted in Ex21, the magnitude of the recycling flow
tion. More specifically, the total freshwater consumption remains depends on the operation type. Fixed flowrate operations O15 and
the same (323.505 t/h) for fOlim ∈ [16.5, 67.471] (40 t/h was the value O16 feature values of 46.163 and 10.296 t/h, respectively, which are
3
used). significantly greater than the recycling flow in fixed load unit O7 .
J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752 3751
Fig. 12. Best solutions found for Ex11 without (left) and with f min = 1 t/h (right).
6.4. Ex11 system to a certain stage, and then considering all possible combi-
nations.
To conclude the discussion, we illustrate the differences between The results have shown that the algorithm with no distinction
the best networks obtained with and without restrictions on the between operation types was the best approach since it could find
minimum stream flowrate, for Ex11, see Fig. 12. The penalty on the optimal solutions in all but one example from the large set of
freshwater consumption is, like for Ex33, very small, 0.017 t/h. thirty six instances considered. It was even more successful than
The major structural change is that we are no longer limited to the global optimization solver BARON, with the added advantage of
freshwater W4. In fact, 15.971 t/h of W3 are being used and W4 con- being orders of magnitude faster. The performance of the new algo-
sumption is down to 88.933 t/h. The similar contaminant concen- rithms has also been compared to that of a closely related heuristic
trations of these streams are probably the reason why the flowrate procedure developed by the present authors. In such previous work,
differences are considerable. The elimination of the 0.004 t/h O3 the unit assignment to calculation stages was made a priori, lead-
recycling stream is compensated by the new W3 –O1 stream, so the ing to a series of linear programs and a significantly larger number
number of interconnections remains the same, 13. The other unde- of starting points resulting from the exhaustive enumeration of all
sirably low flowrate, O4 –O1 , now features the minimum specified possibilities. The results have shown that the new best method is as
value, f min = 1 t/h. effective as its predecessor in escaping local optima, requires signif-
icantly fewer computational resources and can go well beyond the
7. Conclusions previously observed practical limit of seven fixed load units. Large
problem instances can be tackled with fairly reasonable computa-
This paper has presented two new algorithms for the optimal de- tional resources as illustrated with those examples featuring 20 units,
sign of water-using networks with multiple contaminants featuring which took less than 7 min to solve.
fixed contaminant load as well as fixed flowrate units. They can be
viewed as two-phase procedures, where the initialization phase pre- Notation
cedes the optimization of a standard nonlinear programming prob-
lem derived from a superstructure that includes all alternatives for Sets/indices
freshwater use and wastewater reuse and recycling. The novelty is
the generation of starting points based on the sequential solution
of a few mixed integer linear programs, where the binary variables AcO active operations, those that can be assigned to the
are used to select the most appropriate unit in terms of minimum calculation stage under consideration
freshwater consumption. The sequential procedure avoids the need C/c process contaminants
of bilinear terms by ensuring that the concentration of all possible I numbering of units ({1, . . . ,|O|} for M1, {1, . . . ,Ofl |} for
inlet streams to one calculation stage are known beforehand or from M2)
the results from previous stages. The MILP-based procedures differ Off fixed flowrate operations
in the way units are treated, either evenly or according to their type Ofl fixed contaminant load operations
(fixed load or fixed flowrate). While both methods are in essence O/o,j water-using operations
single starting point, it has been found that multiple starting points PvO operations allocated to previous calculation stages
are desirable to avoid local optima. These were generated by fixing W/w freshwater sources
a particular unit to the first calculation stage or the fixed flowrate s initialization point under consideration
3752 J.P. Teles et al. / Chemical Engineering Science 64 (2009) 3736 -- 3752
Parameters Aly, S., Abeer, S., Awad, M., 2005. A new systematic approach for water network
design. Clean Technol. Environ. Policy 7, 154.
in max
co,c maximum inlet concentration of contaminant c, Alva-Argaéz, A., Kokossis, A., Smith, R., 2006. The design of water-using systems
in petroleum refining using a water-pinch decomposition. Chem. Eng. J.,
operation o
out max doi:10.1016/j.cej.2006.10.001.
co,c maximum outlet concentration of contaminant c, Bagajewicz, M., 2000. A review of recent design procedures for water networks in
operation o refineries and process plants. Comput. Chem. Eng. 24, 2093.
wat
cw,c concentration of freshwater w in contaminant c Biegler, L., Grossmann, I., Westerberg, A., 1997. Systematic Methods of Chemical
Process Design. Prentice-Hall PTR, New Jersey, USA.
foin inlet flowrate to fixed flowrate operation o Brooke, A., Kendrick, D., Meeraus, A., Raman, A., 2005. GAMS: A Users Guide. GAMS
folim limiting flowrate of fixed load operation o Development Corporation, Washington, DC.
foout outlet flowrate of fixed load operation o Castro, P., Matos, H., Fernandes, M., Pedro Nunes, C., 1999. Improvements for mass-
exchange networks design. Chem. Eng. Sci. 54, 1649.
f min minimum allowed flowrate for any stream Castro, P.M., Matos, H.A., Novais, A.Q., 2007. An efficient heuristic procedure for the
ps calculation stage of initialization s under considera- optimal design of wastewater treatment systems. Resources Conservation and
tion Recycling 50, 158–185.
Castro, P.M., Teles, J.P., Novais, A.Q., 2009. Linear program-based algorithm for the
tfsMILP total freshwater flowrate for initialization s optimal design of wastewater treatment systems. Clean Techn. Environ. Policy
tfsNLP total freshwater flowrate after the NLP stage for 11, 83–93.
point s Cormos, C.-C., Bandyopadhyay, S., 2008. Water management in process industries
incorporating regeneration and recycle through a single treatment unit. Ind. Eng.
vw value of freshwater w Chem. Res. 47, 1111–1119.
mo,c mass exchange in operation o for contaminant c Doyle, S.J., Smith, R., 1997. Targeting water reuse with multiple contaminants. Trans.
Ps,o position in initialization s of operation o IChemE 75 (B), 181.
Deng, C., Feng, X., Bai, J., 2008. Graphically based analysis of water system with zero
Variables (corresponding parameters) liquid discharge. Chem. Eng. Res. Des. 86, 165–171.
El-Halwagi, M.M., 2006. Process integration. In: George, S., Efstratios, P. (Eds.), Process
in
Co,c inlet concentration of contaminant c for fixed load Systems Engineering, vol. 7. Academic Press, Elsevier.
El-Halwagi, M.M., Gabriel, F., Harell, D., 2003. Rigorous graphical targeting for
operation o resource conservation via material recycle/reuse networks. Ind. Eng. Chem. Res.
out (cout )
Co,c outlet concentration of contaminant c for fixed load
o,c 42, 4319.
operation o El-Halwagi, M.M., Glasgow, I.M., Eden, M.R., Qin, X., 2004. Property integration:
oper componentless design techniques and visualization tools. A.I.Ch.E. J. 50, 1854.
Fj,o flowrate from operation j into operation o Feng, X., Seider, W., 2001. New structure and design methodology for water networks.
Fotot total flowrate into fixed load operation o Ind. Eng. Chem. Res. 40, 6140.
wat
Fw,o flowrate from freshwater w into operation o Gunaratnam, M., Alva-Argáez, A., Kokossis, A., Kim, J.-K., Smith, R., 2005. Automated
tsys tsys design of total water system. Ind. Eng. Chem. Res. 44, 588.
Fo (fo ) total flowrate from operation o to the treatment Hallale, N., 2002. A new graphical targeting method for water minimisation. Adv.
system Environ. Res. 6, 377.
Yo binary variable identifying the selection of operation Hernández-Suárez, R., Castellanos-Fernández, J., Zamora, J.M., 2004. Superstructure
decomposition and parametric optimization approach for the synthesis of
o in a certain calculation stage distributed wastewater treatment networks. Ind. Eng. Chem. Res. 43, 2175.
oper
Yj,o binary variable identifying if the connection Huang, C., Chang, C., Ling, H., Chang, C., 1999. A mathematical programming model
between operation j and o exists for water usage and treatment network design. Ind. Eng. Chem. Res. 38, 2666.
wat Karuppiah, R., Grossman, I., 2006. Global optimization for the synthesis of integrated
Yw,o binary variable identifying if the connection water systems in chemical processes. Comput. Chem. Eng. 30, 650.
between freshwater w and operation o exists Karuppiah, R., Grossman, I., 2008. Global optimization of multiscenario mixed integer
nonlinear programming models arising in the synthesis of integrated water
networks under uncertainty. Comput. Chem. Eng 32, 145–160.
Variables attributes Lim, S.-R., Park, J.M., 2008. Synthesis of an environmentally friendly water network
system. Ind. Eng. Chem. Res. 47, 1988–1994.
Liu, Y., Duan, H., Feng, X., 2008. The design of water-reusing network with a hybrid
structure through mathematical programming. Chin. J. Chem. Eng. 16, 1–10.
l level, either initial value or value in optimal solution Lavric, V., Iancu, P., Plesu, V., 2005. Genetic algorithm optimisation of water
consumption and wastewater network topology. J. Cleaner Prod. 13, 1405.
Manan, Z., Tan, Y., Foo, D., 2004. Targeting the minimum water flow rate using
Acknowledgments water cascade analysis technique. A.I.Ch.E. J. 50, 3169.
Ng, D.K.S., Foo, D.C.Y., Tan, R.R., Tan, Y.L., 2007. Ultimate flowrate targeting with
regeneration placement. Chem. Eng. Res. Des. 85, 1253–1267.
The authors gratefully acknowledge financial support from Fun-
Prakash, R., Shenoy, U., 2005. Targeting and design of water networks for fixed
dação para a Ciência e Tecnologia through project PTDC/AMB/64659/ flowrate and fixed contaminant load operations. Chem. Eng. Sci. 60, 255.
2006 and research grant SFRH/BD/41062/2007. Prakotpol, D., Srinophakun, T., 2004. GAPinch: genetic algorithm toolbox for water
pinch technology. Chem. Eng. Process. 43, 203.
Relvas, S., Matos, H., Fernandes, M., Castro, P., Nunes, C., 2008. AquoMin: a software
Appendix A. Supplementary materials tool for mass-exchange networks targeting and design. Comput. Chem. Eng. 32,
1085–1105.
Supplementary data associated with this article can be found in Takama, N., Kuriyama, T., Shiroko, K., Umeda, T., 1980. Optimal water allocation in
a petroleum refinery. Comput. Chem. Eng. 4, 251.
the online version at doi:10.1016/j.ces.2009.04.051 Teles, J., Castro, P.M., Novais, A.Q., 2008. LP-based solution strategies for the optimal
design of industrial water networks with multiple contaminants. Chem. Eng.
References Sci. 63, 376–394.
Wang, B., Feng, X., Zhang, Z., 2003. A design methodology for multiple-contaminant
water networks with single internal water main. Comput. Chem. Eng. 27, 903.
Agrawal, V., Shenoy, U.V., 2006. Unified conceptual approach to targeting and design
Wang, Y., Smith, R., 1994. Wastewater minimization. Chem. Eng. Sci. 49, 981.
of water and hydrogen networks. A.I.Ch.E. J. 52, 1071–1082.