Académique Documents
Professionnel Documents
Culture Documents
Abstract
Efficient transport of timber for supplying industrial conversion and biomass power plants is a
crucial factor for competitiveness in the forest industry. Throughout the recent years
minimizing driving times has been the main focus of optimizations in this field. In addition to
this aim the objective of reducing environmental impacts, represented by carbon dioxide
depot vehicle routing problem with pick-up and delivery and time windows (MDVRPPDTW)
and a new iterative method is proposed. For the numerical studies, real life instances of
different scale concerning the supply chain of biomass power plants are used. Small ones are
10
taken to validate the optimality of the new approach. Medium and large instances are solved
11
with respect to minimizing driving times and fuel consumptions separately. This paper
12
analyses the trade-offs between these objectives and shows how an additional mitigation of
13
14
Keywords: Green Logistics, Timber Transport, Greenhouse Gas Mitigation, Mixed Integer
15
16
Introduction
17
The forest based sector plays a significant role for several countries. Besides the well known
18
importance for New Zealand, Sweden, Finland, Chile and Canada, it also accounts for a
19
prominent share of the economy in other countries like Austria. In Austria more than a
20
hundred thousand people are working in sectors related to forest, timber and paper industry.
21
From this it follows that efficient timber transports are a main interest of freight forwarders.
22
Zazgornik et al. (2012) mention approximately 500,000 log-truck trips per year that originate
23
in Austrian forests, not considering transports due to timber imports. Several publications
24
give estimations of the percentage of transport costs in relation to the total timber price (e.g.
25
von Bodelschwingh (2001), Murphy (2003), Favreau (2006)). In all of them, a value of
26
27
According to Flisberg et al. (2009) log-truck scheduling is traditionally done manually. This is
28
also the case in companies that the authors have worked with in Austria. Hence, finding
29
efficient ways of planning the transports and thereby reducing the costs is crucial.
30
The tackled problem in this paper is related to the log-truck scheduling problem (LTSP, e.g.
31
Palmgren et al. (2003)) and to the pick-up and delivery problem with time windows (PDPTW,
32
e.g. Ropke & Pisinger (2006)). The pick-up takes place at a wood storage location, where the
33
afore empty log-truck is fully loaded. Afterwards, the delivery location, which is an industrial
34
site, is visited and all transported wood is unloaded. Usually the next transport task follows
35
for the empty log-truck, thus the log-truck visits another wood storage location. Therefore,
36
only one wood storage location and one industrial site is visited within one trip. Due to the
37
specific construction of the used log-trucks, backhauls are unusual and not considered. The
38
given problem is similar to the one described in El Hachemi et al. (2010), but dependencies of
39
the activity of log loaders at wood storage locations and industrial sites as well as
40
consequential waiting times are neglected. Besides this difference, log-trucks have a given
41
home-base or depot, as used in the LTSP, respectively. Each log-truck has to start and end its
42
tour at its depot, whereas it is possible that more than one log-truck is located at one depot.
43
The problem in this paper is referred to as multi depot vehicle routing problem with pick-up
44
and delivery and time windows (MDVRPPDTW), whereas transport tasks are predefined, as
45
46
For solving the MDVRPPDTW a new method, the so-called near-exact solution
47
48
three stages. As a first stage an extended assignment problem (EAP) is solved to generate a
49
reduced transport network. Afterwards, the given network is checked in terms of maximum
50
tour length and time window requirements. Violations lead to the generation and addition of
51
cuts to the EAP, which is thereafter solved again. If no violations occur the MDVRPPDTW is
52
solved on the given network. In order to validate the NE for small problem instances, it is
53
54
55
Sbihi & Eglese (2007) state that most research in vehicle routing and scheduling is done to
56
minimize costs. Besides the economic factor, also the environmental impact of transports can
57
be reduced by efficient planning and the advice of decision support systems. The objectives of
58
minimum costs and minimum environmental impacts are to a certain degree not in conflict
59
with each other. The European Environment Agency (EEA, 2009) states that emissions due to
60
fuel consumption contain the greenhouse gases (GHGs) carbon dioxide (CO2 ), nitrous oxide
61
(N2 O) and methane (CH4 ), as well as particulate matter (PM ), heavy metals, toxic
62
substances, carcinogenic species, ozone precursors and acidifying substances. In this paper the
63
focus is on the emission of GHGs that are presented as carbon dioxide equivalents (CO2 e).
64
The direct proportion of CO2 emissions to fuel consumptions (e.g. ICF Consulting (2006),
65
66
departments in the UK, namely the Department of Energy and Climate Change (DECC) and
67
the Department for Environment, Food and Rural Affairs (Defra), present factors for CH4 and
68
N2 O emissions as CO2 e that are added to the direct CO2 formation in the combustion engine
69
(Defra & DECC, 2011). Furthermore, they give estimations of indirect GHG emissions per
70
liter of diesel. In this paper, direct and indirect emissions are accounted for and converted in
71
kilograms of CO2 e. The aim is to analyze the trade-off for freight forwarders of minimizing
72
73
Eglese & Black (2010) review some possibilities of estimating emissions. The simplest way is
74
to assume an average speed or fuel consumption per kilometer traveled for the whole road
75
network. This approach has flaws, due to the nonlinear speed dependency of fuel consumption.
76
As a more sophisticated approach, they propose different average speeds per road class for the
77
entire road network. This approach is also used in the presented numerical studies of this
78
paper. Based on the computer programme to calculate emissions from road transport
79
(COPERT), developed by Ntziachristos & Samaras (2000), the EEA provides speed dependent
80
formulas for estimating the fuel consumption, depending on the type of truck, its maximum
81
weight limit, its exhaust emission standard, its load factor and the road gradient (EEA, 2009).
82
In a recent comparative study of different vehicle emission models for road freight
83
transportation the COPERT model provided the best estimate for heavy load vehicles with
84
weights of 50 tonnes (Demir et al., 2011). Therefore, this model was used to provide input for
85
calculating the fuel consumptions in this study. An alternative method for calculating CO2
86
emissions is used by Bektas & Laporte (2011) for the pollution-routing problem (PRP). Due to
87
their given problem it is necessary to account for change in vehicle loads during a tour of a
88
truck. This is not the case in the tackled problem of this paper, as there are only two states of
89
a truck - fully loaded or empty. Furthermore, Eglese & Black (2010) mention congestion as an
90
important factor for deviations of the average speed. To account for this factor Sbihi & Eglese
91
(2007) suggest a time dependent vehicle routing and scheduling problem (TDVRSP). However,
92
no data for time dependency are available for the test area used in this paper. Additionally,
93
timber transports are mainly carried out on rural roads, where time dependency is less
94
prominent compared to roads in urban areas. This time dependency is also assumed in the
95
emission minimization vehicle routing problem (EVRP) of Figliozzi (2010). In his study he
96
97
environments.
98
The numerical studies of this paper are carried out with real life data - concerning the supply
99
of biomass power plants - provided by the Institute of Forest Engineering of the University of
100
Natural Resources and Life Sciences, Vienna. The input matrices for solving the
101
MDVRPPDTW contain driving times and fuel consumptions. Both of them have been
102
computed twice to account for shortest paths in terms of driving times and fuel consumptions.
103
Furthermore, all instances, where fuel consumptions and driving times are compared, have
104
105
This paper presents a novel application for a routing problem in timber transport. With the
106
NE it is possible to solve some of the real life problem instances with a size of up to 60
107
transport tasks and 20 available log-trucks with their global optimum. For all other instances
108
solutions close to the global optimum are obtained in a fast manner. Furthermore, a method
109
for calculating and reducing CO2 e emissions is introduced and applied to the presented
110
problem. The objective of this paper is to give the readers an idea of how to implement the
111
minimization of greenhouse gas emissions for transportation in their research or daily business
112
and to introduce a powerful method to optimize the routing of log-trucks. Furthermore, this
113
paper analyses the trade-offs between minimizing driving times and greenhouse gas emissions
114
115
116
This section presents a detailed problem description and outlines the modelling assumptions
117
and data requirements. Special emphasis is given to the proposed solution approach (NE).
118
Problem description
119
120
A fleet of homogeneous log-trucks R that are located at depots H. These depots are
121
typically the homes of the log-truck drivers. However, parts of the fleet can also be
122
123
A set of transport tasks T that start at wood storages W and end at industrial sites S.
124
125
126
where the log-truck is fully loaded. In order to fulfill the task t T, the log-truck drives
127
to a predefined industrial site s S to fully unload its goods. Afterwards, either another
128
task is started by driving to a wood storage location or the log-truck returns to its depot
129
hr .
130
131
Wood storage locations and industrial sites can be visited more than once throughout
the planning period.
132
133
As full-truck loads are assumed, transports between different wood storage locations are
134
not allowed. The same holds for industrial sites. Additionally, all log-trucks have the
135
136
137
A feasible tour has to fulfill constraints in terms of a maximum driving time MT and
time windows. Time windows occur at depots [ah ,bh ] and industrial sites [as ,bs ].
5
138
139
For loading at the wood storage and unloading at the industrial site, service times SWw
and SIs occur.
140
With the transport tasks already defined, the objective of the log-trucks is to choose the
141
optimal empty-truck rides. If this optimization is done in terms of fuel consumption, the
142
driving times DT are exchanged by fuel consumptions FC of empty-truck rides. Due to the
143
given problem structure, choices of following tasks occur at depots or industrial sites only,
144
which - depending on the formulation - reduces the number of constraints and variables
145
markedly, compared to a vehicle routing problem (VRP) without predefined transport tasks.
146
Figure 1 shows an illustration of the given problem with two trucks, five wood storage
147
locations and two industrial sites. In this example one log-truck is located at a depot only. At
148
the beginning of the planning period the starting and end points of a transport task t T as
149
well as its duration T Dt and fuel consumption F Ct are already known. The starting point
150
equals a certain wood storage location w W and the end point is a specific industrial site
151
s S. The duration T Dt of a transport task t equals the driving time from the wood storage
152
location w to the industrial site s of t. The same concept applies to the fuel consumption F Ct
153
of a transport task t, which is the fuel consumption if driving fully loaded from the wood
154
155
Mathematical model
156
The problem is formulated as a standard MDVRPPDTW. The binary decision variable tijr is
157
158
159
160
Afterwards, an arbitrary number of transport tasks may follow. As starting and end points of
161
transport tasks are given, the driving time DTij of task i T to task j T equals the driving
162
time from the industrial site s of i to the wood storage location w of j . If j H, DTij is the
163
driving time of the last industrial site s to the depot hr , which equals a home ride.
164
eir is the completion time of task i by log-truck r. str corresponds to the starting time of the
165
tour of log-truck r and etr marks its end. The overall model formulation is given below:
min
X X
DTij tijr
(1)
iT H jT H rR
X X
i T H
(2)
k T H, r R
(3)
r R
(4)
i T, r R
(5)
i, j T, r R
(6)
j T, r R
(7)
i T, r R
(8)
i, j T H, r R
(9)
eir 0
i T H, r R
(10)
str 0
r R
(11)
etr 0
r R
(12)
s.t.
tijr = 1
jT H rR
tikr
iT H
tkjr = 0
jT H
(DTij + T Dj )tijr +
iT H jT
DTij tijr ) M T
jH
ai eir bi
eir + SIi + DTij + SWj + T Dj M (1 tijr ) ejr
tijr {0, 1}
166
The objective (1) seeks to minimize the sum of the driving times DTij of the empty-truck
167
rides. Constraints (2) ensure that each task i T H is fulfilled. A ride from i to j, with
168
i = j, is only allowed, if i, j H. This means that the log-truck does not leave the depot.
169
(3) force a log-truck to leave for another task after completing task i. (4) make sure that the
170
total driving time of a tour does not exceed the maximum tour length MT. Constraints (5)
171
and (6) guarantee that time windows at the depots and industrial sites are met, whereas
172
constant M has a large integer value, which is introduced to linearize the constraints. If j H,
173
the parameters T Dj , SWj and SIj are 0. (7) and (8) connect the start str and end etr of a
174
tour to the end of a task eir . Constraints (9) ensure the binarity of the decision variables tijr .
175
The last constraints of (10), (11) and (12) contain the non-negativity restrictions of the given
176
decision variables.
177
CO2 e calculation
178
For an estimation of the CO2 emissions, the fuel consumption of each arc within a road
179
network has to be known, as the CO2 emissions are directly proportional to the fuel
180
consumption. Different factors for the conversion of liters of diesel to kilograms of CO2 can be
181
found in the literature. For example the EEA (2009) uses 3.14 kg CO2 per liter diesel if an
182
oxidation of 100 % of the fuel carbon is reached - which is called ultimate CO2 . Complete
183
oxidation of all carbon components is not realistic, as also carbon monoxide (CO),
184
hydrocarbons (e.g CH4 ) and P M are formed. Besides incomplete oxidation there are
185
incompustible species present in the combustion chamber e.g. nitrogen gas (N2 ) or nitrogen
186
oxides (N Ox ) out of the air (EEA, 2009). In terms of GHGs, the by-products of N2 O and
187
CH4 are important. Consequently, Defra & DECC (2011) give values for the formation of
188
these gases as CO2 e. In terms of 100 % mineral diesel 0.0012 kg CO2 e of CH4 and
189
0.0184 kg CO2 e N2 O are emitted per liter. Added to the direct formation of 2.6480 kg of CO2
190
this leads to the emission of 2.6676 kg CO2 e per liter diesel. Besides direct CO2 e emissions
191
Defra & DECC (2011) also provide information of indirect GHGs as CO2 e. They originate
192
from required preceding processes like the extraction and transport of primary fuels or the
193
refining, distribution, storage and retail of finished fuels. As a total, indirect emissions of
194
0.5085 kg CO2 e per liter diesel are reported. Overall direct and indirect emissions add up to
195
3.1761 kg CO2 e per liter diesel. The use of a certain percentage of biofuel reduces the total
196
emissions; e.g. a share of 3.6 vol % biofuel leads to emissions of 3.1073 kg CO2 e per liter
197
diesel. Worth mentioning is that the reported 3.14 kg CO2 per liter diesel for exhaust
198
emissions of vehicles in European countries - which are stated in the annex of EEA (2009) -
199
200
In EEA (2009) many formulas are available to calculate the fuel consumption for different
201
types of vehicles. In this paper only heavy-duty vehicles are focused on, as they are used in
202
timber transport. Different formulas are useable depending on the type of truck, its maximum
203
weight, its exhaust emission standard, its load factor and the road gradient. These formulas
204
provide the fuel consumption in grams per kilometer depending on the speed v. Besides the
205
unknown v, up to five factors (, , , and ) are used within the different formulas. These
206
factors are derived from statistical analyses and are given constants that can be found in the
207
annex of EEA (2009), if the aforementioned specifiers for truck and road are known.
For example the fuel consumption fc of a half loaded truck and trailer with EURO V emission
standard and a maximum weight from 34-40 tonnes on a road with a gradient of +2 % is
calculated with equation (13).
f c = v v
(13)
208
209
210
consumption of 496.05 g/km or 49.605 kg per 100 km, respectively. The consumption of diesel
211
in [g] or [kg] is not as intuitive as an indication in liters. Hence, fuel consumptions are
212
transformed to liters with the mean density of diesel at a temperature of 15 C from the
213
European Standard (EN) 590:2009 of 0.8325 kg/l. So the above calculated use of 49.605 kg per
214
215
By the use of formula (13), it is possible to establish a network, where the specific fuel
216
consumption of each arc is known. Therefore, only the length of the arc and the average speed
217
on it have to be inserted. The use of a single average speed for all the arcs within a network
218
would lead to a result that only differs from a weighting with driving times by a constant
219
factor (Eglese & Black, 2010). This is not suitable for comparing the results of minimizing
220
driving times on the one hand to the results of minimizing CO2 e emissions on the other hand.
221
Hence, arcs are divided into different segments depending on their road class. Each segment
9
222
has a length and a corresponding average speed and by adding up the fuel consumptions of all
223
the segments of an arc, the total fuel consumption from one node to another node is retrieved.
224
From this information a fuel matrix is obtained by taking the minimum fuel consumption from
225
each node to every other node. This matrix differs from the one with minimum driving times.
226
Fuel consumptions and driving times are speed dependent, but in contrast to driving times,
227
the speed dependency of fuel consumptions is not linear. This leads to different routes within
228
min
X X
F Cij tijr
(14)
iT H jT H rR
229
To obtain the total emissions in CO2 e the fuel consumption is multiplied with 3.14 kg CO2 e
230
231
Solution approach
232
The NE takes the structure of the given problem into account and solves the problem either
233
exactly or heuristically, if computing times are unreasonable. By using this approach the
234
MDVRPPDTW becomes solvable for problem sizes that cannot be reached by applying
235
standard model formulations like the one presented before. Both approaches are implemented
236
in the programming language Mosel and are solved with the solver software Xpress 7.2.
237
The iterative approach of NE is displayed in Figure 2. In the first stage an EAP is solved to
238
generate a valid network for the log-trucks. Afterwards, the feasibility of this network is
239
checked by considering maximum tour length and time windows. If violations occur, cuts are
240
generated - which is the second stage - and added to the EAP. Additionally to the initially
241
used constraints of the EAP, the added cuts lead to a further limitation of the solution space.
242
After the addition of cuts, the first stage is repeated and the resulting network gets trimmed in
243
the direction of a feasible one. In the third stage the model formulation of the MDVRPPDTW
244
is solved on the created constrained network. If the MDVRPPDTW still cannot be solved on
245
the given network, a cut is added to the EAP that bans the actual solution. This procedure is
246
repeated until the algorithm finds a network the MDVRPPDTW is solvable on. The solution
10
247
248
The main advantages of this approach are the short computing times for solving the EAP and
249
the reduction of the number of variables for the MDVRPPDTW. On the one hand a number
250
of infeasible arcs can be excluded, because parts of the network cannot be reached by all
251
log-trucks. On the other hand the values of the starting and home rides of the log-trucks -
252
because they are already predetermined by the structure of the network - as well as all the arcs
253
254
The choices for the log-trucks of where to go next occur either at their depots H or at the
255
industrial sites S. Due to that structure, it is possible to solve a standard assignment problem
256
that assigns all transport tasks t T either to an industrial site s or a depot h in a way that
257
either the total empty-truck driving time or the fuel consumption is minimized. Additionally,
258
subcycles have to be avoided and each log-truck r R needs to reach its depot hr . If this can
259
be guaranteed, a valid network is obtained. However, this does not imply that the given
11
260
261
The EAP can be formulated by the following system of equalities and inequalities:
min
X X
DTst astr
(15)
sHS tT H rR
s.t.
X X
astr = 1
t T H
(16)
ahtr = 1
h H
(17)
sHS rR
X X
tT H rR
XX
ahr tr LBT
(18)
rR tT
X X
astr = |T Bs |
s S
(19)
S S
(20)
s S, r R
(21)
s H S, t T H, r R
(22)
tT H rR
X
t(T \
sS
X
S
sS
astr 1
T Bs )H rR
iHS:i6=s tT Bs
aitr =
asjr
jH{T \T Bs }
astr {0, 1}
262
The binary decision variable astr is equal to 1, if transport task t T or depot t H is either
263
264
s,t H, it follows that log-truck r is not used and stays at its associated depot hr . The starting
265
266
267
The objective function (15) minimizes the driving times DTst for the empty-truck rides. If the
268
objective is to minimize the fuel consumption, DTst is simply replaced by F Cst . This leads to
269
the following objective function (23), whereas the remainder of the problem stays the same:
min
X X
F Cst astr
(23)
sHS tT H rR
270
(16) ensure that all transport tasks t T are assigned to a depot s H or an industrial site
12
271
s S and that all log-trucks finish at their depot t = hr . A log-truck r either leaves its
272
depot hr to start with a task t T or stays at home at the depot t H (17). However, at
273
least a certain number of log-trucks LBT has to leave their depots to fulfill the tasks (18).
274
275
This minimum number of log-trucks equals LBT = dT OT /M T e, whereas TOT is the value of
P
the objective function of the actual iteration plus the sum of all task durations tT T Dt .
276
277
much transport tasks t T or home rides t H have to be started from an industrial site s
278
as belong to the industrial site s. T Bs is a subset of T and contains all transport tasks t that
279
deliver to s. (20) ensure that no subcycles occur. Therefore, it has to hold that there is at
280
(21)
least one task t assigned to s S S that does not deliver to industrial sites within S.
281
assure that each log-truck entering the transport tasks of an industrial site also has to leave
282
again. Finally, (22) are the binarity constraints for the given problem.
283
Solving the system of linear equations and inequations above leads to a valid network on which
284
an implementation of the MDVRPPDTW can be carried out. Before solving it, some
285
feasibility checks are done to avoid later violations concerning maximum tour length and time
286
windows. However, as no tours are constructed yet, there is no guarantee that all violations are
287
eliminated by the introduced cuts. Different types of cuts are used and added to the EAP and
288
then the EAP is solved again. If no cuts need to be added at this stage, the MDVRPPDTW is
289
solved. This problem is either feasible, which leads to a solution that is the global optimum, or
290
infeasible. If the MDVRPPDTW is infeasible, cuts that ban the given solution are added to
291
the problem and the EAP is solved again. A detailed description of the introduced cuts would
292
go beyond the scope of this paper, but can be found in Oberscheider et al. (2011).
293
Numerical studies
294
The proposed solution methods were used to solve different problem sizes based on real-life
295
data. Comparisons are done with respect to driving times and fuel consumptions, whereas
296
from the latter the CO2 e emissions are calculated. From August 2009 to February 2011 daily
297
tours of five log-trucks were tracked and stored in a database. From this given information a
298
road network consisting of 99 wood storage locations, 4 biomass power plants and 14 depots
299
for the log-trucks was modelled. According to the given data, biomass power plants are used as
300
industrial sites in the numerical studies. The biomass power plants are located in Austria and
13
301
all the selected depots and wood storage locations are situated within a distance between 2 and
302
138 km away from the power plants (Figure 3). For further information about the situation
and potential of biomass power plants in the investigated area, see Rauch et al. (2010).
Figure 3: Distribution of depots, wood storage locations and biomass power plants in the test area
303
304
The required matrices are derived from a road network that is implemented in a geographic
305
information system (GIS) according to the road classes 0-7 of Holzleitner et al. (2010). The
306
average speeds per road class originate from Ganz et al. (2005), in which data were collected
307
for another region of Austria. The driving times of the recorded data and the data gained by
308
simulations in GIS were compared and afterwards, the average speeds have been reduced by a
309
factor of 1.452. This factor is used to tackle the gap between observed and simulated values in
310
order to make the scenarios more realistic. Three different matrices are extracted from the GIS
311
data. The first one contains the minimum driving times from each node to every other
312
required node. The second and third have the same structure, but comprise fuel consumptions.
313
The second one is obtained for the fuel consumptions of empty-truck rides and the third one
314
for log-trucks with full loads. The global positioning system (GPS) data for generating the
315
input matrices for the algorithm, cover the rides between two locations only. It is assumed
14
316
that the engine of the log-truck is turned off at the wood storage location. Fuel consumptions,
317
due to reversing the log-trucks at the locations or for unloading at the industrial sites are not
318
considered.
319
To choose the correct formula out of the ones that are given in the annex of EEA (2009), input
320
factors are determined. Besides the load of 0 % and 100 %, the type of truck is set to a truck
321
and trailer with 34-40 tonnes maximum weight. EURO III is chosen as emission standard
322
according to the year of manufacturing of the tracked log-trucks. As no digital elevation model
323
(DEM) is used for the GIS data, an overall road gradient of 0 % is taken. Equation (24) is
324
used for calculating the fuel consumptions for empty-truck rides and full-truck rides in [g/km].
f c = + ev + ev
(24)
325
The setting of the parameters ( ) for the different loads can be seen in Table 1. The values
326
of the parameters as well as equation (24) are taken from the annex of EEA (2009).
Table 1: Parameter settings for the calculation of fuel consumptions
Load
0%
100 %
547.36
634.79
0.055
0.029
1,836.32
476,141.21
0.43
1.41
174.44
215.04
327
The received matrices for the fuel consumptions are converted from grams of diesel per path to
328
liters of diesel, before used in the algorithms. This is done in the way it is described in the
329
section about CO2 e calculation. Figure 4 gives an overview of how much the emissions per
330
kilometer vary for the valid range of driving speeds of the given formula (6 to 86 km/h)
331
332
Three different sizes of daily datasets have been used, whereas in all three cases 20 instances
333
have been generated. The smallest dataset contains scenarios with 15 tasks and 5 log-trucks.
334
It is used to validate the NE approach by a comparison to the solution of the standard model
335
formulation of the MDVRPPDTW. Instances with 30 tasks and 10 log-trucks were generated
336
337
Additionally, 20 instances with 60 tasks per day and 20 log-trucks were taken for the tests of
338
the NE.
15
Figure 4: CO2 e emissions [g/km] on roads with a gradient of 0 % of a Euro III truck and trailer with
34-40 tonnes maximum weight depending on its load and driving speed [km/h]
339
For the large instances the home location of log-truck 5 was taken as central depot, due to its
340
geographical location. 7 log-trucks have their origin there. The wood storage locations were
341
picked randomly out of the given 99, whereas the likelihood of a transport to a certain biomass
342
power plant is in relation to its demand for wooden chips in bulk stacked cubic meter (BCM)
343
per year (Table 2), as found in Rauch & Gronalt (2010). Due to the random choice, more than
one transport task can have its origin at a certain wood storage location.
Table 2: Yearly demand of wooden chips of the power plants
Plant
1
2
3
4
Demand in BCM
150,000
150,000
97,000
600,000
Likelihood in %
15.0
15.0
9.7
60.2
344
345
Besides the already mentioned input, the times for loading and unloading are needed. The
346
measurements resulted in durations of 55 minutes for loading at the wood storage location and
347
37 minutes for unloading at the power plant. Power plants can receive deliveries between 7 am
348
and 7 pm. At the earliest, drivers may start at 5 am. The latest feasible arrival at the depot in
349
the evening is at 9 pm [0, 960]. In between this time window, drivers are allowed to have a
350
maximum operation time of 480 minutes. As operation time, driving times are counted only,
16
351
352
Results
353
The following subsections describe the results of the numerical tests. The first one contains the
354
validation of the NE by a comparision to the solution of the standard model formulation that
355
is generated with solver software XPress 7.2 with standard settings. This was done with small
356
problem instances and the objective to minimize the driving time. The following one shows
357
the results for the objective of minimizing the driving time and minimizing the fuel
358
consumption or CO2 e emission, respectively. Additionally, the trade-offs of using one or the
359
360
The NE ran till it either terminated or could not find the global optimal solution after a
361
runtime of 3,600 seconds for the medium sized instances and 7,200 seconds for the large
362
instances. After this duration, the driving time or fuel consumption of the current solution was
363
recorded, respectively. This value serves as lower bound (LB) for the comparison with the
364
actual solution, which can be gained by raising the lower bound of the number of needed
365
log-trucks LBT by one. The deviation from the LB to the solution of the NE equals the
366
maximum deviation from the global optimal solution. The heuristic approach of adding
367
log-trucks in the NE is taken in order to get feasible solutions in reasonable computation times.
368
For all the instances, for which run times are reported, the tests were performed on a single
369
workstation with an Intel Core i7 with 2.8 GHz and 6.00 GB RAM, which runs on
370
371
372
The problem size for solving the model formulation of the MDVRPPDTW with the MIP
373
solver XPress 7.2 using standard settings is restricted, due to computation time. Therefore,
374
instances with 15 tasks and a maximum of 5 log-trucks have been chosen. 19 out of the 20
375
tested instances have the same solution values for both methods. For one instance, the MIP
376
solver did not finish within the maximum computation time of 43,200 seconds. In this case a
377
gap of 7.64 % from the lower bound of 984.51 minutes of driving time to the best found integer
378
solution of 1,066 minutes was recorded. The solution value of the NE of this instance equals
17
379
380
The computation times for the NE as well as for the MIP solver vary markedly. The range for
381
the NE goes from 0.1 seconds to 27,072.1 seconds. The latter shows an exceptional higher
382
computation time than the rest of the instances, due to a large number of iterations that had
383
to be performed. However, also the computation times for the instances that were solvable to
384
optimality by the MIP solver vary from 4.1 seconds to 15,322.8 seconds. The median of the
385
computation time required to solve the instances by NE is 7.5 seconds, whereas for the MIP
386
solver it is 20.3 seconds. It follows that for the given instances the NE is the faster method. It
387
388
389
In Table 3 a summary of the results, which can be found in detail in the Appendix, is given.
390
The means of the presented fuel consumptions and driving times are given for the empty-truck
391
rides only, as they are minimized by the NE. For medium-sized instances of both objectives
392
the NE found the global optimal solution for 16 out of 20 instances. For the remaining
393
instances one log-truck was added after 3,600 seconds of runtime. They also have similar
394
395
From the large instances with the scenario of minimizing driving times, Instance 1 could be
396
solved in a provable exact way only. Out of the remaining 19 instances, 14 have been solved by
397
adding one log-truck, while for five of them the number of log-trucks had to be increased by
398
two. In terms of minimizing fuel consumption, 4 globally optimal solutions could be obtained.
399
For 12 instances it was sufficient to add one truck, whereas for 4 instances two trucks had to
400
be added to the given problem. Runtimes show higher variations as for medium-sized
401
402
Scenario
Mean
FC
[l]
Mean
DT
[min]
Global
opt. found
[no.]
Max. DEV
from LB
[%]
Mean
Trucks
[no.]
DT M
FC M
DT L
FC L
310.8
308.7
585.2
581.2
1,791.2
1,830.3
3,365.0
3,445.3
16
16
1
4
3.10
3.57
2.94
2.97
8.3
8.4
16.4
16.5
Trucks
added
1
2
Range of
runtimes [s]
Min
Max
4
4
14
12
2.7
2.2
594.0
149.4
0
0
5
4
3,603.4
3,607.0
15,183.6
16,853.9
Mean
runtime
[s]
924.7
935.7
9,052.6
7,883.6
For a comparison of the objectives of minimizing fuel consumption and minimizing driving
18
403
time, different input matrices for driving times and fuel consumptions have to be computed.
404
This follows the fact that also the shortest paths between two points within the network may
405
change, due to different objectives. Therefore, the fastest way does not have to be the most
406
efficient one in relation to fuel consumption. Hence, the comparison leads to even higher
407
deviations than an optimization that uses the same input matrices for both scenarios. This is
408
not true for empty-truck rides only, but also for transport tasks with fully loaded log-trucks.
409
Therefore, empty-truck rides and full-truck rides have to be aggregated before the solutions of
410
the two scenarios are compared to each other. The main focus of the comparison is on reducing
411
total CO2 e emissions by changing objectives. Additionally, it is important to report the change
412
of total driving times, as this is the main interest of drivers and freight forwarders. A summary
413
of these results is presented in Table 4, whereas detailed results can be found in the Appendix.
Table 4: Total CO2 e emissions versus total driving times of medium (M) and large (L) instances for
the objectives of minimizing driving time (DT) and minimizing fuel consumption (FC)
Scenario
M
L
2,781.1
5,396.5
14.4
24.7
DT extension [min]
Mean
SD
3,745.1
7,229.6
81.6
170.2
3,663.6
7,059.4
36.1
50.7
414
For the medium sized instances, the exchange of the objective function leads on the one hand
415
to reduced CO2 e emissions, but on the other hand to higher driving times for all tested
416
instances. The average reduction of CO2 e is 33 kg and the extension of the driving time has a
417
mean value of 81.6 minutes. As stated in Table 3 approximately 8 log-trucks have to be used.
418
Hence, changing the objective to minimizing CO2 e emissions leads to an average increase of
419
approximately 10 minutes per driver and a mean reduction of roughly 4 kg CO2 e per tour.
420
According to the tests with large instances, Instance 9 is an outlier (see Appendix), due to a
421
reduced driving time, when minimizing fuel consumption. The reason for this is that two
422
log-trucks had to be added for Instance 9 to solve it within the maximum computation time
423
regarding the minimization of driving times. The remaining instances showed the expected
424
behavior of reduced total CO2 e emissions and increased total driving time. By using the
425
objective of minimizing CO2 e emissions, a mean reduction of 62.3 kg is obtained. The average
426
427
the medium sized instances the CO2 e emissions decrease by approximately 4 kg CO2 e per
428
19
429
Discussion
430
Based on the comparison of the results of the NE and the MIP solver for small instances it can
431
be concluded that the NE is a promising method for solving problems with the given structure.
432
For the tested instances it was enough to raise the lower bound of trucks by a maximum of 10
433
% of the available log-trucks. In order to save computation time it is possible to raise this
434
lower bound already at the beginning of the computation. This is an option for medium and
435
large instances where a solution could not be found within the given maximum time. Hence, if
436
used in praxis, the algorithm should be started parallel with additional log-trucks on one hand
437
and the attempt to solve the problem exactly on the other hand. Thereby, fast solutions can be
438
obtained. In our test case this would lead to a maximum computation time of 2,453.9 seconds
439
for Instance 20 of the large instances with the objective of minimizing fuel consumption.
440
It is likely that some of the presented solutions, for which the global optimal solution cannot
441
be guaranteed, are solved with their global optimum. This arises from the fact that a rather
442
straight forward approach is used to set the lower bound of log-trucks. For example, for solving
443
Instance 7 of the medium sized instances in terms of minimizing driving time, the lower bound
444
of log-trucks is set to eight. The division of the lower bound of the solution by the maximum
445
tour length leads to a minimum number of log-trucks of 7.98. Thus, in the opinion of the
446
authors, it is unlikely that eight log-trucks are sufficient to solve the given problem, because
447
the average driving time of a log-truck would be 478.9 minutes. As single routes are indivisible,
448
an assignment of 30 tasks to eight log-trucks within the maximum tour length is improbable.
449
By using two different objective functions, the potential savings of CO2 e emissions compared
450
to an optimization in terms of driving times were presented. In general, it is the opinion of the
451
authors that the biggest part of possible savings is achieved by the use of a decision support
452
system per se, even if it optimizes driving times instead of fuel consumptions. If optimized in
453
terms of fuel consumption, the potential reduction of approximately 4 kg CO2e emissions per
454
driver and day is bought by an extension of the driving time of approximately 10 minutes per
455
driver and day. Specifications in terms of road gradients would increase the accuracy of the
456
results, but these data were not available. Additionally, factors as the driving behavior or
457
maintenance of the trucks play a crucial role for actual emissions, but are hard to indicate.
458
In Figure 4 the impact of driving speeds on CO2 e emissions is shown. In the given range the
459
emissions per kilometer are decreasing with increasing driving speeds. Therefore, it can be
20
460
advantageous for the vehicle to choose a road with a road class that has a higher average
461
speed, even if this leads to rising driving times of the truck due to longer distances. Especially
462
at driving speeds below 25 km/h the slope becomes steep, which leads to comparatively high
463
increases of CO2 e emissions per kilometer. Additionally, to the choice of roads with higher
464
465
travel speeds would be another interesting input factor for increasing the accuracy of the given
466
results even though they are less relevant in timber transport, since it takes place in rural
467
468
The given problem implies full-truck loads, whereas the model formulation of the LTSP of
469
Palmgren et al. (2003) allows more than one pick-up and/or delivery per route. Therefore, it
470
would be interesting to use the findings of minimizing emissions of Bektas & Laporte (2011) on
471
that problem formulation. By using the objective of minimizing greenhouse gas emissions, the
472
sequence of nodes in a route is also depending on how much is loaded or unloaded at the
473
corresponding node, respectively. As shown in Figure 4 a higher load factor leads to higher
474
475
perform deliveries with higher weights first, even if this leads to longer driven distances. Vice
476
477
478
significant reduction of CO2 e emissions compared to the use of the objective of minimizing
479
driving times. Therefore, the used approach gives an idea of how to implement the
480
minimization of greenhouse gas emissions in timber transport in research or daily business and
481
introduces a powerful method to optimize the routing of log-trucks. Further tests will be
482
performed on available instances (e.g. Hirsch (2011) and Zazgornik et al. (2012)) to validate
483
484
Acknowledgments
485
486
providing very important input in terms of data, the authors want to thank Franz Holzleitner
487
of the Institute of Forest Engineering of the University of Natural Resources and Life Sciences,
488
Vienna.
21
489
References
490
491
492
493
494
495
Defra & DECC (2011). 2011 guidelines to Defra / DECCs GHG conversion factors for
496
company reporting. Tech. Rep. 1.2, Department for Environment, Food and Rural Affairs,
497
498
499
500
Demir, E., Bektas, T. & Laporte, G. (2011). A comparative analysis of several vehicle emission
models for road freight transportation. Transportation Research Part D, 16 (5), 347-357.
EEA (2009). EMEP/EEA air pollutant emission inventory guidebook 2009 - technical
501
502
503
Eglese, R. & Black, D. (2010). Optimizing the routing of vehicles. In: McKinnon A, Cullinane
504
505
506
507
approach to the log-truck scheduling problem. Annals of Operations Research, 184 (1),
508
163-178.
509
Favreau, J. (2006). Six key elements to reduce forest transportation cost. Paper presented at
510
511
http://www.forac.ulaval.ca/fileadmin/docs/EcoleEte/2006/Favreau.pdf, accessed 20
512
December 2011.
513
514
22
515
516
programming and tabu search for routing logging trucks. Computers and Operations
517
518
Ganz, M., Holzleitner, F. & Kanzian, C. (2005). Energieholzlogistik Karnten - Transport von
519
Energieholz [Wood fuel logistics Carinthia - transport of wood fuel]. Tech. rep., Institute of
520
Forest Engineering, University of Natural Resources and Life Sciences, Vienna, Austria.
521
Gronalt, M. & Hirsch, P. (2007). Log-truck scheduling with a tabu search strategy. In:
522
Doerner KF, Gendreau M, Greistorfer P, Gutjahr WJ, Hartl RF, Reimann M (eds)
523
Metaheuristics - progress in complex systems optimization, New York, USA: Springer, 65-88.
524
Hirsch, P. (2011). Minimizing empty truck loads in round timber transport with tabu search
525
526
(2), 15-41.
527
Holzleitner, F., Kanzian, C. & Stampfer, K. (2010). Analyzing time and fuel consumption in
528
road transport of round wood with an onboard fleet manager. European Journal of Forest
529
530
ICF Consulting (2006). Assessment of greenhouse gas analysis techniques for transportation
531
532
533
534
535
Murphy, G. (2003). Reducing trucks on the road through optimal route scheduling and shared
log transport services. Southern Journal of Applied Forestry, 27 (3), 198-205.
Ntziachristos, L. & Samaras, Z. (2000). COPERT III computer programme to calculate
536
emissions from road transport - methodology and emission factors (version 2.1). Tech. Rep.
537
538
Oberscheider, M., Zazgornik, J., Henriksen, C. B., Gronalt, M. & Hirsch, P. (2011). A
539
near-exact solution approach for solving a vehicle routing problem in timber transport,
540
working paper.
541
Palmgren, M., R
onnqvist, M. & V
arbrand, P. (2003). A solution approach for log truck
542
scheduling based on composite pricing and branch and bound. International Transactions in
543
544
Piecyk, M., McKinnon, A. & Allen, J. (2010). Evaluating and internalizing the environmental
545
546
Logistics - Improving the environmental sustainability of logistics, London, UK: Kogan Page
547
Limited, 69-97.
548
549
Rauch, P. & Gronalt, M. (2010). The terminal location problem in a cooperative forest fuel
supply network. International Journal of Forest Engineering, 21 (2), 32-40.
550
Rauch, P., Gronalt, M. & Hirsch, P. (2010). Co-operative forest fuel procurement strategy and
551
its saving effects on overall transportation costs. Scandinavian Journal of Forest Research,
552
25 (3), 251-261.
553
Ropke, S. & Pisinger, D. (2006). An adaptive large neighborhood search heuristic for the
554
pickup and delivery problem with time windows. Transportation Science, 40 (4), 455-472.
555
Sbihi, A. & Eglese, R. W. (2007). Combinatorial optimization and green logistics. Computers
556
557
558
559
24
560
Appendix
561
Notations
H
Set of depots
Set of log-trucks
S
S
Subset of S
T Bs
Set that contains the tasks that end in/belong to industrial site s
MT
, , , ,
562
ai
bi
fc
DTij
F Cij
F Ct
Fuel consumption of transport task t: The fuel consumption from the wood
storage location w to the industrial site s of transport task t
hr
LBT
SIs
SWw
T Dt
Task duration of transport task t: The driving time from the wood storage
location w to the industrial site s of transport task t
T OT
Equals the total driving time (empty-truck and full-truck rides) of a solution
astr
eir
etr
str
tijr
25
563
564
List of abbreviations
BCM
CH4
Methane
CO
Carbon monoxide
CO2
Carbon dioxide
CO2 e
COPERT
DECC
Defra
DEM
EAP
EEA
EN
European standard
EVRP
GHG
Greenhouse gas
GIS
GPS
LB
Lower bound
LTSP
MDVRPPDTW
Multi depot vehicle routing problem with pick-up and delivery and time
windows
NE
N2
Nitrogen gas
N2 O
Nitrous oxide
N Ox
Nitrogen oxides
PDPTW
PM
Particulate matter
PRP
Pollution-routing problem
TDVRSP
VRP
26
27
1,760
1,817
1,533
1,805
1,929
1,895
1,847
1,943
1,588
1,794
1,767
2,003
1,792
1,756
1,668
1,939
1,548
1,690
1,805
1,749
3,225
3,300
3,189
3,195
3,207
3,366
3,451
3,278
3,229
3,322
3,263
3,612
3,229
3,488
3,364
3,234
3,253
3,574
3,238
3,312
L1
L2
L3
L4
L5
L6
L7
L8
L9
L10
L11
L12
L13
L14
L15
L16
L17
L18
L19
L20
[min]
LB
M1
M2
M3
M4
M5
M6
M7
M8
M9
M10
M11
M12
M13
M14
M15
M16
M17
M18
M19
M20
Instance
3,225
3,335
3,219
3,236
3,232
3,438
3,523
3,325
3,324
3,370
3,310
3,660
3,276
3,537
3,459
3,264
3,299
3,623
3,285
3,359
1,760
1,817
1,533
1,805
1,929
1,895
1,894
1,943
1,637
1,794
1,767
2,003
1,843
1,756
1,668
1,939
1,596
1,690
1,805
1,749
0.00
1.06
0.94
1.28
0.78
2.14
2.09
1.43
2.94
1.44
1.44
1.33
1.46
1.40
2.82
0.93
1.41
1.37
1.45
1.42
0.00
0.00
0.00
0.00
0.00
0.00
2.54
0.00
3.09
0.00
0.00
0.00
2.85
0.00
0.00
0.00
3.10
0.00
0.00
0.00
553.0
576.7
557.7
564.9
557.9
601.2
618.3
584.3
566.1
586.1
569.5
643.2
571.8
600.5
608.7
571.5
572.5
634.9
568.3
596.3
298.2
313.1
265.2
317.6
336.7
326.0
331.8
335.7
278.8
313.8
302.8
347.4
323.8
298.8
291.2
342.6
284.3
286.9
310.0
311.6
16
16
16
16
16
17
17
16
17
16
16
17
16
17
17
16
16
17
16
16
8
8
7
8
9
9
9
9
8
8
8
9
9
8
8
9
8
8
8
8
594.0
7,370.3
7,388.0
15,183.6
7,341.5
14,542.4
15,099.1
7,286.3
14,508.7
8,930.9
7,387.3
7,602.4
8,784.8
7,444.0
14,452.0
7,301.1
7,519.2
7,347.7
7,604.4
7,364.4
840.0
2.7
33.8
4.2
14.2
10.6
3,603.4
19.4
3,603.2
5.2
120.7
183.1
3,603.3
4.0
43.4
28.9
3,603.4
680.0
2,085.7
5.0
Runtime
[s]
549.5
568.1
548.2
556.5
550.6
590.3
606.3
570.1
548.7
576.1
563.8
634.9
559.5
588.4
594.2
563.7
556.0
627.3
555.8
582.3
297.6
310.5
261.0
312.7
333.4
324.1
328.3
334.0
269.0
312.1
299.4
344.1
312.6
295.1
289.6
339.9
273.0
283.5
307.3
308.8
556.8
571.8
551.5
559.8
561.0
596.8
612.8
577.4
548.7
593.2
572.2
634.9
566.8
588.4
603.3
568.5
563.3
627.3
572.2
598.0
297.6
310.5
261.0
312.7
333.4
324.1
328.3
334.0
278.6
312.1
308.7
344.1
312.6
295.1
289.6
339.9
282.2
292.8
307.3
308.8
1.33
0.65
0.60
0.59
1.89
1.10
1.07
1.28
0.00
2.97
1.49
0.00
1.30
0.00
1.53
0.85
1.31
0.00
2.95
2.70
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
3.57
0.00
3.11
0.00
0.00
0.00
0.00
0.00
3.37
3.28
0.00
0.00
3,337
3,393
3,302
3,301
3,327
3,529
3,604
3,403
3,270
3,496
3,429
3,776
3,336
3,577
3,549
3,325
3,374
3,691
3,383
3,504
1,791
1,854
1,560
1,851
1,961
1,923
1,938
1,977
1,655
1,830
1,845
2,053
1,846
1,807
1,693
1,987
1,626
1,799
1,828
1,782
16
16
16
16
17
17
17
16
15
17
17
17
16
16
17
16
16
17
17
17
8
8
7
8
9
9
9
9
8
8
9
9
8
8
8
9
8
9
8
8
7,583.3
8,183.1
7,393.6
7,305.2
14,460.1
7,425.6
7,439.5
7,308.2
1,195.6
14,557.7
8,136.2
914.4
7,992.0
1,643.6
7,307.9
9,129.2
7,357.0
149.4
15,336.2
16,853.9
3.5
5.4
1,564.9
2.2
4.2
627.0
10.3
5.6
3,603.1
124.3
3,602.6
389.4
558.0
2.6
2.9
110.0
3,603.1
3,607.0
881.7
6.1
LB
Results
6,767
6,976
6,825
6,743
6,999
7,268
7,232
6,968
6,862
6,949
7,121
7,712
6,838
7,348
7,218
6,940
6,829
7,585
6,966
7,042
3,670
3,768
3,248
3,754
3,962
3,975
3,972
3,988
3,381
3,610
3,880
4,160
3,767
3,653
3,540
4,069
3,268
3,832
3,824
3,581
DT: Total
CO2 e
[kg]
6,920
7,093
6,997
6,906
7,196
7,485
7,423
7,122
6,858
7,153
7,340
7,915
7,043
7,504
7,344
7,121
7,021
7,746
7,138
7,267
3,628
3,694
3,199
3,653
3,892
3,901
3,878
3,925
3,322
3,537
3,743
4,080
3,709
3,547
3,494
3,951
3,204
3,638
3,763
3,513
FC: Total
CO2 e
[kg]
2.26
1.68
2.52
2.42
2.81
2.99
2.64
2.21
-0.06
2.94
3.08
2.63
3.00
2.12
1.75
2.61
2.81
2.12
2.47
3.20
1.16
2.00
1.53
2.76
1.80
1.90
2.42
1.61
1.78
2.06
3.66
1.96
1.56
2.99
1.32
2.99
2.00
5.33
1.62
1.94
5,080
5,262
5,244
5,140
5,411
5,640
5,548
5,358
5,109
5,309
5,475
5,976
5,205
5,500
5,522
5,354
5,184
5,826
5,313
5,473
2,681
2,730
2,421
2,766
2,881
2,926
2,937
2,934
2,447
2,650
2,806
3,058
2,749
2,598
2,635
3,009
2,463
2,765
2,818
2,688
DT versus FC
DEV
DT: Total
DT
[%]
[min]
5,059
5,192
5,173
5,083
5,376
5,553
5,471
5,294
5,036
5,294
5,423
5,907
5,142
5,411
5,473
5,258
5,097
5,725
5,276
5,440
2,697
2,778
2,450
2,803
2,940
2,967
2,982
2,962
2,463
2,676
2,825
3,084
2,814
2,641
2,657
3,056
2,483
2,786
2,840
2,716
FC: Total
DT
[min]
0.42
1.35
1.39
1.12
0.65
1.57
1.40
1.21
1.47
0.28
0.95
1.17
1.22
1.64
0.89
1.82
1.71
1.77
0.70
0.59
0.60
1.75
1.19
1.34
2.06
1.43
1.54
0.94
0.65
0.98
0.67
0.86
2.36
1.67
0.82
1.54
0.82
0.77
0.81
1.06
[%]
DEV