Computers and Operations Research 129 (2021) 105211

Contents lists available at ScienceDirect

Computers and Operations Research

journal homepage: www.elsevier.com/locate/caor

Time-dependent multi-depot green vehicle routing problem with time

windows considering temporal-spatial distance
Houming Fan a,⇑, Yueguang Zhang a, Panjun Tian a, Yingchun Lv a, Hao Fan b
a
College of Transportation Engineering, Dalian Maritime University, Dalian 116026, China
b
School of Transportation and Logistics, Dalian University of Technology, Dalian 116024, China

a r t i c l e i n f o a b s t r a c t

Article history: Reducing distribution costs is one of the effective ways for logistics enterprises to improve their core
Received 3 April 2020 competitiveness. Aiming at the multi-depot vehicle routing problem under the time-varying road net-
Revised 1 January 2021 work, this paper proposes an integer programming model with the minimum total costs by comprehen-
Accepted 2 January 2021
sively considering the fixed costs of vehicles, penalty costs on earliness and tardiness, fuel costs and the
Available online 05 January 2021
effects of vehicle speed, load and road gradient on fuel consumption. A hybrid genetic algorithm with
variable neighborhood search is developed to solve the problem. In the algorithm, the temporal-spatial
Keywords:
distance is introduced to cluster the customers to generate an initial population to improve the quality
Multi-depot green vehicle routing problem
Time-varying road network
of the initial solution. Adaptive neighborhood search times strategy and simulated annealing inferior
Time windows solution acceptance mechanism are used to balance the diversification and exploitation in the algorithm
Temporal-spatial distance iteration process. Numerical results show that the model and algorithm we proposed are rather effective.
Hybrid genetic algorithm with variable The research results not only deepen and expand the vehicle routing problem (VRP) theory research, but
neighborhood search also provide a scientific and reasonable method for logistics enterprises to make the vehicle scheduling
plan.
Ó 2021 Elsevier Ltd. All rights reserved.

1. Introduction portation costs and low service levels for a single depot to serve
customers (Fan et al., 2019). In recent years, energy depletion
With the rapid development of the logistics industry, sharing and the greenhouse effect have become increasingly prominent.
distribution resources and reducing vehicle energy consumption Researchers have begun to consider factors such as fuel consump-
has become the main trend of logistics distribution. At the same tion and carbon emissions that harm the environment in VRP, and
time, the development of an intelligent transportation system pro- research GVRP (Yu et al., 2019; Poonthalir and Nadarajan, 2018). In
vides the possibility to obtain traffic information about road net- existing GVRP literature, the models for calculating fuel consump-
works. Therefore, some researchers have researched the multi- tion can be divided into two categories. One assumes that it is only
depot vehicle routing problem (MDVRP), green vehicle routing related to vehicle type, load, and distance between two nodes, and
problem (GVRP), and time-dependent vehicle routing problem has nothing to do with vehicle speed (Rezaei et al., 2019). The other
(TDVRP). We focus on the multi-depot green vehicle routing prob- is to consider the impact of vehicle speed on fuel consumption (Liu
lem (MDGVRP) with the time-varying vehicle speed and soft time et al., 2019; Xiao and Konak, 2016). Compared with MDVRP,
windows, which is related to the MDVRP, GVRP and TDVRP. research on MDGVRP needs to consider more factors, which
With the continuous expansion of the scale of logistics enter- increases the complexity of the problem and the difficulty of solv-
prises, they have multiple depots in the service area, and informa- ing it. Therefore, researchers have successively proposed effective
tion and orders can be shared among the depots. The traditional methods to solve MDGVRP (Jabir et al., 2017; Li et al., 2018).
research on the vehicle routing problem is no longer applicable In previous researches on VRP, researchers usually assumed
to this distribution model, prompting researchers to study MDVRP that the vehicle speed was constant. However, the traffic environ-
(Salhi et al., 2014; Sadati et al., 2020). The multi-depot distribution ment is dynamically changing, which limits the speed of vehicles,
mode can optimize the distribution scheme through global infor- so that vehicles can only travel slowly during rush hours. In
mation, which can effectively solve the problems of high trans- 1981, Beasley (1981) proposed the TDVRP and studied the VRP in
which the travel time changes with the departure time. Subse-
⇑ Corresponding author. quent research on MDVRP mostly used the time-dependent func-
E-mail address: [email protected] (H. Fan). tion based on travel speed proposed by Ichoua et al (2003), and

https://doi.org/10.1016/j.cor.2021.105211
0305-0548/Ó 2021 Elsevier Ltd. All rights reserved.
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

express the vehicle speed as a piecewise function (Mu et al., 2015; search to solve it. Fan et al. (2019) considered the timeliness
Figliozzi, 2012). The Federal Highway Administration (2015) men- requirements of fresh product distribution and proposed a
tioned that the speed of vehicles changes smoothly, rather than a MDOVRP based on joint distribution mode of fresh food, con-
step-change at a certain moment. Xu et al. (2019) proposed to structed an optimization model aiming at minimizing the total dis-
use a trigonometric function to describe the change of vehicle tribution costs, and designed an ant colony optimization (ACO) to
speed, which is closer to the reality of distribution production solve it. Contardo and Martinelli (2014) proposed an optimization
activities. model for MDVRP under capacity and route length constraints, and
The time-dependent multi-depot green vehicle routing problem a new exact algorithm was adopted to minimize the total travel
with time windows (TDMDGVRPTW) considering temporal-spatial time. Salhi et al. (2014) developed a variable neighborhood search
distance proposed in this paper is based on the historical traffic to solve the heterogeneous fleet MDVRP.
information of the distribution network, and considering the
impact of the fixed costs of vehicles and the temporal-spatial dis- 2.2. Research on GVRP
tance between customers on the total costs, which is a VRP expan-
sion problem that optimizes vehicle routing in the multi-depot In recent years, GVRP has attracted the attention of researchers.
distribution network. The contribution of this paper can be sum- Rezaei et al. (2019) proposed a mixed ILP model considering the
marized as: impact of vehicle type, load, and distance on fuel consumption. A
Considering the constraint of time-dependent travel time, fixed GA and population-based simulated annealing were developed to
costs of vehicles, time window, fuel consumption, and the impact minimize distribution costs and carbon emissions. Aiming at the
of vehicle type, speed, load, and road gradient on fuel consumption, integrated scheduling of production and distribution, Ganji et al.
an integer programming model is proposed for this problem. (2020) proposed a multi-objective MIP model considering the
A developed hybrid genetic algorithm with variable neighbor- heterogeneous fleet, time window, and the relationship between
hood search is applied to solve the presented problem. In the algo- fuel consumption and vehicle type, speed and load, and solved it
rithm, the initial population is generated by the temporal-spatial with three heuristic algorithms. Some researchers have begun to
clustering of customers, and an adaptive neighborhood search study the GVRP of alternative fuel-powered vehicles (AFVs) that
strategy and a simulated annealing inferior solution acceptance may need to visit alternative fuel stations during delivery
mechanism are introduced. (Bruglieri et al., 2019; Andelmin and Bartolini, 2017; Zhang et al.,
The rest of this article is as follows: Section 2 briefly reviews the 2020; Koç and Karaoglan, 2016). Bruglieri et al. (2019) considered
relevant literature. Section 3 describes the problem and formulates the linear relationship between fuel consumption and distance,
a mathematical model. In Section 4, the hybrid genetic algorithm and developed a path-based solution approach to solve the pro-
with variable neighborhood search is developed. Section 5 tests posed problem. Andelmin and Bartolini (2017) modeled GVRP as
the algorithm and summarizes the experimental results. Section 6 a set partitioning problem and designed an exact algorithm to
presents the conclusions. solve it. Zhang et al. (2020) considered the capacity limit of the fuel
tank and the relationship between the travel distance and carbon
emissions, and a two-stage ant colony system was adopted to solve
2. Literature review
the MDGVRP to minimize carbon emissions. Koç and Karaoglan
(2016) developed a heuristic-based exact solution approach to
The research of vehicle routing problem with time windows
solve the proposed GVRP.
(VRPTW) is relatively mature. In this paper, the research on
MDVRP, GVRP, TDVRP, and time-dependent green vehicle routing
2.3. Research on TDVRP
problem (TDGVRP) is mainly reviewed.
Considering the time dependence of vehicle speed in VRP can
2.1. Research on MDVRP effectively avoid traffic congestion and shorten delivery time.
Haghani and Jung (2005) considered the dynamic changes of travel
Researchers have done extensive research on the optimization time between nodes and customer demand, formulated a TDVRP
model and solution approach of the MDVRP. Karakatič and optimization model with the minimum delivery time as the goal,
Podgorelec (2015) summarized various genetic algorithms (GA) and developed a GA to solve it. Wu and Ma (2017) considered
that are designed for solving MDVRP and evaluated their perfor- the time-varying characteristics of the distribution network, pro-
mance. Bezerra et al. (2018) developed a general variable neigh- posed an optimization model for the TDVRP of perishable food
borhood search (GVNS) algorithm to solve the proposed MDVRP, with minimum total costs, and designed a hybrid genetic algo-
its goal is to minimize the costs of optimized vehicle routing. rithm to solve it. Taniguchi and Shimamoto (2004) studied the
Oliveira et al. (2016) proposed a decomposition method for MDVRP advanced intelligent transportation system used in urban distribu-
and designed a cooperative coevolutionary algorithm to solve it. tion, proposed an optimization model for VRP with the time-
Ray et al. (2014) developed a heuristic algorithm to solve the estab- varying vehicle speed, and dynamically updated travel time using
lished ILP optimization model of the MDVRP including depot selec- dynamic traffic simulation data. Mancini (2014) considered traffic
tion and route optimization. Afshar-Nadjafi and Afshar-Nadjafi congestion during rush hours, proposed to use multiple functional
(2014) considered the influence of departure time on travel time expressions to express the change of road speed in one day, and
between nodes and developed a heuristic algorithm to solve the developed greedy randomized adaptive search procedures (GRASP)
formulated mixed integer programming (MIP) model. Soto et al. to solve time-dependent vehicle routing and scheduling problem
(2017) developed a general multiple neighborhood search hybri- (VRSP). Zhang and Zhang (2017) considered the time dependence
dized with a tabu search (MNS-TS) strategy to solve the multi- of vehicle travel time, proposed a heterogeneous fleet TDVRP,
depot open vehicle routing problem (MDOVRP). Liu and Ma and adopted an improved genetic algorithm (IGA) to solve the
(2013) proposed a MIP model for MDOVRP with time window problem to minimize the total costs. Duan et al. (2019) considered
based on vehicle leasing and sharing and developed a hybrid both the stochastic and time-varying of the distribution network
genetic algorithm (HGA) to solve the problem. Li et al. (2016) con- and designed a non-dominated sorting ant colony optimization
sidered the constraints of time window to propose a model with (NSACO) to solve the formulated optimization model of stochastic
the minimum travel costs and developed a HGA with adaptive local TDVRP. Considering that the travel time of nodes is affected by
2
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

traffic congestion and other factors, Sabar et al. (2019) designed a tabu search algorithm. Wang et al. (2019) proposed a bi-
self-adaptive evolutionary algorithm for solving the proposed objective model aiming to minimize total carbon emission and
dynamic VRP with traffic congestion. operating costs for the MDGVRP with shared transportation
resources and developed a hybrid heuristic algorithm to optimize
2.4. Research on TDGVRP vehicle routing.
The above-mentioned related literature is of great reference
With further research on GVRP and TDVRP, researchers began value for the research of this paper. Table 1 summarizes the main
to study TDGVRP. Soysal and Çimen (2017) formulated a model contributions already proposed in the literature. Through combing
for TDGVRP, which assumes that fuel consumption depends on dis- the above-related literature and Table 1, we can summarize: (1)
tance and time-dependent speed. A simulation based restricted Most of the related researches on TDVRP express the vehicle speed
dynamic programming approach (RDP) was proposed to optimize in a certain period time by the average speed, ignoring the smooth
the vehicle routing with minimum costs. Liu et al. (2020) proposed increase or decrease of vehicle speed, which is inconsistent with
a model for VRP under time-varying distribution network, which the actual distribution production activities. (2) Most of the
assumes that carbon emissions depend on vehicle speed, vehicle researches on MDVRPTW, GVRPTW, and TDVRPTW generate the
load, and other factors. An improved ACO was developed to solve initial solution randomly or cluster customers in space without
the proposed problem. Poonthalir and Nadarajan (2018) adopted considering the influence of time, which leads to poor quality of
the objective planning method for the GVRP and designed a parti- initial solution and poor performance of algorithm optimization.
cle swarm optimization to minimize the route costs and fuel con- (3) There is a lack of research on the optimization of vehicle rout-
sumption. Xu et al. (2019) proposed a GVRPTW with time-varying ing considering four aspects: time-dependent vehicle travel time,
speed, formulated an optimization model considering the impact multiple depots, time windows, and fuel consumption. In the exist-
of vehicle load and speed on fuel consumption, and designed a ing research, we only found that Wang et al. (2019) considered
heuristic algorithm for the solution. Xiao and Konak (2016) pro- these four features at the same time, but it assumed that the speed
posed a MIP model for time-dependent green VRSP in which emis- of the vehicle follows a normal distribution, and the average speed
sions were assumed to depend on vehicle type, speed, and load. A at the beginning and end of a certain period is used as the speed of
hybrid algorithm was developed to solve it. Çimen and Soysal the vehicle during the period, without considering the smooth
(2017) developed a heuristic algorithm to solve the proposed increase or decrease of the vehicle speed (Federal Highway
TDGVRP with stochastic vehicle speeds. Qian and Eglese (2015) Administration, 2015; Xu et al., 2019), and ignore the temporal
formulated an optimization model of VRP under time-varying and spatial differences of vehicle speed caused by the different
speed to minimize fuel consumption and solved problems with a types of roads in the distribution network. Besides, external envi-

Table 1
A summary of the selected literature.

Reference Problem features Objectives Solution

TW MD TD Factors in fuel
consumption/carbon
emissions
Liu and Ma (2013) * *
Total costs HGA
Afshar-Nadjafi and * * *
Total costs Constructive heuristic algorithm
Afshar-Nadjafi
(2014)
Qian and Eglese (2015) * * Speed and waiting time Fuel emissions Column generation based tabu search
Li et al. (2016) * *
Distance HGA with adaptive local search
Mancini (2014) * *
– GRASP
Xiao and Konak (2016) * * Vehicle type, load, and CO2 emissions Hybrid optimization
speed
Wu and Ma (2017) * *
Total costs HGA
Soysal, and Çimen * Vehicle type, speed, and Total costs Simulation based RDP
(2017) distance
Çimen, and Soysal * Vehicle type, speed, and CO2 emissions ADP-based heuristic
(2017) distance
Poonthalir and * Speed and distance Route costs and fuel consumption Particle swarm optimization
Nadarajan (2018)
Fan et al. (2019) * *
Total costs ACO
Rezaei et al. (2019) * Load and distance transportation costs and CO2 GA and population based simulated annealing
emissions
Bruglieri et al. (2019) AFVs: distance Distance Path-based solution
Duan et al. (2019) * *
Number of vehicles and travel NSACO
time
Sabar et al. (2019) *
Distance Self-adaptive evolutionary algorithm
Liu et al. (2019) * * Speed, load and gradient Carbon emissions Improved ACO
Xu et al. (2019) * * Speed, load, the road, and Fuel consumption and Improved NSGA-Ⅱ
the vehicle conditions satisfaction
Wang et al. (2019) * * * Speed, vehicle parameters, Carbon emissions and operating Multi-phase hybrid heuristic algorithm
distance costs
Ganji et al. (2020) * Speed, distance, vehicle Costs, tardiness of orders and NSGAII, MOPSO, and MOACO
type customer’s dissatisfaction
Zhang et al. (2020) * Alternative fuel-powered Carbon emissions Two-stage ant colony system
vehicles: distance
This paper * * * Speed, load, vehicle type, Total costs: fixed vehicle costs, Hybrid genetic algorithm with variable neighborhood
and gradient fuel costs, and penalty costs search considering temporal-spatial distance

3
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

ronmental factors such as aerodynamic drag coefficient and air relations. The trigonometric expression between speed (v ) and
density, and many vehicle parameters such as engine friction fac- time (t) can be expressed as follows:
tor and efficiency of the vehicle drive train are considered in the 8
>
> a0 sin½b0 ðt  c0 Þ þ d0 ; t 2 ½0; T 1 
calculation model of carbon emissions. For different types of vehi- >
>
>
> ..
cles, these parameters are different and will change constantly in >
> .
< 
  
the process of distribution, which is difficult to determine and
v ¼ ab sin bb t  cb þ db ; t 2 T b ; T bþ1
>
ð1Þ
has poor practicability in distribution and production activities. >
>
> ..
>
> .
>
>
:
3. Problem and mathematical model an sin½bn ðt  cn Þ þ dn ; t 2 ½T n ; 24
The parameters ab ; bb ; cb ; db ; b 2 f1; 2; :::; ng are determined by
3.1. Problem description road conditions.
According to Yang’s (2018) research on the evolution of urban
This paper is a multi-depot vehicle routing problem. The depot recurrent congestion, the trend of increasing or decreasing vehicle
can be open to all vehicles and accept the final stop of vehicles. The speed on weekdays can be obtained, as presented in Fig. 1.
customer node only accepts the delivery service provided by the The whole day is divided into a plurality of periods according to
vehicle and does not accept the final stop of the vehicle. The prob- the changing trend of vehicle speed, and the functional relation-
lem can be described as: The road network constitutes a complete ship between vehicle speed (v ) and the time (t) is different in each
directed graph G ¼ ðV; EÞ, there are different types of roads, and the period. Assuming that the time T i when the vehicle leaves node i is
vehicle speed of each type of road changes continuously. At a cer- within½T b ; T bþ1 , there are two possibilities for the vehicle to travel
tain moment, the vehicle leaves the depot to visit the customer and from node i to node j, i.e. across the period or not. If
provides service to the customer within the service time window RT
lij 6 T ib þ 1 v ðtÞdt, the vehicle arrives at the node j within ½T i ; T bþ1 
required by the customer. There will be penalty costs for vehicles
arriving early or late. Vehicles can return to any depot after visiting without across the period, the travel time t ij can be obtained by cal-
the last customer, and there is no path connection between the culating the upper limit of integration according to the speed func-
RT
depots. The fuel consumption during the delivery process varies tion relation of the period; if lij > T ib þ 1 v ðtÞdt, the vehicle needs to
with factors such as vehicle speed and load. Table 2 summarizes across periods, assuming that the vehicle travels from node i to
the sets, parameters, and variables used in the model.

node j, the travel time is tij ¼ T bþM1  T i þ tbþM
ij and across M
b bþ1 bþM
periods, the distance traveled in each period is lij ; lij ;   ; lij ,
3.2. Determination of time-dependent function of vehicle speed
and the time t bþM
ij traveled in period M can be obtained by calculat-
ing the upper limit of integration according to the speed function
Most of the existing researches on TDVRP used a step function
to express the vehicle speed of the road section, and the vehicle relation of the period.
speed is abrupt at a certain time. In reality, the vehicle speed varies
continuously, such as the trigonometric function relation 3.3. Calculation of the amount of fuel consumption
v ðtÞ ¼ usinðctÞ þ d (which u; c; d are related to road conditions)
This paper uses the MEET model including the correction coef-
between the vehicle speed (v ) and the time (t) (Xu et al., 2019).
ficients for the road gradient and vehicle load given by Hickman
In this paper, the variation of vehicle speed on all-day roads is
et al. (1999) to calculate vehicle carbon emissions. According to
approximately expressed by a plurality of trigonometric function
the relevant research of Alinaghian and Naderipour (2016) on
3.5 t–7.5 t vehicles, combined with the above-mentioned speed
Table 2 time-dependence function, can use the integral idea to obtain the
Nomenclature of sets, parameters, and variables. fuel consumption F ij of vehicles traveling from the node i to node j.
Notation Meaning Z Tj 
 
V V ¼ D [ C is nodes set, where D is the set of depots and C F ij ¼ k
110 þ 0:000375v 3 þ 8702=v
GC
LC v dt ð2Þ
Ti
represents the set of all customers
E E ¼ fði; jÞji; j 2 V; i–jg is edge set In this equation, k ¼ 0:00043L=g, GCis the road gradient correc-
lij The distance between nodes i andj (Unit: km)
tion coefficient, LC is the load gradient correction coefficient. The
v The speed of the vehicle (Unit: km/h)
F ij The fuel consumption of the vehicle from nodes i toj (Unit: liter) calculation formula of GC and LC is:
K The set of all vehicles
k Any vehicle in the vehicle setK
Q The capacity of the vehicle (Unit: ton)
½T s ; T f  Time window of the depot, where T s indicates the time when the
vehicle leaves the depot and T f is the latest time the vehicle
returns to the depot
½ET i ; LT i  Time window of customeri
di The demand of the customer i (Unit: ton)
sik Handle time of vehicle k at nodei
T ik The time when the vehicle k arrives at the nodei
c1 The fuel price per liter (Unit: CNY/liter)
c2 The fixed costs of a vehicle (Unit: CNY)
c3 Penalty cost per unit time for vehicles arriving earlier than ET i
(Unit: CNY/h)
c4 Penalty cost per unit time for vehicles arriving later than LT i (Unit:
CNY/h)
xijk If the vehicle k reaches the node j from the node i, xijk ¼ 1,
elsexijk ¼ 0
yij If the customer i is served by the depot j, yij ¼ 1, elseyij ¼ 0
Fig. 1. Vehicle speed-time dependence function.

4
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

 
GC ¼ exp 0:0059v 2  0:0775v þ 11:936 n ð3Þ restricts the time when the vehicle returns to the depot from
exceeding the working deadline of the depot. Eqs. (13) and (14)
LC ¼ 0:27x þ 1 þ 0:0614nx  0:0011n3 x  0:00235v x are decision variables.

 ð0:33=v Þx ð4Þ
4. Solution method
In Eq. (3), n is the road gradient in terms of percentage, In Eq.
(4), x is the ratio of vehicle load to vehicle capacity which is a digit The genetic algorithm searches for satisfactory solutions by
between [0, 1]. simulating the natural evolution process in biology. It can evaluate
multiple feasible solutions in the search space at the same time
3.4. Mathematical model and has better robustness and global search performance. How-
ever, the algorithm has the disadvantages of slow convergence
The mathematical optimization model of this paper is formu- speed and easy falling into local optimization. The variable neigh-
lated to minimize the total costs, as follows: borhood search algorithm (VNS) uses multiple different neighbor-
XXX XXX hood structures for systematic search and has a strong local search
minZ ¼ c1 xijk F ij þ c2 xijk capability. In this paper, customers are clustered according to the
i2V j2V k2K i2D j2C k2K
XXX spatial–temporal distance between customers. Combining the
þ c3 xijk maxfðET j  T jk Þ; 0g advantages of GA and VNS, a hybrid genetic algorithm with vari-
i2V j2C k2K
XXX able neighborhood search considering the temporal-spatial dis-
þ c4 xijk maxfðT jk  LT j Þ; 0g ð5Þ tance (HGAVNS_TS) is developed to solve the proposed problem.
i2V j2C k2K The pseudo-code of this algorithm is given below:
s.t
XX Algorithm (HGAVNS_TS)
xijk dj 6 Q ; 8k 2 K ð6Þ
Parameters
i2V j2C

XXX d pop size: population size;

xijk 6 jK j ð7Þ
d MAX gen: maximum number of iterations;
i2D j2C k2K
d N k ¼ fN 1 ; N 2 ; :::; N l g: neighborhood structure N l is the
XX XX lth neighborhood structure;
xijk ¼ xjik ¼ 1; 8j 2 C ð8Þ
i2V i2V
d Sn : adaptive neighborhood search times;
k2K k2K
d max N: maximum number of neighborhood cycles;
XX XX 1. Initialize PðtÞ; % using a genetic algorithm to cluster
xijk ¼ xjik 6 1; 8k 2 K ð9Þ
i2D j2C i2D j2C
customers in temporal-spatial
2. gen ¼ 0;
X 3. whilegen < MAX gen
yij ¼ 1; 8j 2 C ð10Þ
i2D
4. evaluate PðtÞ; %
5. select Pðt þ 1Þ from PðtÞ; % elitist preservation + ro
XX X
xijk 6 jSj  1; 8k 2 K; jSj ¼ xijk ; 8i 2 D; k 2 K ð11Þ ulette wheel selection
i2S j2S j2C 6. reproduce pairs in Pðt þ 1Þ; % order crossover
7. fori ¼ 1 : pop size
XX XX
Ts þ t ij xijk þ sik xijk 6 T f ; 8k 2 K ð12Þ 8. forj ¼ 1 : max N
i2V j2V i2V j2C 9. Individual Pi ðt þ 1Þ disturbed from the first
neighborhood structureN 1 , iter 1;
xijk 2 f0; 1g; 8i 2 V; 8j 2 V; 8k 2 K; andfi 2 Dg \ fj 2 Dg ¼ £ 10. if p P p0 % p0 is the acceptance probability of
the new solution
ð13Þ
11. P i ðt þ 1Þ P0i ðt þ 1Þ;
12. break
yij ¼ f0; 1g; 8i 2 D; 8j 2 C ð14Þ
13. else iter iter + 11;
Eq. (5) represents the objective function, which aims at mini- 14. end
mizing the total costs including the fixed costs of vehicles, fuel 15. until (iter ¼ Sn );
costs, and time window penalty costs. Eq. (6) represents the capac- 16. The individual Pi ðt þ 1Þ continues to be
ity limit of the vehicle. In reality, the number of vehicles owned by disturbed by the next neighborhood structure N l ,
depots is limited. In this paper, the number of vehicles is consid- iter 1;
ered when optimizing vehicle routing. Eq. (7) indicates that the 17. repeat
number of vehicles leaving the depot does not exceed the total 18. until (iter ¼ Sn );
number of vehicles that can be used to ensure that the feasible 19. end
solution meets the actual requirements of logistics enterprises. 20. end
Eq. (8) indicates that each customer is visited only once by one 21. gen ¼ gen þ 1;
vehicle, and it is the vehicle arrival and departure balance con- 22. end
straint. Eq. (9) indicates that the vehicle k may not be selected to 23. Best solution
leave the depot to visit customers. If the vehicle k is selected to
perform the delivery task, the vehicle k can only leave a certain
depot iði 2 DÞ to visit customers and can return to any depot. That In this paper, the genetic algorithm is adopted to generate the initial
is, the vehicle k can return to a depot different from the initial one. solution by clustering customers in temporal-spatial, and then a
Eq. (10) indicates that each customer can only be served by one hybrid genetic algorithm with variable neighborhood search
depot. Eq. (11) is to eliminate the sub-loop constraint. Eq. (12) (HGAVNS) is adopted to solve this problem. Taking an instance with
5
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

customer size of n as an example, if the average total travel time of Eq. (16) is employed to calculate the temporal-spatial distance
each feasible solution is t time and the integration step is p. Then between customers (Qi et al., 2012).
the complexity of each routine is estimated as follows: (1) In step 2 3 2 3 2 3
       
4, the number of times to invoke ‘‘evaluate p(t)” is MAX gen, and lij ¼ a1 4lij  min lmn 5=4 max lmn  min lmn 5 þ a2 4lij  min lmn 5=
ST S S S S T T

its time complexity is OðT 1 Þ ¼ MAX gen  pop size  ðt time=pÞ; (2) m;n2C
m–n
m;n2C
m–n
m;n2C
m–n
m;n2C
m–n
2 3
The time complexity of ‘‘select” is OðT 2 Þ ¼ MAX gen; (3) The time    
complexity of ‘‘order crossover” isOðT 3 Þ ¼ MAX gen  pop size 4 max lT  min lmn 5;
T
a1 þ a2 ¼ 1;i; j 2 C
mn
m;n2C m; n 2 C> m–n
ðt time=pÞ; (4) The complexity of variable neighborhood search is m–n

OðT 4 Þ ¼ MAX gen  pop size  max N  Sn  ðt time=pÞ. The time com- ð16Þ
plexity of the proposed algorithm is OðTÞ ¼ OðT 1 Þ þ OðT 2 Þ þ OðT 3 Þþ
(2) Customer clustering and initial solution construction
OðT 4 Þ ¼ Oð½MAX gen  pop size  max N  Sn  ðt time=pÞÞ
According to the pseudo-code of this algorithm and the above
In this paper, a genetic algorithm is used to cluster customers
analysis, the complexity of variable neighborhood search is the
according to the temporal-spatial distance to minimize the sum
main component of the complexity of the proposed algorithm.
of the temporal-spatial distance from the customers in each cluster
iði ¼ 1; 2;    ; uÞ to the cluster center (Qi et al., 2014). Eq. (17) is the
objective ffunction.
4.1. Generation of the initial population
X
u X ST
MinG ¼ lij ; 8i; j 2 C ð17Þ
(1) Temporal-spatial distance i¼1 j assigned to i

The existing researches on VRPTW mostly consider the time The algorithm in this paper adopts the integer encoding
window and spatial distance of customers separately when clus- method. Set up a virtual depot 0, and divide the vehicle routing
tering customers, ignoring one of the factors, which leads to the according to the solution strategy of ‘‘routing first and then group”
higher penalty costs or transportation costs of the generated initial with the vehicle capacity as the constraint. The process is as
solution. This paper considers the two-dimensional distribution follows:
characteristics of the customer’s time window and spatial distance
when generating the initial population. 1) Select the customer closest to the virtual depot and assign
In this paper, Wang et al. (2018) method of measuring the tem- available vehicles. After visiting this customer, visit other
poral distance between customers is used to quantify the proxim- customers in the cluster in turn according to the principle
ity of customers’ handle time. Assume that the time windows of of temporal-spatial distance from near to far. For customers
customers i and j are ½ET i ; LT i ,½ET j ; LT j , and ET i < ET j . The time that exceed the vehicle capacity constraint, the residual
when the vehicle k arrives at the customer i is T ik 2 ½ET i ; LT i , then capacity of vehicles delivered by adjacent clusters is sequen-
T jk is within½ET i þ thik þ t ij ; LT i þ t hik þ tij . Let ½ET i þ t hik þ tij ; LT i þ tially detected according to the principle of temporal-spatial
t hik þ tij  be ½a; b, there are four relationships between ½ET j ; LT j  distance from near to far to meet the customer demand, and
and ½a; b, as shown in Fig. 2. if there is a cluster that meets the customer demand, it is
Eq. (15) is employed to calculate the temporal distance lij
T added to the cluster; otherwise, divide the customer into a
new cluster.
between customer i and customer j (Wang et al., 2018).
2) Check whether there are any unassigned clusters and if so,
go to step 1).
8 3) After the assignment task division of all clusters is com-
>
> c3 ðET j  bÞ; b < ET j
>
> pleted, calculate the distance between the first and last cus-
< c4 ða  LT j Þ ; a > LT j
T tomers and each depot in the routing of each vehicle.
lij ¼ ð15Þ
>
>
> c5 ðthik þ lij =v Þ; a  ET j < b or a < LT j < b 4) According to the principle of minimum distance, the depot is
>
:
t hik þ lij =v ; ET j < a < b < LT j inserted between the first and last customers and the virtual
depot to obtain the initial solution.

In Eq. (15), c5 is the deviation coefficient. The temporal distance 4.2. Fitness function
T T
is directional, i.e. lij –lji , but the distance is not directional in the
clustering process, so the larger of the two is taken as the time dis- The fitness function of chromosomes can be constructed
tance between customer i and customer j. according to Eq. (5). The fitness function of the chromosome R
can be expressed as follows:
1
fR ¼ ð18Þ
zR
ETi LTi ETj LTj
In Eq. (18), Z R is the objective function value of the chromo-
(1) someR. The smaller the objective value of a chromosome, the
a b
greater it’s fitness, and the greater the probability of being selected.
(2)
a b 4.3. Select operation
(3)
a b
The selection operation adopts the strategy of combining elite
a b reservations with the roulette wheel selection. Specific steps are
as follows: using the roulette wheel selection method, the proba-
(4)
ETi LTi ETj a b LTj bility of each individual being selected is proportional to its fitness
value, i.e. the higher the fitness value, the higher the probability of
Fig. 2. Temporal distance diagram. the individual being selected. And the improved method of the
6
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

elite reservation is adopted. Firstly, an individual A with the high- rithm. In this paper, a variable neighborhood search is applied to
est fitness value in the population is stored in the temporary the local search strategy to strengthen the depth search of the
matrix. After the neighborhood search is completed, if the optimal population.
solution has not changed, the similarity between the individual A
and the sub-generation individuals is calculated and the sub- (1) Neighborhood structure
generation A1 with the highest similarity with the individual A is 1) Insert: Randomly select a customer i from the chromo-
replaced. The variable neighborhood search strategy adopted by some and insert it randomly after the customer j. As
the algorithm in this paper may still achieve the level of the indi- shown in Fig. 4(a), customer 3 is inserted behind cus-
vidual A for the sub-generation A1 by further depth search. In con- tomer 6.
trast, although the fitness value of the optimal individual B in the 2) Exchange: Randomly select two customers i and j, then
sub-generation is lower than that of the individual A, it has some exchange their positions. As shown in Fig. 4(b), exchange
good gene sequences and can provide better genes in the popula- the positions of customer 3 and customer 6.
tion crossing of the next generation. Compared with the sub- 3) 2-OPT: Randomly select customers i andj, and exchange
generation A1 , it has more value of continuous iteration. the order of other customers among customer points. As
shown in Fig. 4(c), the position of customer 3 is kept
unchanged, and customers 4, 5, 7, and 6 are in reverse
4.4. Evolutionary operation
order.
(2) Adaptive mechanism and solution acceptance
The evolutionary operation in this paper adopts an order cross-
over operator. In accordance with Fig. 3, when order crossover is
In this paper, we design an adaptive neighborhood search times
performed on the parent A, the parent B is randomly selected from
strategy and a new solution acceptance rule to balance the breadth
the population. Firstly, points i11 ,i12 ,i21 and i22 are randomly gener-
and depth required for evolution, so that the algorithm jumps out
ated respectively, the part between i11 and i12 of the parent A is
of the local optimum. The disturbance intensity required by the
taken as the first segment of the sub-generation A1 and the subse-
population is different at different stages of the algorithm. The
quent points of the sub-generation A1 are related to the parent B,
strategy of adaptive neighborhood search times in this paper is
i.e. the customer points between i11 and i12 in the parent B are
as follows:
eliminated first. In the elimination process, the position sequence
of the customer points in the parent B is not changed, and then 
the eliminated customer points are arranged as the second seg- gen
Sn ¼ b1 þ b2  ð19Þ
ment of the sub-generation A1 to form the sub-generation A1 and MAX gen
the sub-generation B1 . The order crossover operator used in this
In Eq. (19), Sn is the number of search times when the algorithm
paper is unidirectional, that is, it only accepts the operation in a
runs to the gen generation, b1 is the minimum number of searches,
good direction.
b2 is the adaptive number of searches, b c represents round down.
It can be seen from Eq. (19) that the search times gradually
4.5. Local search strategy increase with the increase of iteration number, which increases
the depth search of the population.
The local search strategy is one of the core functions of the algo- To further improve the perturbation of the population and
rithm, which determines the local search capability of the algo- expand the search space, this paper uses the acceptance rule of

i11 i12

A 1 3 5 6 7 4 8 2 9 A1 5 6 7 4 3 1 9 8 2

B 3 4 1 9 8 2 7 6 5 B1 4 1 9 8 2 7 3 5 6

i21 i22

Fig. 3. The diagram of order crossover operator.

3 4 5 7 6 2 8 4 5 7 6 3 2 8

(a) Insert

3 4 5 7 6 2 8 6 4 5 7 3 2 8

(b) Exchange

3 4 5 7 6 2 8 3 6 7 5 4 2 8

(c) 2-OPT

Fig. 4. Neighborhood structures.

7
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

the solution in simulated annealing algorithm to accept the poor and time windows on the solution result. The structure of Section 5
solution with a certain probability to improve the algorithm’s opti- is shown in Table 3.
mization ability. Eq. (20) is the probability of new solution
acceptance.
5.1. Test the performance of the algorithm
(
1; f R ðxÞ 6 f R ðx0 Þ
p¼   ð20Þ
MDVRP and MDVRPTW are the basic problems of this paper,
f R ðxÞf R ðx0 Þ
exp gen
; f R ðxÞ > f R ðx0 Þ this paper selects MDVRP and MDVRPTW instances to test the
algorithm. The algorithm is programmed with MATLAB2018b in
the Windows10 system, the running memory of the computer is
4G, the CPU is Intel(R)Core(TM)i7-7700, and the dominant fre-
5. Computational experiment and analyses quency is 3.60 GHz. Instances of testing algorithm performance
cover different customer sizes, customer distributions, and time
To test the performance of the proposed algorithm, Section 5.1 window sizes, which requires that different parameters should
uses HGAVNS_TS to solve the basic problem of this paper, and the be set during algorithm operation. Firstly, the function of parame-
solution results are compared with other heuristics. Section 5.2 ters in the algorithm is analyzed theoretically, and then the influ-
designs the instance of this problem and uses HGAVNS_TS to solve ence of parameter values on the running results is analyzed.
it, and sets up five cases to analyze the influence of vehicle speed Finally, the parameter values are determined through repeated

Table 3
The structure of Section 5.

Section Content
5.1 Test the performance (1) Instance analysis of HGAVNS_TS was used to solve MDVRP, and the solution results were compared with adaptive large
of the algorithm MDVRP neighborhood search (ALNS 50 K) (Pisinger and Ropke, 2007) and hybrid genetic search with adaptive
diversity control (HGSADC) (Vidal et al., 2012) to verify the effectiveness of HGAVNS_TS in solving MDVRP.
HGAVNS_TS was used to solve MDVRPTW, and the solution results were compared with the variable
(2) Instance analysis of neighborhood search algorithm (VNS) and hybrid genetic algorithm with variable neighborhood search
MDVRPTW (HGAVNS). Besides, HGAVNS_TS is used to solve the instance of MDVRPTW designed by Zhong et al.
(2006), and compared with other heuristics to verify the effectiveness of HGAVNS_TS in solving
MDVRPTW.
5.2 Case study (1) Network description and Design the case of the problems proposed in this paper and set the relevant parameters.
parameters setting
(2) Optimization results Use HGAVNS_TS to solve the designed case.
(3) Sensitivity analyses Five cases are set to analyze the influence of vehicle speed and time windows on the solution results

Table 4
Comparison of HGAVNS_TS with other heuristics on MDVRP instances.

ALNS HGSADC HGAVNS_TS

Instance n d Pre-best
best Avg t/min best Avg t/min best Avg t/min Gap(%)
P01 50 4 576.87a 576.87 576.87 0.48 576.87 576.87 0.23 577.63 580.60 0.22 0.13
P02 50 4 473.53a 473.53 473.53 0.47 473.53 473.53 0.21 473.53 475.86 0.19 0.00
P03 75 5 641.19b 641.19 641.19 1.07 641.19 641.19 0.43 646.12 652.36 0.87 0.77
P04 100 2 1001.04d 1001.04 1006.09 1.47 1001.04 1001.23 1.94 1016.31 1027.06 1.92 1.53
P05 100 2 750.03b 751.26 752.34 2.00 750.03 750.03 1.06 754.76 758.81 1.42 0.63
P06 100 3 876.50c 876.70 883.01 1.55 876.50 876.50 1.14 895.91 897.11 1.57 2.21
P07 100 4 881.97d 881.97 889.36 1.47 881.97 884.43 1.55 901.36 907.87 2.04 2.20
P08 249 2 4372.78e 4390.80 4421.03 5.55 4372.78 4387.38 10.00 4465.30 4547.15 15.36 2.12
P09 249 3 3858.66e 3873.64 3892.50 6.02 3858.66 3873.64 9.50 3921.52 4094.31 13.83 1.63
P10 249 4 3631.11e 3650.04 3666.85 6.05 3631.11 3650.04 9.82 3687.84 3744.58 14.22 1.56
P11 249 5 3546.06d 3546.06 3573.23 5.95 3546.06 3546.06 7.14 3622.31 3689.53 13.56 2.15
P12 80 2 1318.95c 1318.95 1319.13 1.25 1318.95 1318.95 0.52 1319.46 1324.06 0.83 0.04
P13 80 2 1318.95c 1318.95 1318.95 1.00 1318.95 1318.95 0.57 1325.63 1331.28 1.37 0.51
P14 80 2 1360.12b 1360.12 1360.12 0.97 1360.12 1360.12 0.55 1376.32 1381.61 1.31 1.19
P15 160 4 2505.42b 2505.42 2519.64 4.22 2505.42 2505.42 1.92 2515.87 2543.17 6.14 0.42
Pr01 48 4 861.32b 861.32 861.32 0.50 861.32 861.32 0.17 864.63 869.25 0.20 0.38
Pr02 96 4 1307.34d 1307.34 1308.17 1.72 1307.34 1307.34 0.76 1310.95 1316.50 1.44 0.28
Pr03 144 4 1803.80e 1806.60 1810.66 3.57 1803.80 1803.80 1.91 1836.32 1890.12 3.97 1.80
Pr04 192 4 2058.31e 2060.93 2073.16 4.93 2058.31 2060.93 5.22 2088.10 2103.33 7.89 1.45
Pr05 240 4 2331.20e 2337.84 2350.31 6.20 2331.20 2337.84 9.56 2396.50 2419.33 12.66 2.80
Pr06 288 4 2676.30e 2687.60 2695.74 7.75 2676.30 2685.35 10.00 2793.31 2866.03 17.75 4.37
Pr07 72 6 1089.56b 1089.56 1089.56 0.97 1089.56 1089.56 0.34 1103.63 1109.22 0.76 1.29
Pr08 144 6 1664.85d 1664.85 1675.74 3.45 1664.85 1665.05 2.05 1679.03 1697.42 3.51 0.85
Avg. Gap 0.11 0.44 0.00 0.10 1.32 2.47

Source of instances: http://neo.lcc.uma.es/vrp/vrp-instances/multiple-depot-vrp-instances/.

Computer configurations used to test the algorithms: the ALNS runs on a Pentium Ⅳ 3.0 GHz; the HGSADC runs on an Opteron 2.4 GHz scaled for a Pentium Ⅳ 3.0 GHz.
a
Optimality proved.
b
Results of Cordeau et al. (1997).
c
Results of Renaud et al. (1996).
d
Results of Pisinger and Ropke (2007).
e
Results of Vidal et al. (2012).

8
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

experiments. After testing, the parameter settings are as follows: randomly generating an initial population and initializing the pop-
the maximum number of iterations max gen ¼ 50 when the ulation by clustering customers in temporal-spatial when solving
genetic algorithm generates an initial solution by temporal- Pr02. Fig. 6 shows the iterative process of three algorithms to solve
spatial clustering of customers, a1 ¼ 0:6, a2 ¼ 0:4, Pr02.
pop size ¼ 50  150, MAX gen ¼ 100  500, max N ¼ 30  100, Compared with the results of the MDVRP example, the time
b1 ¼ 1, b2 ¼ MAX gen=10  MAX gen=5, the new solution acceptance window constraint is considered in MDVRPTW, which increases
probability p ¼ 0:98  1. The larger the number of customers, the the time spent to obtain the optimal solution for the same cus-
larger the settings of MAX gen, pop size and max N. tomer size. Fig. 5 shows that the quality of the initial solution
can be improved by initializing the population by spatiotemporal
(1) Instance analysis of MDVRP clustering of customers. It can be seen from Table 5 and Fig. 6 that
HGAVNS_TS has good initial solutionquality and strong local
Table 4 shows the results of the ALNS 50 K, HGSADC and searchability, and the improved algorithm has better solution
HGAVNS_TS. The results of some instances obtained by ALNS and performance.
HGSADC are current the best solution. n is the number of cus- Zhong et al. (2006), Duan and Fu (2008), and Ling and Gu (2017)
tomers, d is the number of depots, Pre-best is the best-known solu- studied the multi-depot open vehicle routing problem with time
tions, best is the best solution obtained by the algorithm, Avg is the windows, which is the basic problem of this paper. To further ver-
average result of 10 runs of the algorithm, t is the time spent in one ify the performance of the algorithm for MDVRP with soft time
experiment (Unit: minutes), Gap is the gap between the best solu- window, HGAVNS_TS is adopted to solve the instance with 3
tion obtained by HGAVNS_TS and the Pre-best, and Avg_gap is the depots and 24 customers. The algorithm parameters are set as fol-
average gap between the average solution and the Pre-best. Table 4 lows: max gen ¼ 50, pop size ¼ 50, max N ¼ 60, b1 ¼ 1,
shows that in the results of the benchmark instances solved by b2 ¼ MAX gen=5, p ¼ 1. Table 6 summarizes the best solutions
HGAVNS_TS. The minimum value of Gap is 0.00% and the maxi- under different time window penalty costs obtained by genetic
mum is 4.73%. The average gap between the best solution and algorithm (GA) (Zhong et al., 2006), tabu search algorithm (TS)
the average solution relative to the best-known solution is 1.32%
and 2.47% respectively. In terms of algorithm running time, when
the number of customers is less than 100, the solution time of
the proposed algorithm is similar to that of ALNS and HGSADC,
but when the number of customers exceeds 100, the solution time
of the proposed algorithm increases, but it is acceptable. The effec-
tiveness of the HGAVNS_TS in solving the MDVRP problem can be
verified by comparing the calculation results and the running time
of the algorithm.

(2) Instance analysis of MDVRPTW

To verify the performance of the proposed algorithm for solving

MDVRP with soft time windows, this paper adopts the VNS,
HGAVNS and HGAVNS_TS to solve 12 MDVRPTW instances, and
compares and analyzes the results. PrA_B represents the first B cus-
tomers of PrA (for example, Pr04_48 represents the first 48 cus-
tomers of Pr04), VNS and HGAVNS randomly generate the initial
population, and VNS adopts three neighborhood structures: insert,
exchange and 2-opt. Allow vehicles to return to the nearest depot,
set vehicle speed v ¼ 1, the transportation cost per unit distance is
1.5, c3 ¼ 1, c4 ¼ 3, c5 ¼ 2. Table 5 summarizes the results of the
algorithm running 10 times, and Best is the best solution obtained
by the three algorithms. Fig. 5 is a quality comparison diagram of Fig. 5. The objective value of the initial solution of Pr02.

Table 5
Comparison of HGAVNS_TS with other heuristics on MDVRPTW instances.

VNS HGAVNS HGAVNS_TS

Instance n d Best
best Avg t/min best Avg t/min best Avg t/min Gap (%)
Pr01 48 4 4343.3 4612.8 5044.0 0.41 4423.0 4574.7 0.73 4343.3 4504.8 0.48 0
Pr11 48 4 2441.0 2657.7 2910.3 0.41 2450.2 2514.8 0.69 2441.0 2467.0 0.45 0
Pr05_48 48 4 5575.4 5931.1 5994.4 0.44 5768.7 5848.2 0.80 5575.4 5683.2 0.58 0
Pr08_48 48 6 4436.9 5273.7 5388.2 0.39 4648.1 4999.3 0.74 4436.9 4478.9 0.54 0
Pr07 72 6 5788.9 5967.0 6156.2 2.02 5970.0 6199.3 3.67 5788.9 5940.1 2.15 0
Pr17 72 6 2988.7 3910.4 4058.8 1.97 3034.9 3365.4 3.41 2988.7 3093.7 1.99 0
Pr06_72 72 4 7892.5 9307.2 9369.7 2.18 8063.6 8527.9 4.02 7892.5 8022.2 2.06 0
Pr08_72 72 6 5276.6 5580.6 5738.4 2.01 5368.4 5615.4 3.50 5276.6 5375.0 2.60 0
Pr02 96 4 7853.1 7902.3 8078.4 4.51 8073.1 8295.7 6.62 7853.1 7972.7 5.15 0
Pr12 96 4 3923.2 4726.3 4841.6 4.31 4363.9 4453.9 6.16 3923.2 4024.9 5.34 0
Pr06_96 96 4 11187.7 11716.9 11916.5 5.02 11214.0 11514.7 6.24 11187.7 11251.9 5.57 0
Pr03 144 4 12139.3 13679.4 13808.6 12.37 12857.4 13206.5 30.04 12139.3 12299.7 20.35 0
Avg_Gap (%)p 11.37 15.04 3.27 7.58 0 1.94

Source of instances: https://neo.lcc.uma.es/vrp/vrp-instances/multiple-depot-vrp-with-time-windows-instances/.

9
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

delivery process can effectively reduce the negative impact of

cargo transportation on the environment. The integer planning
model developed in this paper takes into account factors such as
multi-depots, time-dependent vehicle speed, time windows and
fuel consumption, which is closer to the actual delivery production
activities and ensures the sustainable transportation of goods. In
this paper, the proximity of temporal distance between customers
is quantified and combined with spatial distance to cluster cus-
tomers in the temporal-spatial distance. At the same time, the vari-
able neighborhood search strategy is introduced into the genetic
algorithm to enhance the searchability and improve the perfor-
mance of the algorithm. The performance of the proposed mathe-
matical model and heuristic algorithm to solve this problem is
verified and analyzed by numerical experiments.

(1) Network description and parameters setting

The benchmark instance Pr01 is selected and modified as the

instance of the proposed problem. The instance includes 4 depots
and 48 customers, and the node coordinate position is the same
as Pr01. The rest information of depots and customers is shown
Fig. 6. Iterative process for solving Pr02. in Table 7. Parameter settings are as follows: Q ¼ 5t, sik ¼ 0:5di ,
c1 ¼ 6:9, c2 ¼ 500, c3 ¼ 20, c4 ¼ 30, c5 ¼ 25, n ¼ 0. There are three
types of roads in the distribution network: urban main roads, sec-
(Duan and Fu, 2008), improved ant colony algorithm (IACO) (Ling
ondary roads and branch roads. As shown in Fig. 7, red represents
and Gu, 2017) and HGAVNS_TS. It can be seen from Table 6 that
the main roads, black represents secondary roads and undrawn
the minimum gap between the results obtained by GA, IACO, and
routes are branch roads. Fig. 8 shows vehicle speeds for three types
TS and the best solution obtained by HGAVNS_TS is 0.06%, and
of roads.
the maximum gap is 8.54%.

(2) optimization results

5.2. Case study
The proposed algorithm is used to solve the problem in this
With the development of business, some supermarket chains,
paper. The parameters of the algorithm are set as follows:
express companies and fast-moving consumer goods companies
max gen ¼ 50, pop size ¼ 100, MAX gen ¼ 150, max N ¼ 60,
have chosen to establish multiple depots in the service areas for
b1 ¼ 1, b2 ¼ MAX gen=5, p ¼ 0:99. The distribution scheme is
joint distribution in order to reduce the high costs caused by
shown in Table 8, and vehicle routing is shown in Fig. 9. To verify
long-distance delivery and improve the delivery timeliness. In real-
the effectiveness of considering temporal-spatial distance in this
ity, the traffic environment is dynamic, and traffic congestion may
experiment, HGAVNS is used to solve the problem under the same
make it impossible to provide services to customers within the
conditions. The iterative process of the two algorithms is shown in
expected time. Besides, considering the fuel consumption in the

Table 6
Results of comparison between HGAVNS_TS and GA, TS, IACO.

Parameters Best GA Gap (%) IACO Gap (%) TS Gap (%) HGAVNS_TS Gap (%) t/s
c3 ¼ 0:2,c4 ¼ 2 665.4 665.8 0.06 675.1 1.46 – – 665.4 0 3.05
c3 ¼ 0,c4 ¼ 2 612.5 – – 641.3 4.70 664.8 8.54 612.5 0 3.27

Source of instances: Zhong et al. (2006).

Table 7
Information about depots and customers.

Node di ET i LT i Node di ET i LT i Node di ET i LT i Node di ET i LT i

D1 6 18 10 0.9 7.5 16.5 23 0.7 7.5 16.5 36 0.5 8.5 14

D2 6 18 11 0.35 8.5 13.5 24 0.95 6.5 16.5 37 0.55 7.5 15.5
D3 6 18 12 0.3 7.5 16.5 25 0.7 8.5 13.5 38 0.75 7.5 16.5
D4 6 18 13 0.45 7.5 15.5 26 0.3 6.5 12 39 0.65 6.5 12
1 0.6 6.5 10 14 0.45 6.5 10.5 27 0.8 7.5 16.5 40 0.75 7.5 16.5
2 0.4 7.5 16.5 15 0.2 6.5 9.5 28 0.45 8.5 13.5 41 0.4 7.5 16.5
3 0.8 8.5 15.5 16 1.25 7.5 16.5 29 1 7.5 16.5 42 1.1 7.5 11.5
4 0.25 7.5 16.5 17 0.25 7.5 15.5 30 0.65 6.5 12 43 1.2 7.5 16.5
5 0.6 8.5 13.5 18 0.85 6.5 9.5 31 0.5 7.5 15.5 44 0.15 7.5 16.5
6 0.25 7.5 16.5 19 0.15 7.5 16.5 32 0.8 6.5 10.5 45 1.25 7.5 11.5
7 0.65 6.5 9.5 20 0.8 6.5 12 33 0.95 6.5 9.5 46 0.95 6.5 13.5
8 1 8.5 13.5 21 1.25 7.5 12 34 1.1 7.5 12 47 1.05 7.5 12
9 0.65 6.5 10 22 1.05 7.5 16.5 35 0.7 7.5 16.5 48 0.5 7.5 16.5

10
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

Fig. 7. Distribution network. Fig. 9. Vehicle routing corresponding to the best solution.

Fig. 8. The time-dependent function of speed. Fig. 10. Comparison of the iterative process.

Fig. 10. It can be seen from Fig. 10 that the objective value of the It can be seen from Table 8 that if the time window penalty cost
best individual in the initial population generated by the spa- is ignored, the total cost is 5040.66, that is, the opportunity loss
tiotemporal clustering of customers is better than that of the initial cost of the optimized scheme is 179.49. If the traveling speed of
population generated randomly, and the convergence algebra is vehicles is constant at 30 km/h, 40 km/h and 50 km/h, the corre-
reduced, thus accelerating the convergence speed. It is further sponding total costs of the delivery scheme are 5493.31, 4926.12
proved that the algorithm has a strong optimization ability. and 4642.15 respectively, and the deviations from those when con-

Table 8
The best distribution scheme.

Vehicle Distribution route Distance Costs Fuel costs Penalty costs

1 D1-9-42-46-24-12-47-D4 170.25 741.36 219.86 21.50
2 D3-20-33-13-8-5-29-D3 158.28 726.70 216.70 10.00
3 D4-39-43-21-40-38-D4 136.44 711.14 185.65 25.49
4 D4-30-2-31-14-1-4-19-28-16-D3 232.23 852.93 317.93 35.00
5 D2-22-27-3-6-48-11-45-D2 116.05 710.23 168.73 41.50
6 D1-35-32-44-37-7-34-10-D2 74.80 670.20 140.20 30.00
7 D3-26-18-17-15-25-23-26-41-D1 189.91 807.59 291.59 16.00
Summation 1077.96 5220.15 1540.66 179.49

11
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

Table 9
Parameter setting of sensitivity analysis.

Case1 Case2 Case3 Case4 Case5

Time windows ½ET i ; LT i  – ½ET i ; LT i  ½ET i ; LT i  –
Vehicle speed Time-dependent Time-dependent 30 40 40
Parameters c3 ¼ 20, c4 ¼ 30 – c3 ¼ 20, c4 ¼ 30 c3 ¼ 20, c4 ¼ 30 –

6000.0
5220.2
4928.5
5000.0 4530.9
4300.4 4208.6
4000.0 3500.0 3500.0 3500.0 3500.0 3500.0
3000.0

2000.0 1540.7 1428.5

1005.7
762.5 708.6
1000.0
179.5 0.0 25.2 37.9 0.0
0.0
Case 1 Case 2 Case 3 Case 4 Case 5

Fuel costs Penalty costs Vehicle fixed costs Total costs

Fig. 11. Comparison of sensitivity analysis results.

sidering all features are 5.23%, 5.63% and 11.07% respectively. time. Seeking new optimization strategies and obtaining better
Therefore, it is very important to consider these features when for- distribution schemes can effectively reduce distribution costs. In
mulating optimization models. this paper, the influence of the fixed costs of vehicles, traffic infor-
mation of distribution network and the temporal-spatial between
(3) Sensitivity analyses customers on the total costs are considered comprehensively, an
integer programming model to minimize the distribution costs is
In the previous MDVRP related studies, some researchers
constructed and develops a hybrid genetic algorithm with variable
ignored the time dependence of vehicle speed or did not consider
neighborhood search considering the temporal-spatial to solve the
the time window, and there may be some deviation between the
proposed problem. Numerical results show that the model and
solution result of the corresponding model and the actual cost.
metaheuristic we proposed are rather effective. Obtains the follow-
To analyze the influence of time window and the change of vehicle
ing conclusions:
speed on the formulation of the delivery scheme, five cases are set
up and the results of the delivery schemes solved under different
(1) The formulated model has fully considered the sharing of
conditions are compared and analyzed. Among them, Case1 is
customer and vehicle resources among multiple depots,
the numerical experiment of this paper, Case2 ignores the time
effectively improved the distribution efficiency, and further
window constraint, Case3 and Case4 ignore the time dependence
expanded and deepened the VRP research.
of vehicle speed, Case5 ignores the time window constraint and
(2) The model considers the smooth increase or decrease of the
the vehicle speed is 40 km/h, as shown in Table 9. The proposed
vehicle speed and the influence of vehicle speed, load, and
algorithm is used to solve the models in five cases, and the delivery
road gradient on fuel consumption. Although it increases
schemes and costs in different cases are obtained. The best solution
the complexity of the problem and the difficulty of solving
in 10 runs is shown in Fig. 11.
it, it is closer to the reality of distribution production
It can be seen from Fig. 11 that the cost of the scheme obtained
activities.
when ignoring the time window constraint is 4828.5, which is
(3) The designed hybrid genetic algorithm with variable neigh-
291.7 different from the delivery scheme obtained under Case1.
borhood search considering the temporal-spatial effectively
When the time dependence of vehicle travel time is ignored and
improves the quality of the solution by generating initial
the vehicle travel speed is constant at 30 km/h and 40 km/h, the
solutions through the temporal-spatial clustering of
cost of the delivery scheme obtained by the corresponding model
customers.
is different from the optimization result considering all features
(4) The selection operation adopts the strategy of combining
by 13.20% and 17.62%, respectively. The total cost difference
elite reservations with the roulette wheel selection ensures
between case3 and Case4 is 230.5, and the difference in fuel costs
the effective convergence of the algorithm. The designed
is 243.3. Comprehensive analysis shows that the change of vehicle
adaptive neighborhood search times strategy and simulated
speed and customer time window will have a great impact on the
annealing inferior solution acceptance mechanism can bal-
optimization results. Therefore, it is of great significance to study
ance the diversification and exploitation in the algorithm
the proposed problem in this paper.
iteration process.

6. Conclusions The research in this paper applies to the time-dependent multi-

depot green vehicle routing problem with time windows, which
In reality, more and more enterprises have multiple depots due can further reduce the distribution costs of logistics enterprises
to the integration of resources, and the dynamic changes of the and provide a theoretical solution for the multi-region joint distri-
traffic environment also affect the distribution of goods all the bution of logistics enterprises.

12
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

CRediT authorship contribution statement Technical report. Rue de la Loi 200. European Commission/DG VII, Brussels,
Belgium, p. 1049 (accessed 18 February 2020).
Ichoua, S., Gendreau, M., Potvin, J.-Y., 2003. Vehicle dispatching with time-
Houming Fan: Conceptualization, Methodology, Investigation, dependent travel times. Eur. J. Oper. Res. 144 (2), 379–396. https://doi.org/
Supervision. Yueguang Zhang: Writing - original draft, Data cura- 10.1016/S0377-2217(02)00147-9.
Jabir, E., Panicker, V.V., Sridharan, R., 2017. Design and development of a hybrid ant
tion, Software. Panjun Tian: Writing - review & editing, Validation.
colony-variable neighbourhood search algorithm for a multi-depot green
Yingchun Lv: Writing - review & editing, Validation. Hao Fan: vehicle routing problem. Transp. Res. Part D: Transp. Environ. 57, 422–457.
Writing - review & editing, Validation. https://doi.org/10.1016/j.trd.2017.09.003.
Karakatič, S., Podgorelec, V., 2015. A survey of genetic algorithms for solving multi
depot vehicle routing problem. Appl. Soft Comput. J. 27 (2), 519–532. https://
Declaration of Competing Interest doi.org/10.1016/j.asoc.2014.11.005.
Koç, Ç., Karaoglan, I., 2016. The green vehicle routing problem: A heuristic based
exact solution approach. Appl. Soft Comput. 39, 154–164. https://doi.org/
The authors declare that they have no known competing finan- 10.1016/j.asoc.2015.10.064.
cial interests or personal relationships that could have appeared Li, J., Li, Y., Pardalos, P.M., 2016. Multi-depot vehicle routing problem with time
windows under shared depot resources. J. Comb. Optim. 31 (2), 515–532
to influence the work reported in this paper. https://link.springer.com/article/10.1007/s10878-014-9767-4.
Li, J., Wang, R., Li, T., Lu, Z., Pardalos, P.M., 2018. Benefit analysis of shared depot
resources for multi-depot vehicle routing problem with fuel consumption.
Acknowledgment Transp. Res. Part D: Transp. Environ. 59, 417–432. https://doi.org/10.1016/j.
trd.2018.01.026.
This work was supported by the Special Project of National Ling, H.F., Gu, J.H., 2017. Study on multi-depot open vehicle routing problem with
soft time windows. 53(14), 232-239. http://cea.ceaj.org/CN/Y2017/V53/I14/
Emergency Management System Construction of the National 232.
Social Science Foundation of China (20VYJ024). Liu, C., Kou, G., Zhou, X., Peng, Y.i., Sheng, H., Alsaadi, F.E., 2020. Time-dependent
vehicle routing problem with time windows of city logistics with a congestion
avoidance approach. Knowl.-Based Syst. 188, 104813. https://doi.org/10.1016/
References j.knosys.2019.06.021.
Liu, J.L., Ma, Z.J., 2013. Multi-depot open vehicle routing problem with time
windows based on vehicle leasing and sharing. Syst. Eng. -Theory Pract. 33 (3),
Afshar-Nadjafi, B., Afshar-Nadjafi, A., 2014. A constructive heuristic for time-
666–675 http://www.sysengi.com/CN/Y2013/V33/I3/666.
dependent multi-depot vehicle routing problem with time-windows and
Liu, C.S., Shen, L.Z., Sheng, H.Y., et al., 2019. Research on low-carbon time-dependent
heterogeneous fleet. J. King Saud Univ. – Eng. Sci. 29 (1), 29–34. https://doi.
vehicle routing problem with traffic congestion avoidance approaches. Control
org/10.1016/j.jksues.2014.04.007.
Decis., 1–10 https://doi.org/10.13195/j.kzyjc.2019.0257.
Alinaghian, M., Naderipour, M., 2016. A novel comprehensive macroscopic model
Mancini, S., 2014. Time dependent travel speed vehicle routing and scheduling on a
for time-dependent vehicle routing problem with multi-alternative graph to
real road network: the case of Torino. Transp. Res. Procedia 3, 433–441. https://
reduce fuel consumption: a case study. Comput. Ind. Eng. 99, 210–222. https://
doi.org/10.1016/j.trpro.2014.10.024.
doi.org/10.1016/j.cie.2016.07.029.
Mu, D., Wang, C., Wang, S., et al., 2015. Solving TDVRP based on parallel-simulated
Andelmin, J., Bartolini, E., 2017. An exact algorithm for the green vehicle routing
annealing algorithm. Comput. Integr. Manuf. Syst. 21 (6), 1626–1636. https://
problem. Transp. Sci. 51 (4), 1288–1303. https://doi.org/10.1287/
doi.org/10.13196/j.cims.2015.06.027.
trsc.2016.0734.
Oliveira, F.B., Enayatifar, R., Sadaei, H.J., Guimarães, F.G., Potvin, J.-Y., 2016. A
Beasley, J.E., 1981. Adapting the savings algorithm for varying inter-customer travel
cooperative coevolutionary algorithm for the Multi-Depot Vehicle Routing
times. Omega 9 (6), 658–659. https://doi.org/10.1016/0305-0483(81)90055-4.
Problem. Expert Syst. Appl. 43, 117–130. https://doi.org/10.1016/j.
Bezerra, S.N., Souza, S.R., Souza, M.J.F., 2018. A GVNS algorithm for solving the
eswa.2015.08.030.
multi-depot vehicle routing problem. Electron. Notes Discrete Math. 66 (4),
Pisinger, D., Ropke, S., 2007. A general heuristic for vehicle routing problems.
167–174. https://doi.org/10.1016/j.endm.2018.03.022.
Comput. Oper. Res. 34 (8), 2403–2435. https://doi.org/10.1016/
Bruglieri, M., Mancini, S., Pezzella, F., Pisacane, O., 2019. A path-based solution
j.cor.2005.09.012.
approach for the Green Vehicle Routing Problem. Comput. Oper. Res. 103, 109–
Poonthalir, G., Nadarajan, R., 2018. A Fuel Efficient Green Vehicle Routing Problem
122. https://doi.org/10.1016/j.cor.2018.10.019.
with varying speed constraint (F-GVRP). Expert Syst. Appl. 100, 131–144.
Çimen, M., Soysal, M., 2017. Time-dependent green vehicle routing problem with
https://doi.org/10.1016/j.eswa.2018.01.052.
stochastic vehicle speeds: an approximate dynamic programming algorithm.
Qi, M., Lin, W.-H., Li, N., Miao, L., 2012. A spatiotemporal partitioning approach for
Transp. Res. Part D: Transp. Environ. 54, 82–98. https://doi.org/10.1016/j.
large-scale vehicle routing problems with time windows. Transp. Res. Part E:
trd.2017.04.016.
Logis. Transp. Rev. 48 (1), 248–257. https://doi.org/10.1016/j.tre.2011.07.001.
Contardo, C., Martinelli, R., 2014. A new exact algorithm for the multi-depot vehicle
Qi, M.Y., Zhang, J.J., Ren, L., 2014. Vehicle routing algorithm based on spatiotemporal
routing problem under capacity and route length constraints. Discrete Optim.
clustering. Comput. Sci. 41 (3), 218–222 http://www.jsjkx.com/CN/Y2014/V41/
12 (5), 129–146. https://doi.org/10.1016/j.disopt.2014.03.001.
I3/218.
Cordeau, J.F., Gendreau, M., Laporte, G., 1997. A tabu search heuristic for periodic
Qian, J., Eglese, R., 2015. Fuel emissions optimization in vehicle routing problems
and multi-depot vehicle routing problems. Networks 30 (2), 105–119. https://
with time-varying speeds. Eur. J. Oper. Res. 248 (3), 840–848. https://doi.org/
doi.org/10.1002/(SICI)1097-0037(199709)30:2<105::AID-NET5>3.0.CO;2-G.
10.1016/j.ejor.2015.09.009.
Duan, F.H., Fu, Z., 2008. Multi depot open vehicle routing problem with soft time
Ray, S., Soeanu, A., Berger, J., Debbabi, M., 2014. The multi-depot split-delivery
windows and its tabu search. Comput. Eng. Appl. 44 (36), 42–44 http://cea.ceaj.
vehicle routing problem: Model and solution algorithm. Knowl.-Based Syst. 71,
org/CN/10.3778/j.issn.1002-8331.2008.36.011.
238–265. https://doi.org/10.1016/j.knosys.2014.08.006.
Duan, Z.Y., Lei, Z.X., Sun, S., Yang, D.Y., 2019. Multi-objective robust optimisation
Renaud, J., Laporte, G., Boctor, F.F., 1996. A tabu search heuristic for the multi-depot
method for stochastic time-dependent vehicle routing problem. J. Southwest
vehicle routing problem. Comput. Oper. Res. 23 (3), 229–235. https://doi.org/
Jiaotong Univ. 54 (3), 565–572 http://journal16.magtechjournal.com/Jweb_
10.1016/0305-0548(95)O0026-P.
xnjd/CN/Y2019/V54/I3/565.
Rezaei, N., Ebrahimnejad, S., Moosavi, A., Nikfarjam, A., 2019. A green vehicle
Fan, H.M., Yang, X., Li, D., Li, Y., Liu, P.C., Wu, J.X., 2019. Half-open multi-depot
routing problem with time windows considering the heterogeneous fleet of
vehicle routing problem based on joint distribution mode of fresh food. Comput.
vehicles: two metaheuristic algorithms. Eur. J. Ind. Eng. 13 (4), 507–535 https://
Integr. Manuf. Syst. 25 (1), 256–266 http://journal12.magtechjournal.com/
www.inderscience.com/info/inarticle.php?artid=100919.
Jweb_cims/CN/10.13196/j.cims.2019.01.026.
Sabar, N.R., Bhaskar, A., Chung, E., Turky, A., Song, A., 2019. A self-adaptive
Federal Highway Administration, 2015. Freight performance measure approaches
evolutionary algorithm for dynamic vehicle routing problems with traffic
for bottlenecks, arterials, and linking volumes to congestion report. U.S.
congestion. Swarm Evol. Comput. 44, 1018–1027. https://doi.org/10.1016/j.
Department of Transportation, Washington, DC (accessed 15 July 2020).
swevo.2018.10.015.
Figliozzi, M.A., 2012. The time dependent vehicle routing problem with time
Sadati, M.E.H., Aksen, D., Aras, N., 2020. A trilevel r-interdiction selective multi-
windows: Benchmark problems, an efficient solution algorithm, and solution
depot vehicle routing problem with depot protection. Comput. Oper. Res. 123,
characteristics. Transp. Res. Part E: Logist. Transp. Rev. 48 (3), 616–636. https://
104996. https://doi.org/10.1016/j.cor.2020.104996.
doi.org/10.1016/j.tre.2011.11.006.
Salhi, S., Imran, A., Wassan, N.A., 2014. The multi-depot vehicle routing problem
Ganji, M., Kazemipoor, H., Hadji Molana, S.M., Sajadi, S.M., 2020. A green multi-
with heterogeneous vehicle fleet: Formulation and a variable neighborhood
objective integrated scheduling of production and distribution with
search implementation. Comput. Oper. Res. 52, 315–325. https://doi.org/
heterogeneous fleet vehicle routing and time windows. J. Cleaner Prod. 259,
10.1016/j.cor.2013.05.011.
120824. https://doi.org/10.1016/j.jclepro.2020.120824.
Soto, M., Sevaux, M., Rossi, A., Reinholz, A., 2017. Multiple neighborhood search,
Haghani, A., Jung, S., 2005. A dynamic vehicle routing problem with time-dependent
tabu search and ejection chains for the multi-depot open vehicle routing
travel times. Comput. Oper. Res. 32 (11), 2959–2986. https://doi.org/10.1016/
problem. Comput. Ind. Eng. 107, 211–222. https://doi.org/10.1016/
j.cor.2004.04.013.
j.cie.2017.03.022.
Hickman, J., Hassel, D., Joumard, R., Samaras, Z., Sorenson, S., 1999. MEET
methodology for calculating transport emissions and energy consumption. In:

13
H. Fan, Y. Zhang, P. Tian et al. Computers and Operations Research 129 (2021) 105211

Soysal, M., Çimen, M., 2017. A simulation based restricted dynamic programming Xiao, Y., Konak, A., 2016. The heterogeneous green vehicle routing and scheduling
approach for the green time dependent vehicle routing problem. Comput. Oper. problem with time-varying traffic congestion. Transp. Res. Part E: Logist.
Res. 88, 297–305. https://doi.org/10.1016/j.cor.2017.06.023. Transp. Rev. 88, 146–166. https://doi.org/10.1016/j.tre.2016.01.011.
Taniguchi, E., Shimamoto, H., 2004. Intelligent transportation system based Xu, Z., Elomri, A., Pokharel, S., Mutlu, F., 2019. A model for capacitated green vehicle
dynamic vehicle routing and scheduling with variable travel times. Transp. routing problem with the time-varying vehicle speed and soft time windows.
Res. Part C: Emerg. Technol. 12 (3-4), 235–250. https://doi.org/10.1016/j. Comput. Ind. Eng. 137, 106011. https://doi.org/10.1016/j.cie.2019.106011.
trc.2004.07.007. Yang, H., 2018. Research of urban recurrent congestion evolution based on taxi GPS
Vidal, T., Crainic, T.G., Gendreau, M., Lahrichi, N., Rei, W., 2012. A hybrid genetic data. Harbin Institute of Technology http://cdmd.cnki.com.cn/Article/CDMD-
algorithm for multidepot and periodic vehicle routing problems. Oper. Res. 60 10213-1018895490.htm.
(3), 611–624. https://doi.org/10.1287/opre.1120.1048. Yu, Y., Wang, S., Wang, J., Huang, M., 2019. A branch-and-price algorithm for the
Wang, Y., Assogba, K., Fan, J., Xu, M., Liu, Y., Wang, H., 2019. Multi-depot green heterogeneous fleet green vehicle routing problem with time windows. Transp.
vehicle routing problem with shared transportation resource: Integration of Res. Part B: Methodol. 122, 511–527. https://doi.org/10.1016/j.trb.2019.03.009.
time-dependent speed and piecewise penalty cost. J. Cleaner Prod. 232, 12–29. Zhang, W.H., Gajpal, Y., Appadoo, S.S., Wei, Q., 2020. Multi-depot green vehicle
https://doi.org/10.1016/j.jclepro.2019.05.344. routing problem to minimize carbon emissions. Sustainability 12 (8), 3500.
Wang, X.P., Li, X.Y., Zhang, J., 2018. The optimization research of vehicle routing https://doi.org/10.3390/su12083500.
problem with heterogeneous fleet, simultaneous pickup-delivery considering Zhang, C.Q., Zhang, Y., 2017. Study on Vehicle Route Problem Under Dynamic Road
temporal-spatial distance. Chin. J. Manag. 15 (6), 918–926 System. J. Transp. Eng. Inf. 15 (2), 112–118 https://mall.cnki.net/onlineview/
http://manu68.magtech.com.cn/Jwk_glxb/CN/. MagaView.aspx?fn=jtgc201702*1*.
Wu, Y., Ma, Z.J., 2017. Time-dependent production-delivery problem with time Zhong, S.Q., Du, G., He, G.G., 2006. Study on open vehicle routing problem with time
windows. Syst. Eng. -Theory Pract. 37 (1), 172–181 http://www.sysengi.com/ windows limits and its genetic algorithm. Comput. Eng. Appl. 34, 201–204
CN/Y2017/V37/I1/172. https://mall.cnki.net/onlineview/MagaView.aspx?fn=JSGG200634*1*.

14