You are on page 1of 17

Journal of Cleaner Production 225 (2019) 647e663

Contents lists available at ScienceDirect

Journal of Cleaner Production


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

Development of energy consumption optimization model for the


electric vehicle routing problem with time windows
Yiyong Xiao a, Xiaorong Zuo b, Ikou Kaku c, Shenghan Zhou a, *, Xing Pan a
a
School of Reliability and Systems Engineering, Beihang University, Beijing, 100191, China
b
China Aerospace Academy of System Science and Engineering, Beijing, 100854, China
c
Department of Environmental Management, Tokyo City University, Yokohama, 224-8551, Japan

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

Article history: Electric vehicles (EVs) are promising transportation tools for supporting green supply chain and cleaner
Received 1 August 2018 production. In contrast to traditional fossil fuel-powered vehicles, which usually have a short range at
Received in revised form lower speeds, EVs have a much longer (even double) range when traveling at lower speeds than high
12 March 2019
speeds. This feature has a major impact to the vehicle routing problem when EVs are used in the fleet.
Accepted 29 March 2019
Available online 3 April 2019
This study investigated the electric vehicle routing problem with time window (EVRPTW) considering
the energy/electricity consumption rate (ECR) per unit of distance traveled by an EV as a function of the
speed and load, referred to as EVRPTW-ECR for simplicity. As a consequence, the maximum range of an
Keywords:
Electric vehicle
EV is estimated dynamically according to its speeds and loads along the route. A mixed-integer linear
Green vehicle routing problem programming (MILP) model was developed for EVRPTW-ECR, where the EV’s speed was treated as a
Mixed-integer programming continuous decision variable and the battery capacity, instead of a constant distance, was taken as the
Continuous optimization range restriction. Two linearization methods, i.e., the inner approximation and outer approximation,
were introduced to handle the nonlinear relationship between the traveling speed and travel time with a
given parameter ε to control the maximum permissible error. Computational experiments were carried
out based on Solomon’s instances to test the efficiency and effectiveness of the proposed model and
methods, thereby demonstrating that the MILP model can be solved optimally for up to 25 customers by
the CPLEX solver and partially optimized for large instances of up to 100 customers by using a heuristic
approach.
© 2019 Elsevier Ltd. All rights reserved.

1. Introduction such as the energy-minimizing VRP by Kara et al. (2007), the


emissions-based time-dependent VRP by Figliozzi (2010) and Jabali
The high use of fossil fuel-powered vehicles throughout the et al. (2012), the pollution routing problem by Bektas and Laporte
world in recent decades has led to the emission of large volumes of (2011), the fuel consumption rate consideration VRP by Xiao et al.
CO2, which is a well-known greenhouse gas (GHS), and they ac- (2012), the low carbon routing problem by Zhang et al. (2015),
count for approximately three-quarters of the total emissions due and the green vehicle routing and scheduling by Xiao and Konak
to transportation activities (IEA, 2016). The transportation section (2015, 2016, 2017).
was the second largest contributor to greenhouse gases, where it In recent years, electric vehicles (EV) have been introduced as
was responsible for 23% of the global CO2 emissions in 2014 (IEA, alternative-energy vehicles for commercial and personal uses, and
2016). A number of fuel/emission-oriented VRP models have been their market share has increased at a rapid rate, e.g., by more than
developed that consider the balanced optimization of monetary 50% annually in China where the numbers of EVs sold in 2015, 2016,
costs and environmental concerns for fossil fuel-powered vehicles, and 2017 were 331,000, 507,000, and 777,000, respectively (IEA,
2018). The boom in the popularity of EVs is due mainly to their
environmentally friendly characteristics, such as no local green-
house gas emissions, low noise pollution, and high energy effi-
* Corresponding author. ciency, which can help logistics companies to attain a green image
E-mail addresses: xiaoyiyong@buaa.edu.cn (Y. Xiao), zuoxiaorong710@163.com
from an increasing number of socially and environmentally aware
(X. Zuo), kakuikou@tcu.ac.jp (I. Kaku), zhoush@buaa.edu.cn (S. Zhou), panxing@
buaa.edu.cn (X. Pan). customers (Desaulniers et al., 2016). Government’s regulatory

https://doi.org/10.1016/j.jclepro.2019.03.323
0959-6526/© 2019 Elsevier Ltd. All rights reserved.
648 Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663

Hooks (2012) proposed another version of EVRP, called the green


List of acronyms vehicle routing problem (GVRP), which involves using “alternative
energy vehicles” in the fleet that need to visit alternative fueling
Acronym Meaning stations throughout the trip. In GVRP, alternative fueling stations
GHS Green house Gas are copied multiple times and treated as special customers (dummy
ECR Energy/electricity consumption rate nodes) that can be visited at a maximum of one time. Thus, the
EV Electric vehicle GVRP was converted to be in accordance with the traditional VRP
VRP Vehicle routing problem frameworks and it can be solved by using heuristics developed for
EVRP Electric vehicle routing problem traditional VRPs. Felipe et al. (2014) modeled the EVRP by consid-
EVRPTW Electric vehicle routing problem with time ering different technologies for battery recharging where each
window technology is associated with a constant recharging rate and a fixed
EVRPTW-ECR Electric vehicle routing problem with time unit cost. Schneider et al. (2014) proposed an EVRP model with
window considering energy consumption rate time window constraints (EVRPTW), which aims to minimize both
EVTP Electric vehicle touring problem the number of vehicles employed and the total traveled distance.
GVRP Green vehicle routing problem They developed a hybrid metaheuristics as the solution approach
IEA International Energy Agency by combining a variable neighborhood search with a tabu search
INS Iterative neighborhood search method. Bruglieri et al. (2015) considered a partial recharge policy
IPO Iterative partial optimization in the EVRPTW, where the battery is not always fully recharged
MILP Mixed integer linear programming along the travel route. Hiermann et al. (2016) modeled the EVRP
MIP Mixed integer programming with mixed types of EVs with different capacities, costs, and
NEDC New European Driven Cycling recharging rates. Keskin and Çatay (2016) developed an adaptive
SVR Speed varying range large neighborhood search algorithm for the EVRPTW and used a
partial recharge strategy. Similar strategy can also be found in
Bruglieri et al. (2015) and Hiermann et al. (2016). Montoya et al.
(2017) conducted the first and only study to formulate an EVRP
support is also an important factor that promoted the use of EVs in model by considering a nonlinear charging function. Lu et al. (2018)
both public transportations and individual travels (Li et al., 2016). proposed a multi-objective optimal load dispatch model for
According to the life cycle assessment method by Wu et al. (2018), microgrid that was accessed stochastically with a large-scale of
the total life cycle GHG reduction potential of battery EVs will uncoordinated charging of EVs.
gradually reach 13.4% in 2020. Recent sustainability analysis and In some EVRP models, the EVs are assumed to have swappable
practical estimation methods of life cycle carbon emission by EVs batteries and they are allowed to visit battery-exchange stations to
considering power plant can be found in Fern andez (2018), and swap their batteries in a rapid manner, thereby avoiding a long
Casals et al. (2016), and Xiao et al (2011). In addition to their charging time. Zheng et al. (2014) studied the optimal design of
environmental benefits, EVs are attractive from a cost perspective battery charging/swap stations in a distribution system. Adler and
compared with conventional fossil fuel-powered vehicles. The cost Mirchandani (2014) presented a case study of an online routing
of the electricity consumed by an EV is only about 10%e15% of the system for EVs, where multiple originedestination pairs need to be
cost of the gasoline/petrol consumed by a conventional vehicle optimally routed while considering battery swapping and reser-
when traversing the same distance.1 vation. Yang and Sun (2015) studied the battery swap station
The electric vehicle routing problem (EVRP), which involves the location-routing problem with capacitated EVs, where the routing
use of commercial EVs in the logistics distribution/service field, is plan and selection of battery-exchange stations were optimized
one of the latest extensions of the vehicle routing problem (VRP) simultaneously. Liao et al. (2016) studied the electric vehicle tour-
and it has attracted much attention from researchers in recent ing problem (EVTP) with battery swapping operations, which in-
years (Lin et al., 2014, 2016; Demir et al., 2014). EVs have a limited volves finding a tour that visits all given vertices and returns to the
travel range and recharging is inconvenient during trips, so the origin while stopping at a set of required battery switch stations.
primary challenge related to the EVRP is how to incorporate the They developed polynomial time algorithms for the EVTP with
visits to recharge stations into the route optimization. In some fixed tours (i.e., all vertices visited in fixed orders). Schiffer and
recent EVRP models, the visits to charging stations are optimized Walther (2017) extended the EVRP model proposed by Yang and
during trips, such as in the green vehicle routing problem (GVRP) Sun (2015) to include various practical factors, such as charging
(Erdog an and Miller-Hooks, 2012; Felipe et al., 2014; Andelmin and station visits, time window constraints, and partial recharging
Bartolini, 2017), battery-swapping EVRP (Yang and Sun, 2015; Hof options.
et al., 2017; Jie et al., 2018), partial recharging EVRP (Felipe et al., In all of the previous EVRP models mentioned above, a general
2014; Keskin and Çatay, 2016; Bruglieri et al., 2017), EVRP consid- assumption was made that the energy consumption of an EV over
ering a mix of battery-swapping and partial recharge (Schiffer and an arc is a linear function of the distance traveled. Goeke and
Walther, 2017), and the EVRP with a nonlinear charging function Schneider (2015) conducted the first study to use a more realistic
(Montoya et al., 2017). energy consumption model for modeling the EVRPTW with a
Conrad and Figliozzi (2011) were the first to model the mixed fleet (EVs and fuel vehicles). They developed a nonlinear ECR
recharging vehicle routing problem, assuming that EVs with for EVs by considering various factors, including the vehicle mass,
limited travel ranges are allowed to recharge at certain customer speed, and gradient of the terrain. However, in their EVRPTW
locations in order to complete their deliveries. Erdog an and Miller- model, the travel times/speeds over all arcs were assumed to be
previously known constants. Thus, the electric energy consumption
over an arc was affected only by the weight of the cargo. Other
1
studies also developed energy consumption models for EVs. Yuan
The amount of electricity (or gasoline) consumed by an EV (or a conventional
car) for traveling 100 km at 60 km/h is approximately 10e20 kWh (or 10e20 L),
et al. (2017) developed a physical-based statistical method for
costing 5e10 (or 60e120) RMB according to the average electricity (or gasoline) evaluating the energy consumption of an EV during a driving cycle,
price in China. which was modeled based on the energy flow path, as well as
Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663 649

including the battery losses, electric motor losses, driveline losses, and computational experiments were carried out based on
brake losses and load losses. Fiori et al. (2016) developed an Solomon’s instances (Solomon, 1987) to verify the model and
instantaneous energy consumption model called the comprehen- analyze its characteristics.
sive power-based EV energy consumption model, which can be (4) A heuristic approach was proposed for solving large-sized
used to calculate the instantaneous ECRs of EVs with input variables problems with near-optimal solutions.
including the instantaneous speed, acceleration, and vehicle mass.
Liu et al. (2017) investigated the impact of the road gradient on the The remainder of this paper is organized as follows. In Section 2,
electricity consumption of EV. Fetene et al. (2017) harnessed big the ECR function is modeled as a function of the speed and load,
data to estimate the energy consumption of EVs and showed that and it is linearized with a piecewise linearization method. In Sec-
the ECR might vary greatly in a nonlinear manner with the driving tion 3, the inner and outer approximations linearization methods
patterns and weather conditions. are presented for linearizing the nonlinear relationship between
With the rapid development of battery technology, the range of the travel time and speed. The MILP model for the EVRPTW-ECR is
EVs has grown steadily in the last few years. In 2018, many of the described in Section 4, as well as some theoretical propositions. In
latest EV models sold in China have a maximum range over 400 km Section 5, computational experiments are carried out and the re-
(in NEDC), which are relatively high and close to those of conven- sults are analyzed. Finally, the paper is concluded in Section 6.
tional fossil fuel-powered vehicles (D1EV, 2008). To some extent,
this improvement has alleviated the so-called “range anxiety” 2. Modeling the ECR function for EVs
among drivers. In the near future, it is likely that a logistic EV may
not need to visit a charging (or battery-exchange) station during According to the electricity usage model described in Haaren
the entire daytime because its fully-charged battery (in the depots (2012) and Marmaras et al. (2017), the energy losses of an EV are
at night) is able to last for one whole day of logistics service. distributed into four categories: Aerodynamic, Tires, Drivetrain, and
However, another important feature of EVs, referred to as the Ancillary. Among them the Aerodynamic losses are the major factor
speed-varying range (SVR), is also crucial for the EVRP even though in the high velocity regime (>50 mph), as the air resistance,
the battery capacity has improved. That is, the maximum travel calculated by 12Cd rAv2, is proportional to v2. Therefore, EVs have
range of an EV varies greatly with respect to different driving better energy efficiencies when traveling at lower speeds, which is
speeds, and most EVs have a considerably higher (even double) in contrast to traditional fossil fuel-powered vehicles. Fig. 1 shows
range when driving at a low speed than a high speed. For example, the relationship between the driving range and speed for two EV
the Model S 85 kWh, an EV brand produced by the Tesla Company models (Roadster and Model S 85 kWh) produced by the Tesla
in the USA, has a range of 455 miles when driving at a low speed of Company, which clearly demonstrates that the driving range varies
20 miles/h, but its range may decrease to 230 miles when driving at greatly as the speed increases from low to high. The most efficient
a high speed of 70 miles/h (Musk and Straube, 2012). The BYD e6, speed is around 15e25 mph, which is approximately twice as
another EV brand in China, has a range of 560 km at a speed of efficient as high speeds.
20 km/h but only 252 km when the speed increases to 100 km/h Let the function L(v) represent the driving range of EV at speed v.
(Liu, 2014). It should be noted that the SVR feature of EVs contrasts Then, the ECR can be defined as 1=LðvÞ, which represents the per-
with that of traditional fossil fuel-powered vehicles, which usually centage of the total battery capacity consumed when traveling one
have a relatively short range at low speed but a longer range at high unit of distance at a constant speed, v. The ECR function is nonlinear
speed. In existing literature, the SVR feature of EVs has not been and a downward-canvas curve, which can be linearized using a set of
considered in previously reported EVRP models, although it is polylines (section lines). As shown in Table 1 and Fig. 2, several
important for optimizing the traveling speeds in the EVRP in order discretized points were extracted from the ECR curves of Fig. 1 and
to balance the tradeoff between the travel range and travel time. the linearity between the range and speed was analyzed. It can be
The PRP model of Bektas and Laporte (2011) introduced a fuel observed that the regression line for the ECR-speed relation of
consumption function into VRPTW based on the comprehensive Roadster is: y ¼ 2.695  105x þ 4.075  104, with an R2 value of
modal emission model (CMEM) by Barth et al. (2005) and Barth and 0.9668. For the Model S 85 kWh, the regression line is:
Boriboonsomsin (2008). However, the CMEM model was developed y ¼ 1.783  105x þ 5.786  104, with an R2 value of 0.9386. Table 2
for fossil fuel powered vehicles where the energy efficiency is al- shows the rangeespeed relationship for another EV model, BYD E6,
ways at high speed, whereas the EVs, opposite to such a charac- and its regression line is plotted in Fig. 3. The regression line for this
teristics, have always higher energy efficiency in lower speed. model is: y ¼ 2.634  105x þ 1.1711  103 and R2 ¼ 0.97616.
This study investigated the EVRPTW considering dynamic ECRs Since the R2 values of all these regression lines are very close to
affected by the speeds and loads along the routes (EVRPTW-ECR for 1, so it is reasonable to consider approximately the ECR as a linear
short). Contributions of this study were on four folds listed as follows. function of speed. Thus, a line e ¼ Kv þ B can be used as a surrogate
for the ECR function of these EV models, where e represents ECR, v
(1) The energy/electricity consumption rate (ECR) per unit of is the traveling speed, and K and B are the pre-calculated slope and
traveled distance was introduced into the EVRPTW model intercept, respectively. Let ri (and rj) be the remaining battery po-
and ECR was considered as a joint nonlinear function of the wer (as a percentage of the total battery capacity) for the vehicle
speed and load. Thus, the traveling speed acts as a contin- when arriving at nodes i (and j), and vij and eij be respectively the
uous decision variable to affect the maximum ranges of EVs speed and ECR of the vehicle traveling on selected arc (i, j). Then,
during route optimization. This is the first study to consider the following Eq. (1) can be used to model the electricity con-
the effects of the dynamic range of EVs in the EVRP. sumption flow along the route:
(2) Two linearization methods were introduced, comprising the
8
inner approximation based on secant lines and the outer < ri  rj ¼ Dij eij
approximation based on tangent lines, in order to approxi- e  Kvij þ B cði; jÞ; (1)
: ij
mate the nonlinear relationship between the travel time and s  ri  1
traveling speed.
(3) Based on the above folds (1) and (2), a mixed-integer linear where s is the safety threshold for the remaining power rate when
programming (MILP) model was developed for EVRPTW-ECR returning to the depot. It should be noted that Eq. (1) is only
650 Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663

(A) Range vs. speed (B) Nonlinear ECR curve

Fig. 1. Examples of the relationships between the of range and speed for EVs
(Source: Musk and Straube (2012) at www.tesla.com/blog/model-s-efficiency-and-range).

Table 1
Relationship between the range and speed for EVs.

Constant speed (v) Roadster Model S 85 kWh

miles/h km/h Driving range ECR (1/km) % Driving range ECR (1/km) %

20 32.2 656.5 km 0.15% 731.8 km 0.14%


30 48.3 590.3 km 0.17% 697.5 km 0.14%
40 64.4 499.1 km 0.20% 624.6 km 0.16%
50 80.5 417.0 km 0.24% 540.1 km 0.19%
60 96.6 348.6 km 0.29% 458.0 km 0.22%
70 112.7 289.4 km 0.35% 387.4 km 0.26%
80 128.8 243.7 km 0.41% 323.6 km 0.31%

0.5%
0.4%
0.5%
0.4%
Roadster Model S 85kWh
0.4% 0.3%
ECR (1/km)

0.3%
ECR (1/km)

0.3% 0.2%
0.2%
0.2%
0.1%
0.1% 0.1%
0.0%
0.0% 0 20 40 60 80 100 120
20 40 60 80 100 120 140 Speed (km/h)
Speed (km/h)
Fig. 3. Linearity of the ECR functions of speed for the BYD E6 model EV.
Fig. 2. Linearity of the ECR functions of speed for the Roadster model and Model S
85 kWh.
suitable for an ECR function that can be linearized with an R2 value
very close to 1. In order to approach a general, nonlinear, and
downward ECR function more accurately, a piecewise linearization
Table 2 with multiple secant lines can be applied as a surrogate of the ECR
Relationship between the range and speed for the BYD E6 model EV. curve such that the inequality eij  Kvij þ B in Eq. (1) may be
Constant speed (v) Maximum range (L) ECR (1/km) % replaced with eij  Kq vij þ Bq , where q2Q and Q represent the set of
consecutive secant lines. Note that the piecewise linear function
20 km/h 560 km 0.18%
40 km/h 454 km 0.22% allows the decision variable, e.g., the speed vij, to take value in a
60 km/h 371 km 0.27% continuous domain with controllable error ranges, whereas a dis-
80 km/h 322 km 0.31% cretization approach, e.g., in the PRP model, just allows the decision
100 km/h 252 km 0.40% variable to pick a discrete value from a set of pre-defined
(Source: Liu (2014) at auto.enorth.com.cn/system/2014/08/06/012065564.shtml). candidates.
Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663 651

The EV’s load increases the rolling resistance of its tires against
the road surface and the rolling bearings. The energy consumed due
to the rolling resistance is normally assumed to be proportional to
the total weight of the vehicle and the length of the distance tij=D ij/vij
traveled, but independent of the traveling speed (Bektas and
Laporte, 2011). For an EV used in logistic services, the load may
vary greatly along the delivering/pickup tour, and thus it has a p1
marginal effect on the ECR. It is assumed that the marginal increase
in the ECR function for an EV due to the load effect can be estimated

Time tij
by a constant coefficient 4, which is a parameter related to the EV
itself that indicates the unit increment in the ECR after adding one
unit of additional load. Thus, let fij be the load on the EV when
traveling the arc (i, j), Eq. (1) can then be modified to Eq. (2) to p2
consider the load effect on the ECR for the EV.
8 p3
>
> r  rj ¼ Dij eij
>
< i ...
eij  Kq vij þ Bq þ 4fij cði; jÞ; q2Q (2) pn
>
>
>
:s  r  1
i

vmin Speed vij vmax


3. Bounding the travel time to speed using piecewise
linearization Fig. 4. Outer approximation of the nonlinear timeespeed curve.

In order to take the speed as a continuous decision variable in a


formula in Eq. (5).
linear EVRPTW model, the nonlinear relationship between the
travel time and traveling speed, i.e., the timeespeed curve, must be ( .
kp ¼ m22p v2min
linearized. In the following, two piecewise linearization methods . cp ¼ 1; 2; :::; h (5)
are introduced. The first is called outer approximation piecewise bp ¼ 2m1p vmin
linearization, which uses a set of tangent lines to surrogate the
timeespeed curve. The second is called inner approximation piece- It should be noted that kp and bp in Eq. (5) do not depend on the
wise linearization that uses a set of secant lines to surrogate the distance Dij, so they can be pre-calculated before starting to solve
timeespeed curve. Both methods guarantee that the maximal the problem. All of the arcs with different lengths can use the same
approximation error incurred by linearization is controlled within a set of kp and bp in the linearization of the timeespeed relation if they
given range. have the same level (ε) of accuracy requirements. In Table 3, the
values of kp and bp are provided for ε ¼ 1.0% and 0.1%, respectively,
which can be applied directly to existing EVRP models.
3.1. Outer approximation linearization
3.2. Inner approximation linearization
When traveling a given arc (i, j) with distance Dij, the nonlinear
timeespeed curve, i.e., tij ¼ Dij =vij , can be approximated by using a In addition to the outer approximation, the inner approximation
P
set of tangent lines on the basic curve Xij ¼ 1 cj2N0 , denoted uses a set of consecutive secant lines, denoted as P ¼ {p1, p2, … }, to
ði;jÞ2A surrogate the nonlinear curve tij ¼ Dij/vij, starting from the mini-
as P ¼ fp1 ; p2 ; :::g, starting from the minimum speed (vmin) to the mum speed (vmin) to cover the maximum speed (vmax), as shown in
maximum speed (vmax), as shown in Fig. 4. Fig. 5.
Each tangent line p2P is expressed as tij ¼ kp vij þ bp . If the Similarly, to guarantee that the maximum approximation error
objective involves minimizing a function that depends on the travel rate is within a given percentage, tij )60Dij =vij , the minimum
time, then the travel time tij can be bounded to the traveling speed number of secant lines can be determined by Eq. (6) as follows.
vij by using the following set of linear constraints.
 
lnðvmax =vmin Þ
tij  Dij kp vij þ Dij bp ; cp21; 2; :::; h (3) h¼ (6)
ln m
It should be noted that Eq. (3) can approximate the nonlinear pffiffiffiffiffiffiffiffiffiffiffiffiffi
where m ¼ 1 þ 2ε þ 2 ε þ ε2 . It should be noted that parameter m
curve aj  ai  tij þ gi þ wi þ Mð1  Xij Þcði; jÞ2A0 more accurately
has a different expression compared with that used in the outer
when more number of tangent lines is used. Let t 0 ij be the
approximation in Eq. (4).
approximated time, in order to guarantee that the error rate of t 0 ij is
By starting with the first secant line from the minimum speed
within a given number ε%, i.e., ri  rj  Dij eij  Mð1  Xij Þcði; jÞ2A0 ,
vmin, the following formula can be deduced to determine the slopes
the minimum number of tangent lines h can be determined by Eq.
ði;jÞ ði;jÞ
(4) as follows. and intercepts of the secant lines, t ¼ K p v þ Bp ; p2P, for a given
  arc (i, j) with a distance Dij.
lgðvmax =vmin Þ
h¼ þ1 (4) 8
lgm >
> K
ði;jÞ
¼
1
,Dij
>
< p
pffiffiffi pffiffiffi m 2p1
,v2min
where m ¼ ð1 þ εÞ=ð1  εÞ and QaS indicates the minimum ; cp ¼ 1; 2; :::; h (7)
>
> mþ1
integer that is greater than or equal to a. The slopes and intercepts >
: Bpði;jÞ ¼ p ,D
of the tangent lines can be calculated by the following recursive m ,vmin ij
652 Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663

Table 3
kp and bp for the outer approximation of the traveling speed in [20, 100] km/h.

ε p kp bp ε p kp bp

1.0% 1 2.500000E-03 1.000000E-01 0.1% 9 9.084862e-04 6.028221e-02


2 1.673554E-03 8.181818E-02 10 8.005079e-04 5.658650e-02
3 1.120313E-03 6.694215E-02 11 7.053633e-04 5.311735e-02
4 7.499615E-04 5.477085E-02 12 6.215273e-04 4.986090e-02
5 5.020403E-04 4.481251E-02 13 5.476555e-04 4.680408e-02
6 3.360766E-04 3.666478E-02 14 4.825638e-04 4.393467e-02
7 2.249769E-04 2.999846E-02 15 4.252086e-04 4.124117e-02
8 1.506044E-04 2.454419E-02 16 3.746704e-04 3.871281e-02
9 1.008178E-04 2.008161E-02 17 3.301388e-04 3.633945e-02
18 2.909001e-04 3.411159e-02
0.1% 1 2.500000e-03 1.000000e-01 19 2.563251e-04 3.202031e-02
2 2.202862e-03 9.386931e-02 20 2.258596e-04 3.005725e-02
3 1.941040e-03 8.811448e-02 21 1.990150e-04 2.821453e-02
4 1.710338e-03 8.271246e-02 22 1.753610e-04 2.648479e-02
5 1.507055e-03 7.764162e-02 23 1.545185e-04 2.486109e-02
6 1.327934e-03 7.288165e-02 24 1.361531e-04 2.333693e-02
7 1.170102e-03 6.841351e-02 25 1.199706e-04 2.190622e-02
8 1.031029e-03 6.421929e-02 26 1.057115e-04 2.056322e-02

4.1. Problem description

The proposed EVRPTW-ECR model is described on a directed


tij=D ij/vij graph G ¼ (N, A), where N ¼ {0, 1, 2, …, n} is the set of nodes rep-
resenting the customers and depot (denoted by 0);
A ¼ fði; jÞji; j2N; isjg is the set of customer/depot arcs that an EV
can select to traverse. Set A may not include all combination of node
pairs in N, so graph G may be either complete or incomplete. Each
arc, (i, j), in A is associatedwith a known distance, Dij, and a pair of
Time tij

lower/upper speed limits, vij ;v ij . A fleet of homogenous EVs are to


p1
be dispatched to serve the customers in N by starting from and
returning to a depot. Each customer, i, in N must be served once and
is associated with a demand, ai, a service duration time, gi, and a
time window [Si, Ei] in which the service must start. Each EV has a
p2 maximum battery capacity, CB, and a maximum load capacity, CL. All
EVs leave the depot with fully-charged batteries and return to the
p3 ... depot with residual battery capacities above a safe level, s. Vehicles
are not allowed to recharge outside. All EVs have an identical ECR
function that depends on the continuous decision variables
comprising speed and load. The ECR function with zero load is
vmin Speed vij vmax linearized with a set, Q, of secant/tangent lines, with a pair of
slopes, Kq, and an intercept, Bq, for each line, q, in Q. For each
Fig. 5. Inner approximation of the nonlinear timeespeed curve. additional one unit of load, the ECR of the EV increases by a con-
stant rate, 4. The objective is to minimize the total travel cost,
and bp ¼ mpm,v
Let kp ¼ m2p11,v2 þ1
min
, and thus Eq. (3) can also be including the fixed cost of using EVs and hiring drivers, the variable
min
cost of electricity/energy consumed along the tours, and the vari-
used to bound the travel time tij to the traveling speed vij in the case
able wages of drivers related to the length of the travel time/
of minimizing an objective that is positively related to the traveling
distance.
time. Note that kp and bp are also independent of the distance Dij,
The main challenge when solving the EVRPTW-ECR is that an EV
and thus they can be pre-calculated before starting to solve the
has a highly sensitive ECR, which is affected mainly by the speed
problem. Table 4 shows the values of kp and bp for ε ¼ 1.0% and 0.1%
and load, where a high speed and heavy load may considerably
for using the inner approximation.
lower the maximum range. An EV can choose to either serve more
Note that the outer/inner approximation may result an
customers at lower speed (with a longer total travel time and
approximated travel time (for an arc with given speed) that is
higher variable costs) or to serve fewer customers at higher speed
smaller/greater than (or equal to) the exact travel time, but it
(this will require that more EVs fulfill the deliveries with a higher
guarantees that the maximum deviation is within a given param-
fixed cost). The time-window requirements from customers make
eter ε. In Section 4.3, a post-optimization procedure is presented for
the problem even more complex to optimize. Therefore, the
eliminating the approximation errors caused by the outer/inner
EVRPTW-ECR problem considered in this study involves not only
approximations.
optimizing the vehicle routes in a similar manner to the traditional
VRPs, but also optimizing the traveling speeds and loads in order to
4. Problem description and formulation
ensure an overall optimization of the total cost. In Section 4.2, a
MILP model of EVRPTW-ECR is developed, which is based on the
In this section, the EVRPTW-ECR problem is formally described
linearization approaches introduced in Sections 2 and 3. Some
and an MILP model is developed, as well as several theoretical
theoretical analysis of the MILP model are provided in Section 4.3.
propositions for the solution obtained by the proposed MILP model.
Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663 653

Table 4
Kp and Bp for the inner approximation of the traveling speed in [20, 100] km/h.

ε p Kp Bp ε p Kp Bp

1.0% 1 2.047506e-03 9.095012e-02 0.1% 9 8.532484e-04 5.845004e-02


2 1.373396e-03 7.448838e-02 10 7.518830e-04 5.486838e-02
3 9.212258e-04 6.100617e-02 11 6.625597e-04 5.150620e-02
4 6.179261e-04 4.996420e-02 12 5.838480e-04 4.835005e-02
5 4.144833e-04 4.092081e-02 13 5.144872e-04 4.538730e-02
6 2.780209e-04 3.351424e-02 14 4.533664e-04 4.260609e-02
7 1.864867e-04 2.744825e-02 15 3.995067e-04 3.999531e-02
8 1.250888e-04 2.248018e-02 16 3.520455e-04 3.754451e-02
9 8.390518e-05 1.841133e-02 17 3.102227e-04 3.524389e-02
18 2.733684e-04 3.308424e-02
0.1% 1 2.346807e-03 9.693614e-02 19 2.408924e-04 3.105693e-02
2 2.068008e-03 9.099617e-02 20 2.122745e-04 2.915385e-02
3 1.822330e-03 8.542018e-02 21 1.870564e-04 2.736739e-02
4 1.605838e-03 8.018587e-02 22 1.648342e-04 2.569039e-02
5 1.415065e-03 7.527231e-02 23 1.452520e-04 2.411616e-02
6 1.246956e-03 7.065984e-02 24 1.279962e-04 2.263839e-02
7 1.098818e-03 6.633000e-02 25 1.127903e-04 2.125117e-02
8 9.682795e-04 6.226549e-02 26 9.939086e-05 1.994896e-02

4.2. A mixed-integer linear programming model ri Remaining energy rate (%) of EV when arriving at customer i
ai Arrival time (min) at customer i
The parameters and decision variables used to describe the wi Vehicle’s waiting time (min) at customer i
EVRPTW-ECR are listed as follows. fij Load of vehicle on arc (i, j)
Parameter notations:
The objective function is to minimize a comprehensive cost
N Set of nodes, including the depot function that includes the fixed cost for EVs and drivers, the vari-
N’ Set of customer nodes, N0 ¼ N\f0g able cost by electricity/energy consumption, and the variable cost
A Set of arcs that an EV can select to traverse, A ¼ fði;jÞji2N;j2 related to traveling time (or traveled distance). Thus, the MILP
N; isjg model for EVRPTW-ECR is formulated as follows.
A’ Set of arcs linking customer nodes, A0 ¼ fði;jÞji2N0 ;j2N0 ;isjg, 2
and A0 3A X X X
εMaximum allowed error in the timeespeed linearization. Min: Total Cost ¼ F X0j þ CB fe Dij eij þ fd 4 tij
i, j Index of nodes, i ¼ 0, 1, 2, …, n (the depot is represented by 0) j2N 0 ði;jÞ2A ði;jÞ2A
3
n Total number of nodes (including the depot)
X
Dij Distance (km) of arc (i, j) þ ðgi þ wi Þ5 (8)
vij Minimum traveling speed (km/h) on arc (i, j) i2N0
vij Maximum traveling speed (km/h) on arc (i, j)
di Demand of customer i Subject to:
gi Service time (in minute) for customer i P
[Si, Ei] Time window for starting to serve customer i (1) Xij ¼ 1 cj2N0
ði;jÞ2A
CL Load capacity of EV P
(2) Xij ¼ 1 ci2N0
CB Maximum battery capacity of EV (kWh) ði;jÞ2A
F Fixed cost per EV and driver (RMB) (3) tij  60ðDij kp vij þ Dij bp Þ  Mð1  Xij Þcði; jÞ2A; p2P
fe Cost consumed per unit of battery energy (RMB/kWh) (4.1) aj  t0j  Mð1  X0j Þcj2N 0
fd Driver’s variable wage rate (RMB/min) (4.2) aj  ai  tij þ gi þ wi  Mð1  Xij Þcði; jÞ2A0
P Set of secant lines for the timeespeed linearization, pre- (4.3) aj  ai  tij þ gi þ wi þ Mð1  Xij Þcði; jÞ2A0
determined by parameter ε (5) eij  Kq vij þ Bq þ 4fij  ð1  Xij Þcði; jÞ2A; q2Q
kp, bp Slope and intercept of the pth secant line in P, p2P. (6.1) 1  rj  D0j e0j  ð1  X0j Þcj2N0
Q Set of secant lines for the ECR function of EV (6.2) ri  rj  Dij eij  ð1  Xij Þcði; jÞ2A0
Kq, Bq Slope and intercept of the qth secant line of the ECR (6.3) rj  ri  Dij eij þ ð1  Xij Þcði; jÞ2A0
function in Q, q2Q . (6.4) ri  s þ Di0 ei0  ð1  Xi0 Þci2N0
4Additional ECR for one unit of load (6.5) ri  s þ Di0 ei0 þ ð1  Xi0 Þci2N0
sSafety threshold for residual battery capacity when returning (7) v ij Xij  vij  vij Xij cði; jÞ2A
to the depot (8) Si  ai  Ei ci2N0
P P
M A large number (9) fji  fij ¼ di ci2N0
ðj;iÞ2A ði;jÞ2A

Binary decision variables: (10) fij  CL Xij cði; jÞ2A



ai  0; wi  0; ri  0ci2N0
(11)
Xij 1/0 indicate whether arc (i, j) is traversed by an EV or not Xij 2f0; 1g; vij  0; tij  0; eij  0; fij  0cði; jÞ2A
Non-negative continuous decision variables:
In the model given above, Constraints (1) and (2) ensure that
vij Traveling speed (km/h) on arc (i, j)
each customer is visited only once. Constraint (3) is the piecewise
tij Traveling time (min) spent on arc (i, j)
linearization of the timeespeed relationship surrogated by using a
eij ECR of EV when traveling on arc (i, j) with a load fij
set of lines to approximate (inner or outer) the nonlinear curve of
654 Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663

tij ¼ Dij/vij, as discussed in Section 3. The constant, i.e., 60, in approximation (or outer approximation) is used in the timeespeed
Constraint (3) is used to convert the time unit from hours to mi- linearization. It should be noted that the traveling time must be
nutes. Note that Constraint (3) takes effect only for Xij ¼ 1 and it will minimized before applying the post-optimization procedure. The
always hold for Xij ¼ 0 because M is a large number. Constraint traveling time in the solution is allowed to contain part of the
(4e1) requires that the arrival time at the first customer by a vehicle waiting time because they both make an equally weighted contri-
must be greater than or equal to the traveling time spent on the first bution to the objective value.
arc. Constraints (4e2) and (4e3) strictly determine the time flow
between customers. Constraint (5) bounds the vehicle’s ECR to the X
speed and load for each selected arc, as discussed in Section 2. The Min: Total Travel time ¼ tij (9)
ði;jÞ2A
first two terms, i.e., Kq vij þ Bq , represent the part of ECR contributed
by the speed vij. The third term, i.e., 4fij , represents the part of ECR Subject to Constraints (1)e(11) and Constraint (12), as follows:
contributed by the load fij. The last term, i.e., ð1  Xij Þ, guarantees eij " #
P P P P
to take a positive value for Xij ¼ 1 and to be zero if Xij ¼ 0. Con- (12) F X0j þ CB fe Dij eij þ fd tij þ ðgi þ wi Þ
straints (6e1)e(6e5) strictly calculate the flow of the remaining j2N 0 ði;jÞ2A ði;jÞ2A i2N 0
power rates for the EVs among customers, and they guarantee that  objðpÞ,
the residual battery capacity when returning to the depot is above
the given safe level s. It should be noted that Constraints (6e3) and Constraint (12) guarantees that the new solution p0 based on p is
(6e5) can be removed for a non-strict flow without affecting the still an optimal solution of the MILP model formulated in Eq. (8) and
objective cost. Constraint (7) restricts the traveling speed within Constraints (1)e(11), where obj (p) indicates the objective value of
the lower/upper speed limit range. Constraint (8) ensures that the solution p.
service starting time is within the customers’ time window. Next, the variables xij, vij, and tij are fixed to their current values
Constraint (9) determines the vehicle’s load flow throughout the (i.e., treated as input parameters) and the solver is called to solve
tour. Constraint (10) restricts the EV’s load to not exceeding its the simplified MILP model represented by Eq. (8) and Constraints
maximum load capacity, C, and forces fij to be zero when arc (i, j) is (1)e(2), (4)e(6), and (8)e(11). Note that Constraints (3) and (7)
not selected in any tour. Constraint (11) gives the domains of all the should not be included, in order to obtain an accurate solution p*
decision variables. based on p’. It should also be noted that solving the simplified MILP
It should be noted that all of the expressions in the MILP model model is highly efficient because the routes and speeds are already
formulated in Eq. (8) and Constraints (1)e(11) are linear, and thus fixed, and the solver only needs to re-optimize the rest variables,
the model can be solved directly by commercial optimization such as the departure, arrival, and waiting times, and the ECR. The
solvers such as Lingo and CPLEX. However, due to the approxima- main steps of the post-optimization procedure are outlined in
tion methods used in the linearization, the solution delivered Fig. 6.
directly from an optimization solver needs to be checked of its
Proposition 1. Feasibility: For a feasible solution p to the MILP
feasibility. In the following subsection, a post-optimization pro-
model formulated using Eq. (8) and Constraints (1)e(11) as well as
cedure is provided for eliminating the approximation errors and
with inner approximation on the timeespeed relationship, a feasible
guaranteeing the solution feasibility, as well as some theoretical
solution p* of the original EVRPTW-ECR (i.e., without using the
analysis on the solution optimality.
timeespeed approximation and using a nonlinear ECR function) al-
ways exists that has the same routes and traveling speeds as p and a
4.3. Eliminating approximation errors lower (or equal) objective value.
Proof. Let tij and eij be the travel time and the ECR of a traveled arc,
A solution of the MILP model might not be an accurate solution e.g., (i, j), in solution p. According to the inner approximation method,
of the original EVRPTW-ECR problem because the piecewise line- the inequality tijt*ij holds, where t*ij is the actual traveling time on arc
arization method may cause differences between the approximated (i, j) calculated by t*ij ¼ Dij/vij, and eije*ij, where e*ij is the accurate
travel time (bounded by Constraint (4)) and the actual travel time ECR on arc (i, j) calculated by the nonlinear ECR function. The solution
(calculated by tij ¼ 60Dij/vij). Thus, a post-optimization procedure, p can always be transformed into p* by using the following three steps:
as stated in Fig. 6, can be applied to the obtained solution, which is (1) replace tij with t*ij for all arcs on the routes of solution p, (2) replace
designated as p, to eliminate the difference, and thereby obtain an eij with e*ij for all arcs on the routes of solution p, and (3) add tij  t*ij to
accurate and feasible solution p*. The post-optimization procedure the waiting time at customer/depot nodes in order to keep the arrival
first re-calculates the traveling time (or traveling speed) for all time at all customers unchanged. Thus, p* is a feasible solution of the
selected arcs by using tij )60Dij =vij (or vij )60Dij =tij ) if the inner original EVRPTW-ECR. Furthermore, the total time in p* is the same as
that in p, and e*ij is equal to or smaller than eij, so the objective value of
solution p* is less than or equal to that of p.
//Solution obtained by Eq. (8) and Constraints (1)–(11) According to Proposition 1, the post-optimization procedure in
Fig. 6 can always be applied to convert any feasible solution of the
Post-optimization procedure ( )
MILP model with inner approximation into a feasible solution of the
(1) Minimize the traveling time by solving the MILP model given by Eq. (9) under original EVRPTW-ECR.
Constraints (1)–(12)
(2) For all selected arcs (i, j), let tij 60Dij/vij (for inner approximation only) Proposition 2. Optimality: For an optimal solution p* of the original
EVRPTW-ECR problem, a feasible solution p always exists for the MILP
For all selected arcs (i, j), let vij 60Dij/tij (for outer approximation only)
model formulated using Eq. (8) and Constraints (1)e(11), and with
(3) Fix variables xij, vij, and tij to their current values
outer approximation based on the timeespeed relationship that has
(4) Remove Constraints (3) and (7), and solve the simplified MILP model to obtain
the same routes and traveling time as p*.
*
solution Proof. Let t *ij , v*ij , and e*ij be the travel time, speed, and ECR of a
(5) Return *
traveled arc, e.g., (i, j), respectively, in solution p*. Obviously, the
inequality t *ij ¼ Dij =v*ij holds, and t *ij and v*ij will satisfy Constraint (3) if
Fig. 6. Main steps of the post-optimization procedure. the outer approximation method is used. The e*ij can always be
Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663 655

increased to a larger value in order to satisfy Constraint (5). Thus, the was associated with a demand, a time window during which the
new solution, designated as p, is a feasible solution of the MILP model service must start, and a service time. It was assumed that the
and it has the same routes and traveling speeds as those of p*. distance unit was kilometer and the time unit was minute.
According to Proposition 2, it is possible to use the MILP model The homogeneous EVs that served these customers were assumed
with outer approximation to find a solution, p, that has the same route to be the BYD e6 model and their parameters are listed in Table 5.
and traveling time as those of the optimal solution of the original The maximum ranges of a no-load BYD e6 under different velocities
EVRPTW-ECR problem. In addition, according to the following steps, are shown in Table 2 in Section 2 and the ECR function was line-
the solution p can be transformed into an optimal solution of the arized using the regression line obtained in Fig. 3 which is:
original EVRPTW-ECR problem: (1) update vh  vmax by vij )Dij =tij
and (2) update eij with the accurate value calculated by using the ECR ¼ 2.634  105v þ 1.1711  103 (10)
nonlinear ECR function based on vij and fij . Thus, the new solution is an
optimal solution of the original EVRPTW-ECR problem under the The regression line was explained in Section 2. The vehicle was
condition that all of the speeds on the traversed arcs do not exceed the assumed to have a load capacity of 200 units in order to be
upper speed limit. This rule can be used to judge whether the obtained consistent with Solomon’s VRPTW instances. The cost rates for the
solution is guaranteed to be optimal or not. vehicle/driver and the variable cost rates for electricity consump-
tion and the driver’s wage were initialized based on an investiga-
Proposition 3. Actual Feasibility: For a feasible solution p of the
tion of a real logistics company in Beijing.
MILP obtained with the inner approximation on the timeespeed
relationship, after the post-optimization procedure, p is actually a
feasible solution of the problem EVRPTW if the setting for the safe 5.2. Computational results and analysis
threshold on the EV’s residual battery capacity, e.g., s, satisfies s  ε,
where ε is the maximum deviation in the ECR function linearization. First, the CPLEX solver was used to solve all 29 instances with
Proof. This is straightforward. The maximum difference in the only the first 25 customers, i.e., n ¼ 26 (including the depot) under
approximation error incurred by the linearization of the ECR function the parameter setting of ε ¼ 0.1%. A time limit of 7200 s was set for
can be compensated for by setting the safe threshold s, which gua- all the test instances, such that the best feasible solution found so
rantees that the EV returns safely to the depot before its battery is far would be returned with a relative optimality gap (R. Gap, %) if
empty. Furthermore, the inner approximation guarantees a safe trav- the optimal solution was not found within the time limit. Both the
eling time for each arc so the time window requirements of the cus- outer and inner approximations were used and compared in the
tomers can actually be met. Thus, the solution is actually feasible for experiments. The results are shown in Table 6, where the Obj.
the EVRPTW-ECR problem. column indicates the objective value, i.e., the minimum total cost
(¥), of the final solution, the V. column shows the number of EVs
used, the Dis. column indicates the total distance traveled by all EVs,
5. Computational experiments the Tr. T. column indicates the total traveling time of all EVs,
including the waiting time and service time, the T. column shows
Computational experiments were conducted based on known the consumed CPU time in seconds, and the R. Gap column presents
benchmark instances in order to test the proposed MILP model for the relative optimality gap for the final solution at termination.
EVRPTW-ECR. The MIP solver CPLEX (version 12.6.0.1) was used to As seen from Table 6 that using either the outer or inner
solve the benchmark problem instances. All of the computational approximation with the MIP solver obtained the optimally solved
experiments were performed on a PC server with two 2.2 GHz Intel solutions (indicated by R. Gap ¼ 0.0%) for 20 (out of 29) test in-
16 Core Processors and 110 GB of memory. stances within the given time limit. The other nine instances were
terminated with the best feasible solutions obtained so far and
5.1. Test instances their associated relative optimality gaps. The computational time
used for different problem instances varied considerably, where it
The test instances were derived from three sets of Solomon’s ranged from a few seconds to 2 h depending on the problem
VRPTW problem instances (Solomon, 1987), including nine in- structure. Clearly, the solutions differed with the inner and outer
stances from the R1-type set (randomly generated customer loca- approximations. However, the differences were quite small among
tions), 12 instances from the C1-type set (clustered customer the 20 optimally solved solutions, with an average difference of
locations), and eight instances from the RC1-type set (mixture of only 0.45. For the solutions that terminated at the time limit, the
random and clustered locations). All of the instances comprised 100 differences were relatively large, with an average of 31.72. In
customers that are to be served by a homogeneous fleet of EVs, addition, among the solutions with identical routes (i.e., with the
where they started from and ended at one depot. Each customer same total distance), the solutions obtained based on the outer

Table 5
Parameters of the BYD E6 model EV used in the experiments.

Notation Description Value

4 Additional ECR for one unit of load 1.6E-06


s Minimum remaining power (%) when vehicle returns to depot 5%
CL Load capacity of vehicle 200
CB Battery capacity of vehicle (kWh) 57
vmin Minimum speed (km/h) 20
vmax Maximum speed (km/h) 100
F Fixed cost per electric vehicle plus driver (RMB) 300
fe Cost of consuming battery energy per unit (RMB/kWh) 4
fd Driver’s variable wage rate (RMB/m) 0.3
K Slope of the surrogate line of ECR function 2.6340E-05
B Intercept of the surrogate line of ECR function 1.1711E-03
656 Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663

Table 6
Computational results on the 25-customer instances with the parameter ε ¼ 0.1%.

Problem Ins. Outer approximation Inner approximation

Type ID Obj. V. Dis. Tr. T. T. R. Gap Obj. V. Dis. Tr. T. T. R. Gap

(¥) (km) (m) (s) (%) (¥) (km) (m) (s) (%)

R1-type R101 2759.58 7 601.53 1025.86 1 0.0 2767.79 7 609.67 1019.90 1 0.0
R102 2090.50 5 524.65 897.63 3876 0.0 2090.46 5 524.65 886.70 4277 0.0
R103 1704.08 4 431.77 776.94 7280 35.3 1704.06 4 431.77 767.97 7293 34.0
R104 1410.68 3 445.43 635.43 7297 26.8 1410.59 3 445.43 632.51 7204 26.7
R105 1776.97 4 519.77 746.22 10 0.0 1776.97 4 519.77 738.89 10 0.0
R106 1705.22 4 452.23 719.22 7205 36.4 1705.19 4 452.23 709.35 7209 33.9
R107 1380.80 3 419.97 643.69 7213 27.7 1380.80 3 419.97 639.08 7214 28.0
R108 1334.52 3 376.09 617.82 7283 23.5 1334.54 3 376.09 612.84 7313 23.7
R109 1694.46 4 443.66 725.72 7216 30.1 1694.37 4 443.66 714.07 7216 22.0
R110 1674.48 4 420.92 682.26 7306 38.3 1391.21 3 419.48 616.61 7259 25.8
R111 1391.26 3 429.07 638.46 7234 31.2 1393.22 3 432.89 647.00 7215 24.6
R112 1332.01 3 369.05 582.61 7227 30.6 1332.01 3 369.05 579.53 7206 31.2
AVG 1687.88 3.9 452.84 724.32 5762 23.3 1665.10 3.8 453.72 713.71 5785 20.8
C1-type C101 1818.70 3 237.52 2587.66 6 0.0 1818.67 3 237.52 2582.27 9 0.0
C102 1786.48 3 221.64 2499.27 237 0.0 1786.43 3 221.64 2491.75 372 0.0
C103 1774.10 3 207.89 2483.80 635 0.0 1774.06 3 207.89 2476.75 361 0.0
C104 1771.34 3 204.92 2480.47 3488 0.0 1771.30 3 204.92 2473.52 2453 0.0
C105 1802.45 3 237.52 2517.13 12 0.0 1802.40 3 237.52 2509.07 12 0.0
C106 1835.38 3 251.31 2623.63 9 0.0 1835.35 3 251.31 2618.38 7 0.0
C107 1795.45 3 215.66 2567.97 55 0.0 1795.41 3 215.66 2562.06 24 0.0
C108 1784.27 3 217.71 2513.65 445 0.0 1784.24 3 217.71 2507.54 248 0.0
C109 1779.19 3 214.37 2491.09 910 0.0 1779.14 3 214.37 2483.82 884 0.0
AVG 1794.15 3.0 223.17 2529.41 644.11 0.0 1794.11 3.0 223.17 2522.79 486 0.0
RC1-type RC101 1326.65 3 365.23 646.19 2 0.0 1326.59 3 365.23 636.90 3 0.0
RC102 1300.24 3 337.57 619.99 170 0.0 1300.19 3 337.57 612.31 1134 0.0
RC103 1291.22 3 328.73 614.20 362 0.0 1291.18 3 328.73 607.23 4080 0.0
RC104 1275.77 3 308.25 601.65 902 0.0 1275.73 3 308.25 594.06 122 0.0
RC105 1389.98 3 421.84 674.04 4739 0.0 1389.97 3 421.84 666.32 5028 0.0
RC106 1302.34 3 341.98 629.29 5602 0.0 1302.28 3 341.98 618.30 70 0.0
RC107 1259.55 3 297.38 578.00 390 0.0 1259.51 3 297.38 570.62 3972 0.0
RC108 1257.17 3 294.99 581.77 218 0.0 1257.11 3 294.99 571.76 971 0.0
AVG 1300.37 3.0 337.00 618.14 1548 0.0 1300.32 3.0 337.00 609.69 1923 0.0

Note: Bold face indicates the optimally solved values.

approximation always had slightly lower objective values than which were solved with the parameter ε ¼ 0.1% and using inner
those obtained based on the inner approximation. However, all of approximation. It was observed that the selected arcs were opti-
the differences were controlled within the range specified by the mized with different traveling speeds, the battery capacities of EVs
parameter ε. In particular, the R101 instance was solved optimally were consumed along the routes, and the remaining power rate (rj)
based on both the outer and inner approximations, but the objec- declined but was guaranteed to be above the safe threshold (i.e., 5%)
tive gap (i.e., 8.22) was much higher than the average (i.e., 0.45). when returning to the depot. In this solution, three vehicles were
This is because the R101 instance had tighter time windows which used to serve the customers at a fixed cost of 900, an electricity
leaded to the two solutions (based on the inner and outer consumption cost of 57  4  67.63% ¼ 154.19, a traveling time-
approximation respectively) had different routes. related cost of 0.3  (281.04 þ 100 þ 0.05) ¼ 114.33, and a total
In Table 7, a detailed solution of the R101 instance was provided cost of 1168.52.
as an example for readers to take a look into the scheduling details, Next, the performance of the MILP model was tested based on

Table 7
Detailed solution for the R101 instance with 10 customers (Obj. ¼ 1168.52, n ¼ 11, ε ¼ 0.1%, inner).

Veh. Arc (i, j) Distance Load fij Speed vij (km/h) Tra. T. tij (min) Ari. T. ai T. Window (Sj, Ej) Ser. T. gi (min) Waiting T. wi (min) En. used Remaining rj
No. Dij (km) Dij  eij

1 (0, 2) 18.00 10 55.01 19.63 60 (50, 60) 10 0.03 4.74% 95.26%


(2, 6) 16.40 3 33.97 28.97 99 (99, 109) 10 0.00 3.40% 91.86%
(6, 0) 11.18 0 55.01 12.19 0 (0, 230) e 0.00 2.93% 88.93%
2 (0, 5) 20.62 66 55.01 22.49 44 (34, 44) 10 0.01 5.62% 94.38%
(5, 7) 20.62 40 45.83 26.99 81 (81, 91) 10 0.01 5.03% 89.35%
(7, 8) 12.21 35 68.08 10.76 101.76 (95, 105) 10 0.00 3.69% 85.66%
(8, 10) 26.25 26 70.84 22.23 134 (124, 134) 10 0.00 8.08% 77.58%
(10, 1) 15.56 10 54.91 17.00 161 (161, 171) 10 0 4.10% 73.48%
(1, 0) 15.23 0 55.01 16.61 0 (0, 230) e 0 3.99% 69.49%
3 (0, 9) 32.02 48 55.01 34.92 97 (97, 107) 10 0 8.63% 91.37%
(9, 3) 15.00 32 58.60 15.36 122.36 (116, 126) 10 0 4.15% 87.22%
(3, 4) 25.00 19 56.36 26.62 158.98 (149, 159) 10 0 6.72% 80.50%
(4, 0) 25.00 0 55.01 27.27 0 (0, 230) e 0 6.55% 73.95%
Total 253.07 281.04 100 0.05 67.63%
Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663 657

Table 8
Experimental results with different values of ε (n ¼ 26, outer approximation).

ID ε ¼ 5% ε ¼ 1% ε ¼ 0.5% ε ¼ 0.1% ε ¼ 0.5% ε ¼ 0.01%

Obj. T.D. % Obj. T.D. % Obj. T.D. % Obj. T.D. % Obj. T.D. % Obj. T.D. %

C101 1816.28 3.95 1818.61 0.57 1818.49 0.40 1818.65 0.06 1818.63 0.04 1818.66 0.01
C102 1783.95 5.00 1786.38 1.00 1786.29 0.50 1786.41 0.10 1786.40 0.05 1786.43 0.01
C103 1771.73 5.00 1774.01 1.00 1773.93 0.50 1774.04 0.10 1774.02 0.05 1774.05 0.01
C104 1769.00 5.00 1771.25 1.00 1771.17 0.50 1771.28 0.10 1771.26 0.05 1772.16 0.01
C105 1799.74 5.00 1802.34 1.00 1802.25 0.50 1802.38 0.10 1802.36 0.05 1802.39 0.01
C106 1824.97 3.09 1835.16 0.67 1835.23 0.35 1835.32 0.06 1835.31 0.03 1835.33 0.01
C107 1793.48 2.95 1795.37 0.60 1795.30 0.31 1795.39 0.06 1795.38 0.03 1795.41 0.01
C108 1782.04 4.42 1784.19 0.89 1784.10 0.46 1784.21 0.09 1784.20 0.04 1784.23 0.01
C109 1776.74 5.00 1779.09 1.00 1779.00 0.50 1779.12 0.10 1779.11 0.05 1779.13 0.01
AVG 1790.88 4.38 1794.04 0.86 1793.97 0.45 1794.09 0.09 1794.07 0.04 1794.20 0.01

Table 9
Experimental results with different values of ε (n ¼ 26, inner approximation).

ID ε ¼ 5% ε ¼ 1% ε ¼ 0.5% ε ¼ 0.1% ε ¼ 0.5% ε ¼ 0.01%

Obj. E.R. % Obj. E.R. % Obj. E.R. % Obj. E.R. % Obj. E.R. % Obj. E.R. %

C101 1820.26 1.54 1818.69 0.05 1818.78 0.16 1818.67 0.01 1818.69 0.01 1818.67 0.00
C102 1787.62 0.18 1786.44 0.00 1786.46 0.00 1786.43 0.00 1786.47 0.00 1786.44 0.00
C103 1775.04 0.00 1774.06 0.00 1774.08 0.00 1774.06 0.00 1774.09 0.00 1774.06 0.00
C104 1772.27 0.00 1771.30 0.00 1771.32 0.00 1771.30 0.00 1771.33 0.00 1771.30 0.00
C105 1803.52 0.00 1802.41 0.00 1802.42 0.00 1802.40 0.00 1802.44 0.00 1802.40 0.00
C106 1836.69 0.92 1835.54 0.29 1835.39 0.12 1835.35 0.02 1835.37 0.01 1835.34 0.00
C107 1796.25 0.00 1795.42 0.00 1795.43 0.00 1795.41 0.00 1795.44 0.00 1795.41 0.00
C108 1785.49 0.38 1784.26 0.00 1784.29 0.02 1784.24 0.01 1784.26 0.01 1784.24 0.00
C109 1780.16 0.00 1779.15 0.00 1779.16 0.00 1779.14 0.00 1779.17 0.00 1779.14 0.00
AVG 1795.26 0.34 1794.14 0.04 1794.15 0.03 1794.11 0.00 1794.14 0.00 1794.11 0.00

the nine C1-type instances with different values of ε. All of the outer approximation (or inner approximation). A smaller value of
instances were solved to optimality by using the outer and inner parameter ε will result in a closer approximated travel time to the
approximations with the parameters ε ¼ 5.0%, 1.0%, 0.5%, 0.1%, real travel time. The outer and inner approximations have the same
0.05%, and 0.01%, respectively. The post-optimization procedure level of accuracy so either of them can be used to approximate the
was not applied in order to demonstrate how the parameter ε nonlinear timeespeed relationship.
affected the accuracy of the piecewise linearization used for the Next, an additional set of 10 new test instances were generated
timeespeed relationship. The experimental results are shown in in this paper, which were denoted as RA-type, where the customers
Tables 8 and 9 for outer and inner approximations, respectively. The were randomly scattered in a larger rectangular region, i.e.,
T.D. % column entries were calculated as R:Tr:T:Tr:T:  100%, and they 180  180, in order to test the EV’s driving range limit. This is
R:Tr:T:
indicate the time deviation causing by piecewise linearization, because the customer locations of the Solomon’s instances are all
P within the EV’s driving range so they do not reach the range limit.
where Tr. T. is the approximate travel time obtained by tij and
ði;jÞ2A Each new instance had 10 customers with an equal demand of 10. It
P
R. Tr. T is the accurate travel time calculated by ðDij =vij Þ. It can was assumed that the customers did not have time windows and
ði;jÞ2A that the vehicle’s load capacity was unlimited. Thus, the EV’s
be observed that the time deviations were always under the control driving range was the only restrictive condition. First, the 10 new
of the parameter ε, as the traveling times was always lower (or instances were solved with flexible traveling speeds (i.e., speed was
higher) than the accurate value by a maximum of eε (or ε) for the the decision variable). Next, the instances were solved with

Table 10
Solutions obtained under flexible and fixed traveling speeds (n ¼ 11, ε ¼ 0.1%, inner).

Fixed traveling speed Fixed driving

Problem Flexible 20 km/h 40 km/h 60 km/h 80 km/h 100 km/h Range (308 km)

Instance Speed Cost Dev.% Cost Dev.% Cost Dev.% Cost Dev.% Cost Dev.% Cost Dev.%

RA101 952.96 1341.67 40.8 1166.06 22.4 1171.57 22.9 1552.88 63.0 1979.18 107.7 1479.84 55.3
RA102 841.59 952.31 13.2 1125.75 33.8 1109.48 31.8 1506.52 79.0 1969.37 134.0 1103.22 31.1
RA103 964.53 1417.36 46.9 1209.25 25.4 1578.52 63.7 1622.57 68.2 2152.64 123.2 1572.47 63.0
RA104 733.12 878.05 19.8 1044.63 42.5 1030.86 40.6 1051.64 43.4 1493.00 103.7 1025.89 39.9
RA105 1236.98 1482.89 19.9 1258.19 1.7 1620.43 31.0 1667.88 34.8 2546.50 105.9 1614.05 30.5
RA106 765.72 902.47 17.9 1114.27 45.5 1111.80 45.2 1506.82 96.8 1575.36 105.7 1134.56 48.2
RA107 1162.39 1386.70 19.3 1186.29 2.1 1168.65 0.5 1531.11 31.7 2005.83 72.6 1496.32 28.7
RA108 874.33 969.08 10.8 1174.71 34.4 1182.21 35.2 1563.09 78.8 1614.03 84.6 1510.75 72.8
RA109 1218.04 1460.16 19.9 1241.07 1.9 1596.72 31.1 1630.42 33.9 2115.28 73.7 1590.69 30.6
RA110 887.91 976.23 9.9 1141.22 28.5 1124.45 26.6 1526.55 71.9 1574.69 77.3 1492.56 68.1
AVG 963.76 1176.69 21.8 1166.14 23.8 1269.47 32.9 1515.95 60.2 1902.59 98.8 1402.04 46.8
658 Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663

Table 11
Average performance of the solutions for the RA-type instances (n ¼ 11, ε ¼ 0.1%, inner).

Average based on 10 test instances

Fixed Speed Vehicles used Total dis. Total time Total cost Fixed cost Energy cost Time cost CPU time

vij ¼ 100 3.7 753.41 452.05 1902.59 1110 656.97 135.61 16.8 s
vij ¼ 80 2.9 660.43 495.32 1515.95 870 497.35 148.60 33.9 s
vij ¼ 60 2.3 620.16 620.16 1269.47 690 393.42 186.05 42.7 s
vij ¼ 40 2 586.34 879.50 1166.14 600 302.29 263.85 67.8 s
vij ¼ 20 1.5 559.24 1677.72 1176.69 450 223.38 503.32 21.8 s
Fixed range 2.7 640.32 698.42 1402.04 810 382.51 209.53 198.8 s
Flexible 1.3 547.49 1034.66 963.77 390 263.36 310.40 97.7 s

constant speeds of 20 km/h, 40 km/h, 60 km/h, 80 km/h, and increase in costs dwarfed the cost savings due to the shortened
100 km/h. Then, the EVs were assumed to have a fixed driving delivery time. The solutions obtained with a fixed driving range
range, i.e., 308 km (the number given by the manufacturer and also had much higher costs, with an average of 46.8%, compared
measured under the New European Driven Cycling (NEDC) stan- with the solutions produce with flexible speed. This is because
dard), and the instances were solved with an additional constraint more EVs were used. This indicates that the effect of considering
which restricts the total distance traveled by each EV not exceeding flexible speed in EVRPTW has significant improvements on tradi-
the fixed driving range. tional solutions with fixed speeds, with average cost saving rates
The computational results are shown in Table 10, Table 11, and of 18.10%, 17.35%, 24.08%, 36.43%, and 49.34% when
Fig. 7, and they demonstrate that the variations in the traveling compared to the cost obtained under fixed speeds of 20 km/h,
speed had very large effects on the transportation costs. The solu- 40 km/h, 60 km/h, 80 km/h, and 100 km/h, respectively. The
tions produced with a fixed traveling speed always had higher average cost saving rate could also be as high as 31.26% when
objective values (costs) than the solutions obtained with flexible compared to the solution of assigning a fixed range to EVs.
speed, with average cost deviations of 21.8%, 23.8%, 32.9%, 60.2%, The average performance levels in the experiments are sum-
and 98.8% at fixed speeds of 20 km/h, 40 km/h, 60 km/h, 80 km/h, marized in Table 11, which shows the average number of EVs used,
and 100 km/h, respectively. This is because a higher speed total distance traveled, and total delivery time. The Total Cost col-
increased the ECR though, but it decreased the EV’s driving range umn indicates the objective function. The Fixed cost, Energy cost,
more significantly (and thus more EVs were required). The overall and Time cost columns are related to the first, second, and third

1200

1000

Fixed Cost
800
Cost

600

Energy Cost

400

Time Cost
200

0
20 40 60 80 100
Fixed travel speed (km/h)
Fig. 7. Cost variation trends with respect to different traveling speeds.
Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663 659

parts of the objective function. Fig. 7 shows the cost variation Customers that are close to each other in terms of the time window
trends of the fixed cost, the variable cost, and the time cost under (although they may be far from each other in terms of distance) are
different traveling speeds. In general, a higher travel speed will lead more likely to be served contiguously in the route. The main steps
to a higher fixed cost and a higher energy cost, but a lower time of DPO-INS is described in Fig. 8 as follows.
cost. Thus, the optimal trade-off of total cost is determined by their In Fig. 8, the DPO-INS algorithm starts with an initial solution,
cost rates. The experimental data showed that the speed, 20 km/h, S0, constructed by dispatching each customer with a vehicle (in
resulted the lowest total cost. Step 1), and then S0 is taken as the incumbent solution, S. The
non-improvement counter B and the current operator O are
5.3. A dynamic heuristic solution initialized in Step 2. After that, a loop is launched in step 3 to
implement an INS, including from Step 3.1 to Step 3.15. In the INS
A heuristic solution approach was proposed for EVRPTW-ECR, loop, one of the four operators was applied in turn to select a set,
which was based on the iterative neighborhood search (INS) algo- W, of nodes into set U (in Steps 3.1e3.7). And then, the CPLEX
rithm by Xiao and Konak (2016). The basic idea of INS is to apply solver is applied to partial optimization on the selected nodes
partial optimization (PO) to the routing variable, i.e., the variable Xij, while fixing the rest of other nodes (in Steps 3.8e3.10). In Steps
by mixing using multiple local search operators. Moreover, a dy- 3.11e3.12, the number W is adjusted dynamically to ensure that
namic strategy is used in the algorithm, by which the scope of the solving time used by CPLEX is between tmin, e.g., 0.1 s, and
partial optimization, i.e., the number of selected customers, is tmax, e.g., 10 s. That is, the number W will increase by one if the
adjusted dynamically according to the instant CPU time consumed. last solution time is less than 0.1 s and will reduce by one if the
Thereby, the proposed heuristics is named as dynamic partial last solution time is greater than 10 s. In Steps 3.13 and 3.14, the
optimization with iterative neighborhood search (DPO-INS for incumbent solution is updated and the non-improvement
short). In the heuristics, four operators were used in turn to select a counter B is increased by one if no better solution is found. The
part of the variable instances of Xij, and a MIP solver was used to loop is stopped when no further improvement can be made to
optimize the selected variable instances while fixing the remaining the incumbent solution after Bmax continuous attempts (i.e.,
variable instances at their current values. Three of the operators B  Bmax) or the elapsed time is greater than the given time limit
were proposed by Xiao and Konak (2016), i.e., nearest neighbor Tmax (in Step 3.15). Finally, S is the best solution found and it is
selection (OP-1), single range selection (OP-2), and double range the output of the algorithm. In the INS loop, the MIP solver will
selection (OP-3). The fourth operator, named as time neighbor se- consume most of the consumed CPU time (in Step 3.10), so the
lection (OP-4), was newly proposed to select customer nodes that algorithm’s computational complexity depends on the times of
have neighbored time windows. This is because the time windows partial optimizations implemented by the MIP solver, which is
are strong restrictions on the solution’s feasibility in EVRPTW-ECR. controllable as an input parameter by the user.

Algorithm of DPO-INS :
Input parameters: Tmax, Bmax, W, tmin, tmax
Output result: S
1) Let S0 Initialization (): find a feasible incumbent solution S0
2) Let S S0, B 1, O 1
3) Loop Begin
3.1) Let U null
3.2) If O = 1 Then apply OP-1 to select a number W of nodes from set N into set U
3.3) If O = 2 Then apply OP-2 to select a number W of nodes from set N into set U
3.4) If O = 3 Then apply OP-3 to select a number W of nodes from set N into set U
3.5) If O = 4 Then apply OP-4 to select a number W of nodes from set N into set U
3.6) If O = 4 Then let O 1
3.7) Else let O O+1
3.8) Fix all Xij decision variables for i N and j N
3.9) Unfix Xij decision variables for i U and j U
3.10)Call the MIP solver to get a new solution S and save the elapsed CPU time tc.

3.11)If tc tmax Then let W W 1


3.12)Else If tc tmin Then let W W+1
3.13)If S improves upon S Then let B 0 and S S
3.14)Else let B B+1
3.15)If B Bmax or elapsed-time Tmax Then Break: jump out of loop and stop
4) Loop End

Fig. 8. Main steps of the DPO-INS algorithm.


660 Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663

Table 12
Performance of the DPO-INS heuristics (Bmax ¼ 100, n ¼ 26, ε ¼ 0.1%, inner).

10 runs of the DPO-INS heuristics CPLEX

Problem ins. Best AVG Worst Dev. % AVG t. Obj. T. R. Gap%

R1-type R101 2767.78 2767.79 2767.79 0.0% 27 2767.79 1 0.0


R102 2090.46 2235.71 2382.72 0.0% 81.9 2090.46 4277 0.0
R103 1704.06 1708.79 1720.50 0.0% 74.8 1704.06 7293 34.0
R104 1670.71 1671.71 1675.56 18.4% 154.6 1410.59 7204 26.7
R105 1779.23 1779.47 1779.53 0.1% 51.1 1776.97 10 0.0
R106 1705.18 1750.99 2028.03 0.0% 89.6 1705.19 7209 33.9
R107 1380.80 1559.46 1696.33 0.0% 85.7 1380.80 7214 28.0
R108 1334.54 1441.64 1647.61 0.0% 350.3 1334.54 7313 23.7
R109 1694.37 1703.11 1739.38 0.0% 62.8 1694.37 7216 22.0
R110 1391.33 1628.11 1703.51 0.0% 82.3 1391.21 7259 25.8
R111 1393.22 1649.68 1699.59 0.0% 82.9 1393.22 7215 24.6
R112 1332.01 1472.31 1658.91 0.0% 99.8 1332.01 7206 31.2
AVG 1686.97 1780.73 1874.95 1.5% 103.6 1665.10 5785 20.8

C1-type C101 1818.67 1818.67 1818.67 0.0% 36.5 1818.67 9 0.0


C102 1786.43 1787.23 1794.40 0.0% 109.5 1786.43 372 0.0
C103 1774.06 1794.36 1808.54 0.0% 75.1 1774.06 361 0.0
C104 1771.30 1776.52 1786.83 0.0% 107.2 1771.30 2453 0.0
C105 1802.40 1802.40 1802.40 0.0% 32.8 1802.40 12 0.0
C106 1835.34 1835.34 1835.34 0.0% 39.9 1835.35 7 0.0
C107 1795.41 1799.61 1802.40 0.0% 39.2 1795.41 24 0.0
C108 1784.24 1795.85 1808.76 0.0% 46.9 1784.24 248 0.0
C109 1779.14 1779.14 1779.14 0.0% 60.6 1779.14 884 0.0
AVG 1794.11 1798.79 1804.05 0.0% 60.9 1794.11 486 0.0

RC1-type RC101 1326.59 1326.59 1326.59 0.0% 29.6 1326.59 3 0.0


RC102 1300.19 1300.22 1300.42 0.0% 50.1 1300.19 1134 0.0
RC103 1291.17 1291.18 1291.18 0.0% 54.2 1291.18 4080 0.0
RC104 1275.73 1283.75 1345.39 0.0% 53.1 1275.73 122 0.0
RC105 1670.21 1670.23 1670.34 20.2% 61.3 1389.97 5028 0.0
RC106 1302.28 1302.43 1303.81 0.0% 51.7 1302.28 70 0.0
RC107 1259.51 1259.53 1259.70 0.0% 50.1 1259.51 3972 0.0
RC108 1257.11 1257.33 1258.26 0.0% 70.8 1257.11 971 0.0
AVG 1335.35 1336.41 1344.46 2.5% 52.6 1300.32 1923 0.0

Note: Bold face indicates the best values found. In the header, Bmax ¼ 100 indicates the stop condition, n ¼ 26 indicates total number of nodes, ε ¼ 0.1% and inner indicate the
maximum allowed error ratio in the inner-approximation for the time-speed relation, respectively.

The DPO-INS algorithm was repeatedly run for 10 times to the time window constraint is more restrictive. For the C-type and
obtain 10 solutions for each of the test instances. The experimental RC-type instances, the distance differences among clustered cus-
results were shown in Table 12 and they were also compared with tomers are even more trivial in comparison to the time window
the solutions obtained by CPLEX with inner approximation shown differences, that is why the Op-1 operator performed even worse
in Table 6. It can be observed that DPO-INS required much less CPU on the C-type than the RC-type or R-type. In Fig. 1A and Table 1A of
time to obtain solutions that are only slightly worse than the so- the Appendix section, a detailed algorithm progress of solving the
lutions by CPLEX. There are 25 (out of 29) instances found by DPO- C101 instance of C1-type by the DPO-INS algorithm is provided.
INS that have met the best-known solutions. The average CPU time Table 13 compares the performance of DPO-INS and CPLEX for
was 127 s for DPO-INS and 3042 s for CPLEX. Note that some so- the 29 Solomon’s instances with full customer sizes, i.e., 100 cus-
lutions by the heuristics appear to be slightly better than the tomers. The parameter settings were Bmax ¼ 400, n ¼ 101, and
optimal solutions by CPLEX, e.g., solutions 2767.78 and 2767.79 for
the R101 instance, but they are the same one. This is because of two
reasons: (1) The CPLEX solver itself has a numerical accuracy Average mes of
setting for integer programming, and (2) the ε-linearization of the Improvements
time-speed relation may lead to an approximate error. Both of these
values should be considered as BEST. That is, the values 2767.78 and 10.0
2767.79 in first row of Table 12 are all the best value for R101
8.0
instance.
Fig. 9 shows the average times of improvements by using 6.0
different operators. The OP1 operator (i.e., nearest neighbor selec-
tion), which obtained the best performance according to Xiao and 4.0
Konak (2016), however showed the worst performance on the 2.0
EVRPTW-ECR problem. The new time neighbor selection operator R-type
(i.e., the OP4) obtained moderate performance compared with the 0.0 C-type
other two operators. The reason for why the OP-1 operator made a OP-4 RC-type
OP-3
bad performance is largely because the Solomon’s VRPTW in- OP-2
OP-1
stances have much tighter time windows than that of PRP instances
of Bektas and Laporte (2011), such that optimizing the connections
between nearby neighbors may not work out good result because Fig. 9. Average times of improvements by different operators.
Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663 661

Table 13
Performance of the DPO-INS heuristics (Bmax ¼ 400, n ¼ 101, ε ¼ 0.1%, inner).

Problem ins. 10 runs of the DPO-INS heuristics CPLEX

Best AVG Worst Dev. % AVG t. Obj. T. R. Gap %

R1-type R101 6909.85 7051.50 7585.80 8.9 348 6343.34 1178 0.0
R102 6148.54 6362.19 6508.46 0.7 837 6191.18 7201 42.0
R103 5101.52 5494.90 5748.37 11.8 1119 5783.76 7202 41.7
R104 3999.54 4457.06 4927.66 24.8 915 5318.18 7201 34.8
R105 5770.30 5789.04 5828.32 4.7 611 5511.49 7201 33.3
R106 5109.21 5419.08 5791.12 6.4 1010 5460.81 7201 37.5
R107 4635.39 4823.78 5020.76 13.7 1054 5370.34 7201 38.6
R108 3953.52 4290.76 4662.96 15.4 964 4672.18 7203 30.8
R109 4978.87 5205.21 5693.82 21.8 742 6370.90 7201 47.6
R110 4947.48 5107.91 5307.40 22.6 751 6394.50 7204 49.2
R111 4348.53 4787.10 5283.80 22.8 844 5633.15 7202 41.9
R112 4309.16 4621.28 4931.67 24.0 816 5666.70 7207 43.1
AVG 5017.66 5284.15 5607.51 13.1 834 5773.84 6276 33.9

C1-type C101 6492.30 6636.32 6910.95 0.0 359 6492.31 61 0.0


C102 6492.30 7097.39 7475.34 0.0 612 6492.31 7202 1.4
C103 6525.44 6950.55 7306.77 0.5 815 6492.31 7203 4.9
C104 6738.25 6892.15 7059.38 3.8 978 6492.31 7203 5.4
C105 6492.30 6916.70 7788.39 0.0 1124 6492.31 169 0.0
C106 6492.30 6795.14 7299.86 0.0 576 6492.31 7201 3.8
C107 6492.30 6738.77 7019.31 0.0 495 6492.31 7200 3.7
C108 6492.30 6730.25 6987.13 0.0 633 6492.31 7202 4.6
C109 6513.73 6807.27 7443.06 0.3 545 6492.31 7202 5.3
AVG 6525.69 6840.50 7254.47 0.5 682 6492.31 5627 3.2

RC1-type RC101 5992.82 6249.75 6681.27 6.9 748 6434.91 7201 34.9
RC102 5552.55 5637.20 5924.72 7.4 852 5993.40 7202 35.1
RC103 4819.79 5154.47 5471.57 17.5 904 5843.81 7202 35.2
RC104 4425.00 4738.18 4830.17 15.3 840 5224.64 7210 28.2
RC105 5901.01 5967.14 6261.86 12.0 753 6704.98 7201 41.1
RC106 5543.10 5658.76 5925.73 11.3 596 6248.91 7202 38.1
RC107 5119.91 5318.45 5522.18 44.1 826 9157.25 7202 58.9
RC108 4461.91 4785.69 5099.43 37.6 980 7153.60 7203 47.8
AVG 5227.01 5438.71 5714.62 19.0 812 6595.19 7203 39.9

Note: Bold face indicates the best values found.

ε ¼ 0.1%, and inner approximation was employed. It can be observed traveling speeds of 20 km/h, 40 km/h, 60 km/h, 80 km/h, and
that the obtained solutions by the DPO-INS heuristics are better than 100 km/h, respectively. Two piecewise linearization methods were
those by CPLEX, with average improvements of 13.1% and 19.0% also proposed, comprising the inner and outer approximations for
for the R1 and RC problem types, respectively. For the C1 problem the nonlinear timeespeed relationship, with a controllable range of
type, the solutions obtained by DPO-INS were almost the same as the approximation errors. Based on the theoretical analysis and
near-optimal solutions found by CPLEX, which were solved with experimental comparisons, two generalized remarks can be
small R. Gap values. The main aim of this experiment was to show concluded: (1) SVR is an important feature of EVs, having a sig-
that the proposed MILP model could be solved by heuristic algo- nificant impact on efficient utilization of electricity energy and cost
rithms for large problems within controllable CPU time, so the four reduction, and (2) the proposed MILP model and heuristic approach
operators were used with an equal chance in the DPO-INS heuristics. for EVRPTW-ECR are able to solve optimal solutions for problem
However, using a competitive strategy among the operators is ex- instances up to 25 customers and relatively good solutions for
pected to improve the solution efficiency substantially (Dellaert and problem instances up to 100 customers. Future studies may
Jeunet, 2000; Xiao and Konak, 2016). consider more practical factors related to the ECR function, such as
the air condition status and environmental temperature, as well as
6. Conclusion developing more efficient solution approaches for large-sized
EVRPTW-ECR problems.
This study investigated the EVRPTW by considering a new
feature of EVs, i.e., the speed-varying travel range (SVR). Thus, the
maximum traveling range of EVs was not considered as a constant
Acknowledgments
limit, as found in most existing EVRP models, but instead it was
dynamically determined by ECRs (energy/electricity consumption
This study was partly supported by the National Natural Science
for traveling per unit distance) of EVs throughout the tours. An
Foundation of China under Grant Nos. 71871003, 71571004, and
energy consumption optimization model was developed for the
71672006.
EVRPTW, where the ECR of an EV was taken as a nonlinear function
of the continuous speed with the linear effect of the load (referred
to as EVRPTW-ECR). Comparative experiments based on real situ-
ation simulations showed that the dynamic range settings with Appendix
flexible speed achieved great cost reductions, with averages
of 17.9%, 19.2%, 24.8%, 37.6%, and 49.7% compared with the
solutions produced when considering a constant range and fixed
662 Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663

Table 1A
A detailed algorithm steps of solving the C101 instance of C1-type by the DPO-INS algorithm.

SN B Op. W Obj. SN B Op. W Obj. SN B Op. W Obj.

1 0 2 8 35869.41 54 0 3 8 15491.20 107 0 3 8 9706.30


2 2 1 8 35199.96 55 2 2 9 15419.38 108 2 2 10 9673.49
3 0 2 8 34561.52 56 0 3 8 15402.40 109 6 1 8 9663.57
4 2 1 8 33601.97 57 2 2 9 15385.26 110 3 1 8 9658.24
5 0 2 8 32730.14 58 0 3 8 15360.03 111 0 2 10 9648.04
6 0 3 10 31964.97 59 1 1 8 14765.10 112 0 3 8 9640.01
7 0 1 8 31003.50 60 1 3 8 14746.21 113 3 3 8 9626.75
8 0 2 8 30671.29 61 1 1 8 14433.92 114 2 2 10 9593.61
9 0 3 10 30587.82 62 0 2 9 14129.05 115 4 3 8 9539.64
10 1 1 8 29678.24 63 2 1 8 14101.38 116 1 2 10 9246.42
11 0 2 8 29111.65 64 0 2 9 13807.54 117 2 1 8 9230.74
12 0 3 10 29086.71 65 0 3 8 13771.48 118 5 3 8 9227.39
13 0 1 8 28409.67 66 3 3 8 13768.28 119 10 2 10 8889.70
14 0 2 8 27464.71 67 1 1 8 13729.95 120 0 3 8 8884.48
15 0 3 10 26484.32 68 0 2 9 13624.27 121 6 2 10 8857.67
16 0 1 8 26271.34 69 0 3 8 13350.53 122 24 3 8 8835.71
17 0 2 8 25641.22 70 1 1 8 13104.63 123 23 3 8 8832.29
18 0 3 10 25532.90 71 0 2 9 13047.50 124 29 1 8 8830.09
19 0 1 8 24923.79 72 0 3 8 13043.09 125 8 2 11 8529.14
20 0 2 8 24037.97 73 2 2 9 12790.49 126 2 1 8 8479.70
21 0 3 10 23939.99 74 0 3 8 12780.21 127 5 3 8 8478.63
22 0 1 8 23712.93 75 1 1 8 12731.01 128 22 2 11 8471.74
23 0 2 8 23377.12 76 4 2 9 12098.46 129 8 3 8 8471.65
24 0 3 10 23338.74 77 0 3 8 12054.95 130 34 2 11 8428.51
25 1 1 8 22973.29 78 3 3 8 12007.21 131 16 3 9 8427.89
26 1 3 10 22933.57 79 2 2 9 11931.66 132 11 3 9 8403.82
27 1 1 8 22861.84 80 0 3 8 11920.24 133 6 2 11 8386.40
28 0 2 8 22170.41 81 1 1 8 11826.43 134 20 3 9 8354.36
29 0 3 9 22147.91 82 0 2 9 11516.27 135 27 3 9 8353.74
30 0 1 8 21818.51 83 2 1 8 11493.20 136 3 3 9 8319.43
31 0 2 8 21546.94 84 1 3 8 11461.65 137 2 2 11 8315.44
32 2 1 8 21246.58 85 1 1 8 11452.29 138 6 1 10 7965.33
33 0 2 8 21228.18 86 1 3 8 11442.55 139 7 1 10 7958.15
34 0 3 9 20909.36 87 3 3 8 11429.61 140 28 2 11 7943.86
35 1 1 8 20214.40 88 1 1 8 11413.94 141 7 2 11 7938.22
36 0 2 8 20214.01 89 0 2 9 11379.60 142 0 3 9 7923.10
37 0 3 9 19766.96 90 0 3 8 11304.67 143 5 1 10 7923.06
38 1 1 8 19455.35 91 2 2 9 10958.82 144 13 3 9 7888.23
39 0 2 8 18509.32 92 2 1 8 10627.95 145 45 1 10 7849.44
40 0 3 9 18137.86 93 3 1 8 10626.64 146 64 2 11 7824.42
41 1 1 8 17728.63 94 8 2 9 10324.24 147 8 3 9 7822.79
42 0 2 9 17364.10 95 0 3 8 10294.17 148 9 1 10 7814.14
43 0 3 9 17321.60 96 6 2 9 10290.00 149 7 1 10 7787.09
44 1 1 8 17159.58 97 0 3 8 10279.31 150 100 2 11 7727.64
45 0 2 9 16858.70 98 2 2 9 10234.96 151 11 2 11 7377.00
46 0 3 8 16835.10 99 0 3 8 10194.06 152 12 3 9 7335.65
47 2 2 9 16487.72 100 5 1 8 10123.80 153 30 2 11 7315.48
48 0 3 8 16450.49 101 1 3 8 10081.66 154 34 1 10 6901.06
49 1 1 8 16137.45 102 2 2 9 10052.25 155 45 3 9 6898.53
50 0 2 9 15832.16 103 4 3 8 10035.98 156 55 3 9 6885.57
51 0 3 8 15818.76 104 5 1 8 10031.66 157 114 2 12 6825.01
52 1 1 8 15614.85 105 21 3 8 10031.10 158 15 2 12 6492.31
53 0 2 9 15569.71 106 30 2 10 9712.49

Note: SNdSequence number of improvement; BdNumber of continuous attempts without improvement; WdNumber of nodes selected to be unfixed; Obj.dobjective
function value.
Y. Xiao et al. / Journal of Cleaner Production 225 (2019) 647e663 663

Haaren V. R., 2012. http://large.stanford.edu/courses/2012/ph240/prokopiak1/docs/


HowFarWeDrive_v1.2.pdf [accessed 4/4/2019].
Hiermann, G., Puchinger, J., Ropke, S., Hartl, R.F., 2016. The electric fleet size and mix
35000
vehicle routing problem with time windows and recharging stations. Eur. J.
Oper. Res. 252 (3), 995e1018.
Hof, J., Schneider, M., Goeke, D., 2017. Solving the battery swap station location-
30000 routing problem with capacitated electric vehicles using an AVNS algorithm
for vehicle-routing problems with intermediate stops. Transp. Res. Part B
Methodol. 97, 102e112.
25000 International Energy Agency (IEA), 2016. CO2 emissions from fuel combustion-
Objective function

Highlight. www.iea.org/publications/freepublications/publication/co2-
emissions-from-fuel-combustion-highlights-2016.html. (Accessed 8 July 2017).
20000 International Energy Agency (IEA), 2018. Global EV outlook 2018. www.
indiaenvironmentportal.org.in/files/file/globalevoutlook2018.pdf. (Accessed 25
July 2018).
Jabali, O., Van Woensel, T., Kok, A.G. de, 2012. Analysis of travel times and CO2
15000 emissions in time-dependent vehicle routing. Prod. Oper. Manag. 21 (6),
1060e1074.
Jie, W., Yang, J., Zhang, M., Huang, Y., 2018. The two-echelon capacitated electric
10000 vehicle routing problem with battery swapping stations: formulation and
efficient methodology. Eur. J. Oper. Res. 272 (3), 879e904.
Kara, I., Kara, B.Y., Yetis, M.K., 2007. Energy minimizing vehicle routing problem. In:
5000 Dress, A., Xu, Y., Zhu, B. (Eds.), Combinatorial Optimization and Applications:
FIRST International Conference, LNCS 4616. Springer, Berlin/Heidelberg,
pp. 62e71, 2007.
0 Keskin, M., Çatay, B., 2016. Partial recharge strategies for the electric vehicle routing
0 100 200 300 400 problem with time windows. Transport. Res. C Emerg. Technol. 65, 111e127.
Solution time (s) Li, Y., Zhan, C., Jong, de M., Lukszo, Z., 2016. Business innovation and government
Fig. 1A. The converging progress of solving the C101 instance of C1-type by the DPO- regulation for the promotion of electric vehicle use: lessons from Shenzhen,
INS algorithm. China. J. Clean. Prod. 134 (A), 371e383.
Liao, C.S., Lu, S.H., Shen, Z.J.M., 2016. The electric vehicle touring problem. Transp.
Res. Part B Methodol. 86, 163e180.
References Lin, C., Choy, K., Ho, G., Chung, S., Lam, H., 2014. Survey of green vehicle routing
problem: past and future trends. Expert Syst. Appl. 41 (4), 1118e1138.
Adler, J.D., Mirchandani, P.B., 2014. Online routing and battery reservations for Lin, J., Zhou, W., Wolfson, O., 2016. Electric vehicle routing problem. Transport. Res.
electric vehicles with swappable batteries. Transp. Res. Part B Methodol. 70, Procedia 12, 508e521.
285e302. Liu, Y., 2014. Test of BYD e6 EV–A Relatively Reliable Choice. auto.enorth.com.cn/
Andelmin, J., Bartolini, E., 2017. An Exact algorithm for the green vehicle routing system/2014/08/06/012065564.shtml. (Accessed 26 July 2018).
problem. Transport. Sci. 51 (4), 1288e1303. Liu, K., Yamamoto, T., Morikawa, T., 2017. Impact of road gradient on energy con-
Barth, M.J., Boriboonsomsin, K., 2008. Real-world CO2 impacts of traffic congestion. sumption of electric vehicles. Transport. Res. Part D 54, 74e81.
Transport. Res. Rec.: J. Transport. Res. Board 163e171, 2058. Lu, X., Zhou, K., Yang, S., Liu, H., 2018. Multi-objective optimal load dispatch of
Barth, M.J., Younglove, T., Scora, G., 2005. Development of a Heavy-Duty Diesel microgrid with stochastic access of electric vehicles. J. Clean. Prod. 195,
Modal Emissions and Fuel Consumption Model. California PATH Program, 187e199.
Institute of Transportation Studies, University of California at Berkeley. Tech- Marmaras, C., Xydas, E., Cipcigan, L., 2017. Simulation of electric vehicle driver
nical Report. UCB-ITSPRR-2005-1. behaviour in road transport and electric power networks. Transport. Res. C
Bektas, T., Laporte, G., 2011. The pollution-routing problem. Transport. Res. Part B Emerg. Technol. 80, 239e256.
45, 1232e1250. Montoya, A., Gue ret, C., Mendoza, J.E., Villegas, J.G., 2017. The electric vehicle
Bruglieri, M., Pezzella, F., Pisacane, O., Suraci, S., 2015. A variable neighborhood routing problem with nonlinear charging function. Transp. Res. Part B Meth-
search branching for the electric vehicle routing problem with time windows. odol. 103, 87e110.
Electron. Notes Discrete Math. 47, 221e228. Musk, E., Straube, J.B., 2012. Model S efficiency and range. www.tesla.com/blog/
Bruglieri, M., Mancini, S., Pezzella, F., Pisacane, O., Suraci, S., 2017. A three-phase model-s-efficiency-and-range. (Accessed 26 July 2018).
matheuristic for the time-effective electric vehicle routing problem with partial Schiffer, M., Walther, G., 2017. The electric location routing problem with time
recharges. Electron. Notes Discrete Math. 58, 95e102. windows and partial recharging. Eur. J. Oper. Res. 260 (3), 995e1013.
Casals, L.C., Martinez-Laserna, E., García, B.A., Nieto, N., 2016. Sustainability analysis Schneider, M., Stenger, A., Goeke, D., 2014. The electric vehicle-routing problem
of the electric vehicle use in Europe for CO2 emissions reduction. J. Clean. Prod. with time windows and recharging stations. Transport. Sci. 48 (4), 500e520.
127, 425e437. Solomon, M.M., 1987. Algorithms for the vehicle routing and scheduling problems
Conrad, R.G., Figliozzi, M.A., 2011. The recharging vehicle routing problem. In: IIE with time window constraints. Oper. Res. 35 (2), 254e265. Dataset available at.
Annual Conference. Proceedings. Institute of Industrial and Systems Engineers www.cba.neu.edu/~msolomon/problems.htm.
(IISE). Reno, Nevada, vol 1. Wu, Z., Wang, M., Zheng, J., Sun, X., Zhao, M., Wang, X., 2018. Life cycle greenhouse
D1EV, 2008. List of EVs with Ranges over 400 Km. car.d1ev.com/find/00_2-100_5_ gas emission reduction potential of battery electric vehicle. J. Clean. Prod. 190,
1_00_00_00_A.html. (Accessed 26 July 2018). 462e470.
Dellaert, N., Jeunet, J., 2000. Solving large unconstrained multilevel lot-sizing Xiao, Y., Konak, A., 2015. A simulating annealing algorithm to solve the green
problems using a hybrid genetic algorithm. Int. J. Prod. Res. 38, 1083e1099. vehicle routing & scheduling problem with hierarchical objectives and
Demir, E., Bektas, T., Laporte, G., 2014. A review of recent research on green road weighted tardiness. Appl. Soft Comput. 34, 372e388.
freight transportation. Eur. J. Oper. Res. 237 (3), 775e793. Xiao, Y., Konak, A., 2016. The heterogeneous green vehicle routing and scheduling
Desaulniers, G., Errico, F., Irnich, S., Schneider, M., 2016. Exact algorithms for electric problem with time-varying traffic congestion. Transport. Res. E Logist. Trans-
vehicle-routing problems with time windows. Oper. Res. 64 (6), 1388e1405. port. Rev. 88, 146e166.
Erdog an, S., Miller-Hooks, E., 2012. A green vehicle routing problem. Transport. Res. Xiao, Y., Konak, A., 2017. A genetic algorithm with exact dynamic programming for
E Logist. Transport. Rev. 48 (1), 100e114. the green vehicle routing & scheduling problem. J. Clean. Prod. 167, 1450e1463.
 Ortun
Felipe, A., ~ o, M.T., Righini, G., Tirado, G., 2014. A heuristic approach for the Xiao, Y.-Y., Zhang, R.-Q., Kaku, I., 2011. A new approach of inventory classification
green vehicle routing problem with multiple technologies and partial re- based on loss profit. Expert Syst. Appl. 38, 9382e9391.
charges. Transport. Res. E Logist. Transport. Rev. 71, 111e128. Xiao, Y., Zhao, Q., Kaku, I., Xu, Y., 2012. Development of a fuel consumption opti-
Ferna  2018. A more realistic approach to electric vehicle contribution to
ndez, R.A., mization model for the capacitated vehicle routing problem. Comput. Oper. Res.
greenhouse gas emissions in the city. J. Clean. Prod. 172, 949e959. 39 (7), 1419e1431.
Fetene, M.G., Kaplan, S., Mabit, L.S., Jensen, F.A., Prato, G.C., 2017. Harnessing big Yang, J., Sun, H., 2015. Battery swap station location-routing problem with capaci-
data for estimating the energy consumption and driving range of electric ve- tated electric vehicles. Comput. Oper. Res. 55, 217e232.
hicles. Transport. Res. Transport Environ. 54, 1e11. Yuan, X., Zhang, C., Hong, G., Huang, X., Li, L., 2017. Method for evaluating the real-
Figliozzi, M.A., 2010. Vehicle routing problem for emissions minimization. Trans- world driving energy consumptions of electric vehicles. Energy 141, 1955e1968.
port. Res. Rec.: J. Transport. Res. Board 2197, 1e7. Zhang, J., Zhao, Y., Xue, W., Li, J., 2015. Vehicle routing problem with fuel con-
Fiori, C., Ahn, K., Rakha, H.A., 2016. Power-based electric vehicle energy consump- sumption and carbon emission. Int. J. Prod. Econ. 170, 234e242.
tion model: model development and validation. Appl. Energy 168 (15), Zheng, Y., Dong, Z.Y., Xu, Y., Meng, K., Zhao, J.H., Qiu, J., 2014. Electric vehicle battery
257e268. charging/swap stations in distribution systems: comparison study and optimal
Goeke, D., Schneider, M., 2015. Routing a mixed fleet of electric and conventional planning. IEEE Trans. Power Syst. 29 (1), 221e229.
vehicles. Eur. J. Oper. Res. 245 (1), 81e99.

You might also like