Simulation

Special Issue Article

Simulation: Transactions of the Society for
Modeling and Simulation International
2018, Vol. 94(7) 625–636
The location optimization of electric Ó The Author(s) 2018
DOI: 10.1177/0037549717743807
vehicle charging stations considering journals.sagepub.com/home/sim

charging behavior

Zhihui Tian, Wenbin Hou, Xiaoning Gu, Feng Gu and Baozhen Yao

Abstract
The electric vehicle is seen as an effective way to alleviate the current energy crisis and environmental problems.
However, the lack of supporting charging facilities is still a bottleneck in the development of electric vehicles in the
Chinese market. In this paper, the cloud model is used to first predict drivers’ charging behavior. An optimization model
of charging stations is proposed, which is based on waiting time. The target of this optimization model is to minimize the
time cost to electric vehicle drivers. We use the SCE-UA algorithm to solve the optimization model. We apply our
method to Dalian, China to optimize charging station locations. We also analyze the optimized result with or without
behavior prediction, the optimized result of different numbers of electric vehicles, and the optimized result of different
cost constraints. The analysis shows the feasibility and advantages of the charging station location optimization method
proposed in this paper.

Keywords
Behavior prediction, cloud mode, optimization of charging station location, SCE-UA algorithm, waiting time

1. Introduction construction of charging stations for the development of

EVs cannot be ignored.
In recent years, the electric vehicle (EV) has been seen as an In recent years, research on the location problem of EV
effective way to alleviate the current resource crisis, energy charging stations has attracted many scholars. He and col-
crisis, and environmental problems due to its high efficiency, leagues1 incorporated the local constraints of supply and
energy savings, low noise, and zero emissions. EVs have demand on public EV charging stations into facility loca-
been widely promoted and used throughout the world, and tion models. This paper compared three classic facility
are in line with China’s current development ethos. location models—the set covering model, the maximal
Many countries are actively promoting EVs by incen- covering location model, and the p-median model—and
tive policies and tax subsidies (such as no purchase tax, showed that the p-median solutions are more effective than
unlimited purchase number, and half parking fees, etc.). the other two models. Frade and colleagues2 used a maxi-
China is a populous country, but large resource consump- mal covering model to optimize siting of public charging
tion and energy issues have become bottlenecks restricting stations for Lisbon, Portugal. In this paper, nighttime
the rapid and healthy development of China. Because of demand and daytime demand were estimated separately to
the advantages of EVs, the Chinese government has made correspond to different residential and business uses. Xi
great efforts to promote their development by issuing pre- and colleagues3 developed a simulation-optimization
ferential policies and lowering relevant consumption taxes. model to determine the locations of slow chargers. These
However, the lack of charging stations has become a con-
straint on the development of new EVs. Compared with fuel
vehicles, EVs are dependent on charging facilities. Users’ Automotive Engineering College, Dalian University of Technology, PR
travel decisions are very affected by the distribution of China
charging facilities. At present, some domestic and foreign
Corresponding author:
parking lots have been built with a variety of charging Baozhen Yao, Automotive Engineering College, Dalian University of
equipment, such as charging piles. However, these charging Technology, Dalian 116024, PR China.
facilities cannot meet the demand of EVs. Therefore, the Email: [email protected]
626 Simulation: Transactions of the Society for Modeling and Simulation International 94(7)

authors further explored interactions between the optimiza- different. Each charging station has a different number of
tion criterion and the budget available. Cavadas and col- customers at the same time, and the average waiting time
leagues4 split the day into time intervals to take into at each charging station is different. So it is necessary to
account the effects of peak hours. They consider the effect predict the driver’s choice behavior. To account for the
of demand transference. Results show that the method can uncertainty in the process of changing choice, cloud model
greatly expand the satisfied coverage. Hiwatari and col- theory, which is a new cognitive model proposed by Li
leagues5 analyzed the effective layout of charging stations and colleagues,14,15 and successfully applied in lots of dif-
based on the location of EVs running out of electricity in a ferent research16,17, was adopted in this paper.
road traffic simulator. Capar and colleagues6 applied a We use the cloud model to predict the driver’s charging
flow-refueling location model to charging station problems behavior. When the driver’s charging behavior is deter-
under mild assumptions such as charging stations being mined, the average waiting time for each charging station
incapacitated. The new formulation reduced the average is also determined. An optimization model of charging sta-
solution time by over 70%. Cruz-Zambrano and col- tion location is proposed, which is based on the waiting
leagues7 formulated two methods for locating fast-charging time. The target of this optimization model is to minimize
stations (maximizing service coverage and maximizing the time cost to EV drivers.
profit), and discussed the performance of the two proposed
methods. Wang and colleagues8 optimized the location of
battery exchange stations to serve tourism transport for 2. Electric vehicle driver charging
electric scooters. A battery exchanging scheme has rarely behavior prediction
been put forward alone, and is often mentioned together 2.1. Cloud model
with fast-charging technology. You and colleagues9 pro- 2.1.1. Introduction of the cloud model. The cloud model was
posed an O–D trip-based model which simultaneously proposed by Deyi Li, a member of the Chinese Academy of
determines the locations and types of recharging station Engineering. It is an uncertain transformation model deal-
and alternative-fuel vehicles (AFVs) recharging quantity at ing with qualitative concepts and quantitative description.
each recharging station. Chung and colleagues,10 Nie and col- There are three numerical characteristics in cloud
leagues,11 Sathaye and colleagues,12 Dong and colleagues,13 mode,18 expected value (Ex), entropy (En), and hyper
among others, proposed other methods for the problem. entropy (He). Here Ex is the most representative point of
Most studies about locating charging stations consider this qualitative concept. This point is the highest point of
the distance between the demand point and the supply the cloud. And the degree of membership of Ex is 1. Here
point. There are few studies that regard the waiting time
En represents a range that can be measured as qualitative
for EVs to queue in charging station as a factor. Most
concepts. The greater the entropy, the more macroscopic
researchers assume that EV drivers will choose the nearest
the concept, the wider the range that can be measured. In
charging station when they decide to charge. In the actual
addition, En reflects the uncertainty of the qualitative con-
situation, the nearest charging station is not always the best
cept, also called fuzziness; He is the entropy of entropy, it
choice. The charging time of EVs is long. It will take more
is used to express uncertainty. It represents the random-
than 20 min even in fast-charging mode. When there are a
ness of the sample, namely the discrete degree of cloud
lot of EVs to be charged, the drivers need to wait in line.
droplets. The three numerical methods are as follows:
For the EV drivers, the time consumed by waiting for
charging is more unbearable than the charging time. At
1X n
this time, if drivers choose a charging station with a subop- Ex = xi ð1Þ
timal distance, there will be little time to wait for charging. n i=1
Then the total time cost will be less than the situation of rffiffiffiffi
choosing the nearest charging station. Therefore, the wait- p 1X n
En = 3 jxi  Exj ð2Þ
ing time at the charging station should also be an impor- 2 n i=1
tant consideration when optimizing the charging station pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
location. He = S 2  En2 ð3Þ
The location of the charging station is designed to pro-
vide better service to EV drivers. And because of the 1 X n
S2 = (xi  Ex) ð4Þ
importance of the waiting time at the charging station and n  1 i=1
long charging time, the EV driver will choose the best
charging station for himself. Before optimizing the loca-
tion of the charging station, the driver’s choice behavior is 2.1.2. Normal cloud model. The normal cloud model is the
predicted so that the charging station can provide better most important cloud model. The normal cloud theory is
service for more drivers. Under different charging station based on universality of the normal distribution and uni-
location schemes, the driver’s choice behavior will be versality of the normal membership function. m =
HOU et al. 627

(xi  Ex)2
exp½  is defined as the normal curve of the
2En2
normal cloud (X , m).
The normal cloud is calculated as follows.

1. Generate the normal random number xi of expecta-

tion Ex and standard deviation En.
2. Generate the normal random number Eni of expec-
tation En and standard deviation He.
(xi  Ex)2
3. Calculate mi = exp½ , where ðxi , mi Þ is
2Eni 2
the cloud droplets.
Figure 1. One-dimensional cloud generator.

2.2. Prediction of EV drivers’ charging behavior

based on the cloud model
2.2.1. Classification of EV drivers’ charging selection
behavior. The tolerance of each driver of remaining power
is different. Some drivers may choose to charge when the
remaining power is 20%, while some drivers may do so at
10%. It is impossible to determine the number of EVs
charging at the same time, so the number of the customers
at charging stations cannot be determined. In this context,
we use a one-dimensional cloud model to classify the
charging behavior of EV drivers.
In this stage, the remaining power is the only factor. So
a one-dimensional cloud model is used to predict it. First
of all, using the distribution graph of the survey data, the
three digital characteristics are generated by a backward
cloud generator. Then prediction droplets are generated
through the forward cloud generator. The cloud generators
are shown in Figure 1.
Figure 2. Planar cloud generator: (a) planar backward cloud
generator; (b) planar forward cloud generator.
2.2.2. Prediction of EV drivers’ behavior
(1) Planar cloud model.

In real life, people often face a lot of data. But they cannot
extract the specific rule knowledge from the data because
of the lack of effective processing methods and techniques.
We often need such a specific rule as the foundation when
we do some related content forecasts. We always need to
rely on some prior knowledge to partition the data when
extracting an algorithm for data processing. But this divi-
sion is subjective and cannot reflect the distribution of Figure 3. Two-dimensional rule generator.
data. When the data dimension is high, it is more difficult
to divide the high-dimensional data space because of the A two-dimensional X conditional cloud generator and a
coupling relationship among the dimensions. Therefore, one-dimensional Y condition cloud generator can create a
Yang and colleagues19 presented a planar cloud model. complex qualitative rule generator. For example, the rule ‘‘A
The planar cloud model reflects a complex qualitative con- and B, then C’’ cloud generators are shown in Figure 3.
cept that comprises two qualitative atomic concepts. A multi-rule generator is formed when a number of
The forward and backward generator of the planar such two-dimensional cloud single-rule generators are
cloud are shown in Figure 2. combined.
628 Simulation: Transactions of the Society for Modeling and Simulation International 94(7)

Figure 5. Diagram of EVs queuing system in a charging station.

optimization model is established, and the model takes the

minimum time cost as the objective.

3.1. Calculation of electric vehicle driver waiting time

based on queuing theory
Figure 4. EV drivers’ selection of charging stations.
Here R = Q0i =sev ; Q0i = remaining power of electric vehicle i; 3.1.1. Basic introduction and assumptions. In the queuing sys-
sev = consumption power unit distance of each EV. tem of the charging station location problem, the input pro-
cess mainly refers to the behavior of EVs arriving at the
charging station. The charging behavior of EV drivers has
(2) Prediction of drivers’ selection behavior based on greater flexibility and randomness. This article assumes that
the planar cloud model. the behavior of EVs arriving at the charging station obeys
the Poisson distribution with parameter l. The queuing rules
We can know whether drivers choose to charge EVs at the of EVs in charging stations are those of a multi-service win-
moment through the predictions of the cloud model. When dow waiting system, and EV drivers obey the first-come
the EV driver chooses to charge his vehicle, there are sev- first-served rule. Assume that each charging station has C
eral optional charging stations in the range of the distance service windows (charging piles), and each charging pile is
that the remaining power can support, as shown in Figure independent of the others. Assume that the charging time
4. In order to determine the flow of each alternative charg- obeys the negative exponential distribution. Parameter m is
ing station, we need to predict which charging station EV the average service rate. Under normal circumstances, it will
drivers will choose. take 20 min for a charging pile to fill an EV from 20%
If EV drivers go to the nearest charging station to remaining power in fast-charging mode. Parameterm is three
charge, they may need to wait for charging. It will be a cars per hour. Queuing mode is as shown in Figure 5.
long waiting time even if the EV drivers choose fast-
charging mode. If some drivers choose the second nearest Nvj
l= ð5Þ
charging station to charge, they may not need to wait. N1 3 max tij
j
Then the total time cost for charging will be less than that
for the nearest station. Therefore, in this paper, the waiting Here Nvj = The number of EVs to charging stations j and
time is also considered as an influencing factor when con- N1 = The number of charging piles. Each charging pile’s
sidering the driving distance. price is 3–5 million. Compared to construction costs, the
cost of the charging pile is very small. Thus we assume
that the number of charging piles in each charging station
3. Charging station location optimization is the same.
model Here max tij = the longest traveling time of EV i from
j
Through the cloud model discussed above, we can predict
the current location to the charging station j.
the number of customers of each charging station. The ser-
vice capability of each charging station is limited. When
the number of EVs is more than the number of charging 3.1.2. Queuing model. From the above content we can see
piles, the drivers need to wait for charging. It will take that the charging station queuing system mode is M/M/C.
much of the drivers’ time. In this paper, the queuing the- The state transition of the system is as shown in Figure 6.
ory model is used to calculate the waiting time. The According to the queuing theory20:
HOU et al. 629

Figure 6. System state transition diagram.

" 
X       #1 1, if a charging station is built at alternative station j
C 1
1 l n 1 l C l 1 yj =
0, otherwise
P0 = + 1 ð6Þ
n=0
n! m C! m Cm
ð10Þ
8 X
h i1  l C N= yj ð11Þ
>
> 2
>
> Lp = ð C  1 Þ!ð Cm  l Þ lm P0
>
> m j2J
>
>
>
> l
> Here tij = traveling time of EV i from the current loca-
< Ls = Lp + m
>
ð7Þ tion to charging station j;
> Lp Q0i = remaining power of EV i;
>
> Wp =
>
> l e, = electricity consumption per unit time of EV;
>
>
>
> Ls s = charging speed of charging station;
> Ws =
>
>
: l hj = construction cost of charging station j;
m5l
wj = average waiting time in charging station j;
where Lp is the average queue length, that is the number Q = total power of an EV;
of waiting EVs in the system; N = total number of charging stations;
Ls is the average number of customers, that is the num- H = total investment cost; and
ber of EVs in the system; M = total number of EVs.
Wp is the average waiting time; Subject to:
and Ws is the average sojourn time, that is the dwell X
time of a EV in a system. xij 4 1 ð12Þ
j2J
In order to ensure the normal operation of the charging
station, the service rate is assumed be greater than the arri- X
yj hj 4 H ð13Þ
val rate. j2J
XX
xij 5 90%M ð14Þ
3.2. The optimization model of charging station i2I j2J

location Constraint (12) represents that an EV can be served by

This paper proposes an optimization model based on queu- not more than one charging station; constraint (13) con-
ing theory. The model can ensure that the government can strains the total construction cost; constraint (14) expresses
balance investment in the construction of charging stations the number of serviced EVs, which accounted for more
and can minimize the charging time cost of EVs: than 90% of all EVs.
XX XX
min Z = xij wj + xij tij 4. Model solution
j2J i2I i2I j2J
XX ð8Þ 4.1. Predictive solution on electric vehicle driver
+ xij ½Q  (Q0I  e, tij )=s
i2I i2J
behavior
4.1.1. The solution on charging behavior classification based on

1, if EV i choose charging station j to charge the one-dimensional cloud model. This paper uses question-
xij = ð9Þ naire surveys on the relationship between the remaining
0, otherwise
power and the probability of selecting charging in the light
Here xij is determined by the queuing model in of classification of charging behavior classification. Three
Section 3.1. digital characteristics of the one-dimensional cloud mode
630 Simulation: Transactions of the Society for Modeling and Simulation International 94(7)

are generated according to the survey results, and then the generator and the Bernoulli distribution, we randomly gen-
required number of cloud droplets are generated through erate 0 or 1 to indicate the randomness of whether to
the positive cloud model generator. charge at each charging station or not. Among them, 1 rep-
Different people have different degrees of tolerance for resents that the EV will be charged at some charging sta-
remaining power. For example, some people think that tion, otherwise 0. The greater the probability obtained
EVs must be charged when the remaining power is lower from the cloud model, the greater the probability that 1
than 30%, while others disagree. Therefore, in order to can be randomly generated.
ensure the randomness of the classification and to exclude Everyone has a different perception on waiting time
the contingency, this paper conducts cloud model simula- and distance. For example, some people think that a wait
tion behavior 10 times to get the average value of the of 20 min is too long and they cannot accept it; some peo-
simulation results. On this basis, we use the Bernoulli dis- ple may not think so. This paper predicts driver behavior
tribution to randomly generate 0 or 1 to indicate random- 10 times to obtain the average value of the simulation
ness in the decision of whether to charge or not. Here, 1 results. Moreover, it can ensure the accuracy of the predic-
represents that the EV will be charged, otherwise 0. The tion results and the randomness of samples, and exclude
greater the probability obtained from the cloud model, the contingency.
greater the probability that 1 can be randomly generated.

4.2. SCE-UA algorithm for electric vehicle charging

4.1.2. The solution on charging behavior classification based on
the two-dimensional cloud model. This paper forecasts the
station location
passenger flow of each charging station through the two- The optimization model of EV charging station location is
dimensional cloud model, considering waiting time and a nonlinear integer programming problem; it cannot be
distance. First, we use questionnaire surveys on waiting solved with the traditional analysis method. This paper uses
time and distance, and then divide waiting time into the SCE-UA algorithm to solve the optimization model.
‘‘excellent,’’‘‘good,’’‘‘average,’’ and ‘‘poor,’’ and divide The SCE-UA algorithm is also known as a hybrid evo-
the distance into ‘‘farthest’’ and ‘‘farther’’‘‘near,’’ and lutionary algorithm. Similar to other heuristic algorithms,
‘‘close.’’ The rules are as follows: the SCE-UA algorithm has also been researched and
applied by a number of scholars.21–23 It is a global optimi-
Rule 1: if the waiting time is ‘‘excellent’’ and the distance zation algorithm, and it combines the characteristics of
is ‘‘close,’’ the average probability of charging is P1 ; complex method, random search, and biological competi-
Rule 2: if the waiting time is ‘‘good’’ and the distance tive evolution.
is ‘‘close,’’ the average probability of charging is P2 ; The SCE-UA algorithm has strong capabilities for glo-
Rule 3: if the waiting time is ‘‘average’’ and the distance bal optimization. In the probabilistic optimization of
is ‘‘close,’’ the average probability of charging is P3 ; hydrological model parameter subspace, Kucezara24 com-
Rule 4: if the waiting time is ‘‘poor’’ and the distance is pared the search performances of the SCE-UA algorithm,
‘‘close,’’ the average probability of charging is P4 ; genetic algorithm, and MSX algorithm (using simplex
Rule 5: if the waiting time is ‘‘excellent’’ and the distance method and quasi-Newton method respectively). The
is ‘‘near,’’ the average probability of charging is P5 ; results showed that the SCE-UA algorithm had better
Rule 6: if the waiting time is ‘‘good’’ and the distance robustness and better convergence. Although the genetic
is ‘‘near,’’ the average probability of charging is P6 ; algorithm had a fast convergence rate in the initial stage of
Rule 7: if the waiting time is ‘‘average’’ and the distance evolution, it could not find the effective optimal solution
is ‘‘near,’’ the average probability of charging is P7 ; in the region near the optimal solution.
Rule 8: if the waiting time is ‘‘excellent’’ and the distance The essence of the SCE-UA algorithm is to regard the
is ‘‘farther,’’ the average probability of charging is P8 ; global search process as a natural evolutionary process.
Rule 9: if the waiting time is ‘‘good’’ and the distance The sample points first form a group. Then the group is
is ‘‘farther,’’ the average probability of charging is P9 ; divided into several subgroups, which are complexes.
Rule 10: if the waiting time is ‘‘excellent’’ and the dis- Each complex evolves independently, searches in different
tance is ‘‘farthest,’’ the average probability of charging directions, and gradually evolves into a certain stage in its
is P10 ; best search direction. After several generations of evolu-
Rule 11: if the waiting time is ‘‘good’’ and the distance tion, the groups are mixed. A new group is generated
is ‘‘farthest,’’ the average probability of charging is P11 . through the mixing process and the evolution process is
restarted. This process enhances the survival ability of the
The probability that EV drivers choose which station to population and gradually converges the solution to the glo-
charge at is forecasted through the two-dimensional multi- bal optimum by sharing the information obtained from
rule generator. Using the two-dimensional multi-rule each independent group.
HOU et al. 631

Specific steps of minimization problems of the algo-

rithm solution SCE-UA are as follows21:

Step 1: Parameters setting. Initialize the number of

complexes p ( p 5 1) and the number of points m in
each complex. Compute the sample size as s = pm.
Step 2: Generate a random sample of s points
X½x1 , x2 , x3 ,    , xm  in feasible solution space; compute
the function value fi of the optimization model at each
point xi .
Step 3: Rank the s function values fi (i = 1, 2, . . . , s) of
the optimization model in ascending order and store
them in array D = f(xi , fi ), i = 1, 2, . . . . . . , sg such that Figure 7. Charging network for the small case.
i = 1 represents the smallest function value.
Step 4: Partition D points into p complexes A1 , A2 ,
. . . , Ap , each containing m points. established, the smaller the average waiting time, and the
Step 5: Generation of offspring. Evolve each complex smaller the total time cost. Due to the total construction
Ak , k = 1, 2, . . . , p using the competitive complex evo- cost constraint, the time cost is minimized when eight
lution algorithm. charging stations are established. The minimum time cost
Step 6: Evolve each complex and shuffle them into is 1173.1 min. So, in this small case, the optimal charging
points, then generate new complexes; repeat step 5. station location scheme is to establish eight charging sta-
Step 7: Check for convergence. If convergence criteria tions, namely number 2, 3, 4, 5, 6, 7, 9, 10 alternative
are satisfied, stop; otherwise return to Step 3. charging stations.

5. Numerical tests
5.2. Case of Dalian
In this paper, two examples are given to verify the feasi-
Dalian is an important coastal city in east China, located
bility of the model and algorithm. We first verify the feasi-
in the south of Liaodong peninsula. It covers an area of
bility through a small network, then we optimize the EV
13,237 km2 and has a population of 6.69 million. Dalian is
charging station location in Dalian.
in the first batch of cities in China to popularize new EV
cars. The generous subsidies for new EVs is up to 147,500
5.1. Simple case yuan, which is about 50% of the price of a common elec-
tric car. New electric cars have the favor of many people
As shown in Figure 7, it is a small network consisting of 50
due to there being no-purchase-tax policies for them.
EVs and 10 alternative charging stations. Optimize the charg-
ing station location on this basis. We must ensure that 90% More than 2000 new-energy vehicles have been put in the
of all EVs are served. The objective, minimum total time cost fields of bus, postal, sanitation, units and the private pur-
including waiting time, should be obtained under the con- chase in Dalian. The vehicles include pure electric passen-
straint of total construction cost. The results are as follows. ger cars, plug-in hybrid buses, plug-in hybrid electric
Table 1 shows the optimal charging station location passenger cars, and others. There is great demand for EV
plan when establishing 4–8 charging stations. As shown in charging in Dalian. Therefore, it is necessary to optimize
Table 1, the greater the number of charging stations the charging stations location in the city.

Table 1. Result of small case.

Scenario Number of EVs Number of CSs Project Time cost (min) Average waiting time (min) Cost (million $)

1 50 4 3,4,5,10 1867.3 6.72 1900.26

2 50 5 3,4,5,6,10 1719.5 4.12 2325.64
3 50 6 3,4,5,6,9,10 1500.3 4.66 2821.72
4 50 7 2,3,4,5,6,9,10 1386.6 4.96 3262.87
5 50 8 2,3,4,5,6,7,9,10 1093.1 2.14 3704.98

CS, charging station; EV, electric vehicle.

632 Simulation: Transactions of the Society for Modeling and Simulation International 94(7)

Figure 8. Distribution of charging demand in the study area on different days.

5.2.1. Data collection and parameter calibration. In this of EV drivers. Predicted results are shown in Figure 11.
paper, gas stations existing in Dalian are selected as alter- When the remaining power is less than 23% of total electric-
native sites for charging stations. While the location of ity, the probability of the driver choosing to charge is very
EVs is changing, within a certain period of time it reached, high; when the remaining power is more than 23% of total
to a certain extent, a dynamic balance (as visualized in electricity, the probability will decrease with the increase of
Figure 8). The distribution of EVs and alternate charging remaining power. In Figure 11, the charge selection probabil-
stations in Dalian are shown in Figure 9. ity of the EV driver obtained by the cloud model is approxi-
Other parameters are shown in Table 2. mately the same as that of the survey data in Figure 10. The
choice probability in Figure 11 also preserves the random-
ness of driver selection behavior.
5.2.2. Results presentation. In the forecast of driver behavior,
Similarly, by the questionnaire survey and planar cloud
we first carried out the questionnaire survey and made pre-
model, we get the EV drivers’ charging behavior in alter-
dictions accordingly. The results of the survey about the
native stations. Then, through the SCE-UA algorithm, EV
charging choice behavior of the EV drivers are shown in
charging station location is optimized. We assume that
Figure 10. From the questionnaire, the cloud model of charg-
each charging station has 10 charging piles.
ing choice behavior classification of EV drivers is a decline
In the fixed cost constraints, we assume that the num-
cloud model. The expected value is 23%. Programmed with
ber of EVs is 1000. We optimize charging station location
MATLAB, we used the cloud model to predict the behavior
without the driver behavior prediction mode and with the
driver behavior prediction mode respectively. The optimi-
Table 2. Value of parameters.
zation results are shown in Table 3. If we do not predict
driver behavior, which assumes that the driver chooses the
Parameters Values nearest charging station when they need to charge, the
total time cost is 27.28% more than the total time cost in
Number of candidate charging stations 58
Standard speed of electric vehicle 50k mph driver behavior prediction mode, and the average waiting
Electricity consumption speed of electric 8.5 kW time is 1.68 min more. Therefore, before optimizing the
vehicle charging station location, it is necessary to forecast the
The initial electricity of electric vehicles 200 kWh charging behavior of the drivers.
Electricity consumption speed of battery 40 kW When the total number of EVs that need to be served is
distribution vehicle
different, the optimization result will be different. So we
HOU et al. 633

Figure 9. The distribution of EVs and alternate charging stations in Dalian.

Figure 10. Survey data.

Table 3. Optimized result with or without behavior prediction of EV drivers.

Scenario Behavior Number of Project Time cost (min) Average Cost (million $)
prediction charging waiting
stations time (min)

1 N 13 1,3,6,15,17,19,22,25,27,33,42,44,56 1.9365e + 005 4.43 56.3754

2 Y 13 1,3,15,17,19,22,25,27,33,36,42,51 1.5214e + 005 2.75 55.2123
634 Simulation: Transactions of the Society for Modeling and Simulation International 94(7)

Table 4. Optimized result of different number of EVs.

Scenario Number of Number of Project Time cost (min) Cost (million $)

electric charging
vehicles stations

1 500 10 1,3,15,17,22,25,27,33,42,51 0.9432e + 005 41.6077

2 1000 13 1,3,15,17,19,22,25,27,33,36, 1.5214e + 005 55.2123
42,51
3 1500 15 1,3,7,15,17,19,22,25,27,33,36,42,51,52,56 2.5823e + 005 67.1341
4 2000 18 1,3,7,13,15,17,19,22,25,27,33,36,37,42,46,51,52,56 3.7214e + 005 75.8935
5 2500 19 1,3,7,13,15,17,19,22,25,27,33,36,37,42,46,51,52,54,56 5.2156e + 005 8673.69

Table 5. Optimized result of different cost constraints.

Scenario Cost Number of Project Time cost (min) Average Cost (million $)
constraint charging waiting
stations time (min)

1 50 11 1,3,15,17,22,25,33,42,46,51 1.9175e + 005 3.58 47.8463

2 60 13 1,3,15,17,19,22,25,27,33,36,42,51 1.5214e + 005 2.75 55.2123
3 70 15 1,3,7,13,15,17,22,25,27,33,36,42,43,56 1.3523e + 005 2.06 68.5738

should also be a factor when optimizing the charging sta-

tion site. Table 5 shows the optimal scheme under differ-
ent investment cost constraints, when the number of EVs
is 1000.
As can be seen from Table 5, with the increase in the
cost constraint value, the number of charging stations will
increase, while the average waiting time and the total time
cost will decrease. When the cost constraint of investment
is $60 million, the number of charging stations is 13, the
time cost is 1.5214e + 005 min, and the average wait time
is 2.75 min. Compared to the case when the cost constraint
is $50 million, although the cost of investment increased
by 15.39%, the total time cost is reduced by about $20.7%,
and the average waiting time is reduced by 0.83 min.
Compared to the case when the cost constraint is $50 mil-
lion, although the total time cost increased by 12.50%, and
Figure 11. Cloud model.
the average waiting time increased by 0.69 min, the cost
of investment increased by 15.39%. Therefore, through the
comprehensive comparative analysis, it is more reasonable
optimize charging station location against the background when the upper limit of the investment cost is set at $60
of different numbers of EVs. The optimization results are million.
shown in Table 4.
As shown in Table 4, when more EVs need to be
charged, more charging station need to be constructed, and
6. Conclusions
the construction costs increase. The EV is seen as an effective way to alleviate the current
In this paper, when construct the optimization model, energy crisis and environmental problems. However, the
we assume that construction costs cannot be more than the lack of supporting charging facilities is still a bottleneck in
total cost constraints. When the government invests in EV the development of EVs in the Chinese market. In this
charging stations, it needs to ensure that the cost of invest- paper, the cloud model is used to predict drivers’ charging
ment is reasonable. Therefore, the cost of investment behavior. An optimization model of charging stations is
HOU et al. 635

proposed, which is based on the waiting time. The SCE- 8. Wang YW. Locating battery exchange stations to serve tour-
UA algorithm is used to solve the optimization model. The ism transport: a note. Transp Res D 2008; 13(3): 193–197.
method is applied to Dalian China to optimize the charging 9. You PS and Hsieh YC. A hybrid heuristic approach to the
station location. Through an example, we prove the neces- problem of the location of vehicle charging stations. Comput
sity of EV driver behavior prediction and the feasibility of Ind Eng 2014; 70(1): 195–204.
10. Chung SH and Kwon C. Multi-period planning for electric
the charging station location optimization method pro-
car charging station locations: a case of Korean
posed in this paper. In the case of optimization with or Expressways. Eur J Oper Res 2015; 242(2): 677–687.
without behavior prediction of EV drivers, the total time 11. Nie Y and Ghamami M. A corridor-centric approach to plan-
costs without behavior prediction of EV drivers is 27.28% ning electric vehicle charging infrastructure. Transp Res B
more than the total time cost in driver behavior prediction 2013: 57: 172–190.
mode, and the average waiting time is 1.68 min more. 12. Sathaye N and Kelley S. An approach for the optimal plan-
Therefore, before optimizing charging station location, it ning of electric vehicle infrastructure for highway corridors.
is necessary to forecast the charging behavior of drivers. Transp Res E 2013; 59(11): 15–33.
We also analyzed the constraints on investment cost. It is 13. Dong J, Liu C and Lin Z. Charging infrastructure planning
believed that the methodology followed here can provide for promoting battery electric vehicles: an activity-based
approach using multiday travel data. Transp Res C 2014;
government with good insight into charging facilities and
38(1): 44–55.
the development of EVs. 14. Li D, Cheung D, Shi X, et al. Uncertainty reasoning based on
cloud models in controllers. Comput Math Appl 1998; 35(3):
Acknowledgments 99–123.
15. Li D, Liu C and Gan W. A new cognitive model: cloud
We would like to express our sincere thanks to the reviewers for
model. Int J Intell Syst 2009; 24(3): 357–375.
their valuable evaluation and useful suggestions.
16. Yan F and Xu K. Application of a cloud model-set pair anal-
ysis in hazard assessment for biomass gasification stations.
Funding PloS One 2017; 12(1): e0170012.
17. Yu B, Li T, Kong L, et al. Passenger boarding choice predic-
This research received no specific grant from any funding agency
tion at a bus hub with real-time information. Transportmetrica
in the public, commercial, or not-for-profit sectors.
B 2015; 3(3): 192–221.
18. Li D, Liu C and Gan W. A new cognitive model: cloud
References model. Int J Intell Syst 2010; 24(3): 357–375.
1. He SY, Kuo YH and Wu D. Incorporating institutional and 19. Yang Z and Li D. Planar model and its application in predic-
spatial factors in the selection of the optimal locations of pub- tion. Chin J Comput 1998; 21(11): 961–969.
lic electric vehicle charging facilities: a case study of Beijing, 20. Tavakkolimoghaddam R, Vazifehnoshafagh S, Taleizadeh
China. Transp Res C 2016; 67: 131–148. AA, et al. Pricing and location decisions in multi-objective
2. Frade I, Ribeiro A, Gonc xalves G, et al. Optimal location of facility location problem with queuing systems. Eng Optimiz
charging stations for electric vehicles in a neighborhood in 2016; 49(1): 136–160.
Lisbon, Portugal. TRR Journal 2011; 2252: 91–98. 21. Yao BZ, Hu P, Lu XH, et al. Transit network design based
3. Xi X, Sioshansi R and Marano V. Simulation–optimization on travel time reliability. Transp Res C 2014; 43: 233–248.
model for location of a public electric vehicle charging infra- 22. Lee S and Kang T. Analysis of constrained optimization
structure. Transp Res D 2013; 22(4): 60–69. problems by the SCE-UA with an adaptive penalty function.
4. Cavadas J, Correia GHDA and Gouveia J. A MIP model for J Comput Civil Eng 2015; 30(3): 04015035.
locating slow-charging stations for electric vehicles in urban 23. Jeon JH, Park CG and Engel BA. Comparison of perfor-
mance between genetic algorithm and SCE-UA for calibra-
areas accounting for driver tours. Transp Res E, 2015; 75:
tion of SCS-CN surface runoff simulation. Water 2014;
188–201.
5. Hiwatari R, Ikeya T and Okano K. A road traffic simulator to 6(11): 3433–3456.
24. Kuczera G. Efficient subspace probabilistic parameter opti-
analyze layout and effectiveness of rapid charging infrastruc-
mization for catchment models. Water Resour Res 1997;
ture for electric vehicle. In: Vehicle Power and Propulsion
33(1): 177–185.
Conference (VPPC), Chicago, IL, USA, September 6–9, 2011,
pp. 1–6. New York: IEEE.
6. Capar I, Kuby M, Leon VJ, et al. An arc cover–path–cover Author biographies
formulation and strategic analysis of alternative fuel station
locations. Eur J Oper Res 2013; 227(1): 142–151.
Zhihui Tian, is a PHD’s degree candidate at Dalian
7. Cruz-Zambrano M, Corchero C, Igualada-Gonzalez L, et al. University of Technology, Dalian, China.
Optimal location of fast charging stations in Barcelona: a
flow-capturing approach. In: Proceeding of 2013 10th Wenbin Hou, a professor of State Key Laboratory of
International Conference on the European Energy Market Structural Analysis for Industrial Equipment, School of
(EEM), Stockholm, Sweden, May27–31, 2013, pp. 1–6. New Automotive Engineering Dalian University of
York: IEEE. Technology, Dalian, China.
636 Simulation: Transactions of the Society for Modeling and Simulation International 94(7)

Xiaoning Gu, is a master’s degree candidate at Dalian Dalian University of Technology, Dalian, China. She is
University of Technology, Dalian, China. also the Dean of Institute for Automotive Service
Engineering. Her current research interests include public
Feng Gu, is a PHD’s degree candidate at Dalian transportation, Vehicle Automation, swarm intelligence
University of Technology, Dalian, China. and vehicle routing problem.

Baozhen Yao, received her PhD degree from Beijing

Jiaotong University, Beijing, China in 2011. Currently she
is a professor in school of Automotive Engineering,