Kendaraan Listrik 2

JID: TRB
ARTICLE IN PRESS [m3Gsc;March 16, 2017;19:29]
Transportation Research Part B 0 0 0 (2017) 1–19
Contents lists available at ScienceDirect
Transportation Research Part B

journal homepage: www.elsevier.com/locate/trb
Joint charging mode and location choice model for battery

electric vehicle users
Min Xu a, Qiang Meng a,∗, Kai Liu b, Toshiyuki Yamamoto c
a
Department of Civil and Environmental Engineering, National University of Singapore, Singapore 117576, Singapore
b
School of Transportation and Logistics, Dalian University of Technology, PR China
c
Institute of Materials and Systems for Sustainability, Nagoya University, Japan 464-8603
a r t i c l e i n f o a b s t r a c t
Article history: This paper aims to investigate the choice for charging mode and location with the revealed
Received 12 June 2016 preference data of battery electric vehicle (BEV) users in Japan. Three alternatives including
Revised 3 March 2017
the normal charging at home (for private BEVs)/company premise (for commercial BEVs),
Accepted 4 March 2017
normal charging at public charging stations and fast charging at public charging stations
Available online xxx
are defined. A mixed logit model is developed to investigate what and how factors influ-
Keywords: ence BEV users’ choice of charging mode (normal or fast) and location (home/company or
Battery electric vehicles public stations), by identifying an appropriate instrumental variable to correct the serious
Charging mode and location endogeneity problem caused by the midnight indicator. The parameters estimation and re-
Endogeneity sults interpretation are conducted for private and commercial BEVs respectively. They sug-
Mixed logit model gest that the battery capacity, midnight indicator, initial state of charge (SOC) and number
of past fast charging events are the main predictors for users’ choice of charging mode
and location, that the day interval between current charging and next trip positively af-
fects the normal charging at home/company. In addition, with the increasing of vehicle-
kilometres of travel (VKT)/travel duration on former/next travel day, the probability of nor-
mal charging at home/company is increased for commercial BEVs, while is decreased for
private BEVs. The findings obtained herein have provided new insights into the realization
of power peak-load shifting and operation strategy for public charging stations, as well as
inspired the development and application of new models and methodologies to determine
the density and deployment of public charging stations.
© 2017 Elsevier Ltd. All rights reserved.
1. Introduction
Battery electric vehicles (BEVs) have been touted as one of the promising ways to achieve urban sustainable development
because of their competences to fossil fuel dependency and reduce the tailpipe emissions. Over the past decades, many gov-
ernments over the world have affirmed their commitments to vehicle electrification by providing considerable investments
in charging infrastructure, consumer incentives as well as research and development (Trigg et al., 2013). BEVs can be re-
plenished at charging facilities installed at homes, working places or public charging stations. Their charging magnitudes
are generally classified into three levels: Levels 1, 2 and 3. In the literature, Levels 1 and 2 are collectively referred to as
the “normal” or “slow” charging mode with charging time ranging from 4 to 12 h to get a depleted 24 kwh battery fully
∗
Corresponding author.
E-mail address: ceemq@nus.edu.sg (Q. Meng).
http://dx.doi.org/10.1016/j.trb.2017.03.004
0191-2615/© 2017 Elsevier Ltd. All rights reserved.
Please cite this article as: M. Xu et al., Joint charging mode and location choice model for battery electric vehicle users,
Transportation Research Part B (2017), http://dx.doi.org/10.1016/j.trb.2017.03.004
JID: TRB
2 M. Xu et al. / Transportation Research Part B 000 (2017) 1–19
replenished, while Level 3 is known as the “fast” charging mode due to their great charging efficiency with charging time
ranging from 0.5 to 2 hours for a full charge (Trigg et al., 2013). Although fast charging stations provide high charging effi-
ciency, they are currently limited to public infrastructure due to high voltage requirement, and mainly for emergency usage
of BEVs during their long trips. Rather, slow chargers can be installed at home and working places, thus they provide great
convenience and flexibility to charge battery when BEVs are parked overnight at home or all days in working places.
Although BEVs deliver sizable social and environmental benefits, the large-scale adoption of BEVs will inevitably have
a significant impact on the power grid load. To ensure that BEVs and power grid can interface with each other safely
and reliably, the peak power capacity may need to be increased by upgrading the power distribution transformers. For-
tunately, the desirable effect of peak-load shifting can be achieved by increasing the off-peak electricity utilization by BEVs
(Hadley, 2006). The deep study on the BEV users’ charging behaviour, especially their choice in charging mode and location,
will enable us to understand two critical issues: (i) whether the advent of a further mass-market of BEVs is viable without
largely upgrading the current power distribution transformers and (ii) whether it is possible to encourage BEV users charge
at off-peak hours by fiscal incentives.
Apart from the added load on the power grid, how to deploy charging infrastructures intelligently in prospect of large-
scale BEVs’ uptake, especially in public locations, is also one of the most pressing challenges for any local governments.
Even excluding the expense for additional electrical or construction works, the simple procurement and installation of one
public charging station can easily cost up to USD 50 0 0 to 15,0 0 0 (Trigg et al., 2013). However, without any information
regarding charging demand and behavior, early attempts to merely maximize the coverage of charging stations result in
the low or almost no utilisation of some charging points (Hagerty and Ramsey, 2011; Russo, 2015). When the demand for
BEVs is uncertain, especially for the countries in the initial stage to demonstrate vehicle electrification, how to determine
the appropriate number of public BEV charging stations and deploy them intelligently remains the most prominent issue
going forward along the way to achieve widespread jumpstart of BEVs. Given the limited knowledge and experiences, one of
the best practices to address the aforementioned problem is to analyse the available historical data about BEV users’ travel
pattern, charging behavior and infrastructure utilization.
The recovery of BEVs has brought in the extensive studies during the last few decades. For example, Brownstone et al.
(20 0 0), Dagsvik et al. (20 02), Hackbarth & Madlener (2013) and Glerum et al. (2013) analysed household’s or individual’s
preferences for BEVs by applying different discrete choice models. Jensen et al.(2013) investigated the effect of experiencing
an electric vehicle on individual’s preferences for electric vehicle. Fetene et al. (2016) recently proposed a microeconomic
model of charging at working place and gave a rationale for the public initiatives to boost the provision of charging at
working place. Moreover, on the one hand, some studies sought for strategic or operational decision-makings to accelerate
the adoption of BEVs, including (i) how to distribute the budget to charging infrastructure, incentives, etc. to better facilitate
the uptake of BEVs (Nie et al., 2016) and (ii) how to optimally deploy charging infrastructure to sustain the mass-adoption
of BEVs (He et al., 2013; Nie and Ghamami, 2013; Sathaye and Kelley, 2013; Mak et al., 2013; He et al., 2015; Arslan and
Karaşan, 2016; Ghamami et al., 2016; Li et al., 2016a; He et al., 2016; Hof et al., 2017; Liu and Wang, 2017) as well as (iii)
how to address the routing, touring, fleet deployment or relocation problem of BEVs to facilitate their integration in urban
logistics and shared mobility (Li et al., 2016b; Liao et al., 2016; Boyacı et al., 2017). Understanding how the incentives affect
the adoption of BEVs and how to deploy the charging infrastructure calls for a tangible behavioral model that can capture
the charging pattern of BEV users. On the other hand, others investigated the impact of jumpstart of BEVs on the power
grid, environment and climate, traffic assignment, etc. (Clement-Nyns et al., 2010; Waraich et al., 2013; Jiang et al., 2012;
He et al., 2014; Xu et al., 2017). These impact analyses are closely related to the deployment of charging infrastructure and
resultant charging behavior of BEV users. In the other words, the existing studies generally assume that the deployment of
charging infrastructure and charging pattern/behavior are exogenously given.
Although the charging behavior modeling is of great importance, it has been examined by a few studies due to the lack
of real revealed preference data. Smart et al. (2013) studied various driving and charging behavior across electric vehicles
enrolled in a project conduced in 21 metropolitan areas in the USA. Sun et al. (2015a) applied a mixed logit model with
unobserved heterogeneity in an effort to investigate the timing choice for the normal charging of private and commercial
BEV users in Japan after their last trip of a day. They found that the normal charging behavior was driven by electricity
demand and BEV users preferred to charge at night for the sake of low electricity tariff. Jabeen et al. (2013), explored the
BEV users’ charging preference to home, work and public places with the stated choice experiments conducted in Australian.
They showed that BEV users preferred to charge their vehicles at night and they were sensitive to charging cost and duration
generally. However, their models did not differentiate the charging level of public charging stations, and the interaction of
driving and charging behaviors was not revealed due to the lack of explanatory variables concerning the driving behavior.
With data collected from a real-world fleet of instrumented plug-in hybrid electric vehicles (PHEVs) deployed in the USA,
Zoepf et al. (2013) adopted a mixed logit model to study whether PHEV users charge at the end of each trip. They found
that PHEV users generally preferred to charge their PHEVs after a day’s last trip ending at home with the next trip occurring
in more than 3 hours later. Daina (2013) used the multinomial logit model to investigate the choice of charging timing and
trip adaptation with data from the stated preference survey conducted in UK for the non-electric vehicle drivers.
To the best of our knowledge, no research has been conducted so far to investigate the joint choice of charging mode
(normal or fast) and location (home/company or public) with revealed preference data. By considering the charging modes
and locations jointly and allowing the correlation between alternatives, the specified choice situation is more realistic, re-
sulting in more credible estimation results. More importantly, the distinct behavior of BEV users towards normal charging
JID: TRB
M. Xu et al. / Transportation Research Part B 000 (2017) 1–19 3
Fig. 1. Non-residential charging stations by slow/fast in 2012 (Trigg et al., 2013).
at home or work place and public charging stations, as well as the difference between normal and fast charging at public
charging points can be therefore revealed explicitly. It should be pointed out that defining the choice set as “charging at
home” and “charging at public stations” or “normal charging” and “fast charging” cannot capture those differences. More-
over, all the aforementioned studies on the charging behavior modelling of BEV users have overlooked the endogeneity issue
although some of them may suffer from such problem.
In light of the above considerations, this study will focus on the model development for joint charging mode and location
choice of BEV users by identifying an appropriate instrumental variable to correct the endogeneity. The revealed preference
data collected over an entire year from nearly 500 BEV users in Japan will be used. Based on these real data, a mixed logit
model will be built to describe the charging behavior of BEV users. Specifically, what and how factors influence their choice
of charging mode (normal or fast) and location (home/company or public) are investigated jointly. The endogeneity issue
caused by the midnight indicator is resolved by the control function approach with an appropriate instrumental variable.
The final results provide important clues for the realization of peak-load shifting and operational strategy for public charging
stations, as well as the determination of density and deployment of public charging stations.
In all, the contributions of this study are fourfold. First, this study makes the first attempt to investigate the joint choice
of charging mode (normal or fast) and location (home/company or public) using the real revealed preference data, which is
more reasonable than isolating the choice for charging mode and location. Second, this study proposes a tangible procedure
to pre-process the probe BEV data to facilitate model building for the charging behavior of BEV users. The procedure can be
applied to any newly-collected probe BEV data with ease. Third, this study develops a mixed logit model for the charging
mode and location choice behavior of BEV users by identifying an appropriate instrumental variable for the endogenous
variable to solve the endogeneity issue. Last, the developed models provide useful insights in the charging behavior of BEV
users as well as ideas for policy-makers to boost BEV market.
The remainder of this study is organized as follows. First, the data profile and probe data pre-process procedure are first
introduced in Section 2, followed by a discussion of the observed BEV driving and charging behavior at the vehicle level.
Section 3 elaborates the research methodology for the data analysis including the choice set construction, explanatory vari-
ables and instrument identification as well as the specification for mixed logit model. Results interpretations and discussions
are presented in Section 4. Lastly, Section 5 wraps this study up with conclusions and future research.
2. Probe BEV data pre-processing procedure and observed BEV driving & charging behavior in Japan
As shown in Fig. 1, Japan has held the largest stock of non-residential fast charging stations by 2012, and Japanese
government aims to deploy 2 million slow chargers and 50 0 0 fast charging points by 2020. Driven by the fast growing
number of charging stations, Japan has become the spot with largest market share of BEVs in world by 2013, followed
by the United States and China (Trigg et al., 2013). In view of the leading role of Japan to adopt BEVs, it should be more
representative to model the charging behavior of BEV users by using the revealed preference data of Japan.
The data used by this study is fortunately provided by the project dedicated to the consign technology development for
rational use of energy by the Ministry of Economy, Trade and Industry (METI) in Japan. Two-years probe data from nearly
JID: TRB
500 private and commercial BEVs, scattered throughout Japan (almost all the prefectures), were collected from February
2011 to January 2013. The number of private and commercial probe BEVs is about 250. There are two types of BEVs in
terms of their driving ranges, roughly 120 km and 180 km respectively. The information of each BEV was recorded at each
minute when it was in the state of moving or charging, including vehicle ID, GPS location, odograph reading, state of charge
(SOC), charging mode (normal or fast). Additionally, the prefecture where BEVs were registered and the type of business
for each commercial vehicle (e.g., whether it was from the electricity company) were also revealed in the dataset. For more
detailed elaboration of this project and data, interested readers may refer to the studies by Sun et al. (2015a,b).
2.1. Probe BEV data pre-processing procedure
Table 1 shows the raw data of a particular BEV (GPS locations are not presented in the table because of privacy). It can
be seen that detailed information including the ODO (km), SOC (%), GPS location and their correspondent clock time for
each travel or charging event is recorded every minute. We can easily figure out that this BEV had first made two trips (in
state of “moving” ) on the 1st March 2012, and was then normal-charged to SOC of 100% at midnight. In the following day,
another two trips were made in the morning and afternoon respectively, followed by a fast charging activity taking about
25 min.
Since this study focuses on the charging behavior modelling of BEV users, the raw probe BEV data are first processed
to obtain the event-driven data of each BEV comprising three events with start and end times: travelling, normal charging
and fast charging. For example, Table 2 shows the raw event-driven data generated directly from the raw probe BEV data
presented in Table 1.
The raw event-driven data generated from probe BEV data are not immune to the errors involved at the collecting and
receiving end. For instance, 0.3% of the raw event-driven data of the Japan project are the traveling events starting with SOC
less than −5% or ending with SOC larger than 105%, which are obviously impossible. There are nearly 30% of the raw event-
driven data for the traveling events with trip distance being zero (measurement accuracy for distance is a kilometre). 6.6%
of the raw event-driven data are for the charging events with charging duration less than two minutes or SOC difference
less than 1% between starting and ending of event. It is thus necessary to eliminate these error-prone event-driven data.
The probe BEV data pre-processing procedure is presented in Fig. 2 as follows.
Table 1
Raw data of a particular BEV.
Basic information State ODO SOC GPS location Starting time Current time
∗∗
moving 3727 47 2012/3/1 08:13 2012/3/1 08:14
Vehicle ID, ∗∗
moving 3728 46 2012/3/1 08:13 2012/3/1 08:15
moving … … … … …
∗∗
Type (private/commercial), moving 3748 33 2012/3/1 08:13 2012/3/1 09:18
∗∗
moving 3748 32.5 2012/3/1 08:13 2012/3/1 09:19
∗∗
moving 3749 32.5 2012/3/1 16:37 2012/3/1 16:38
∗∗
Type of vehicles in terms moving 3749 32 2012/3/1 16:37 2012/3/1 16:39
of driving range moving … … … … …
∗∗
(120 km/180 km), moving 3769 28 2012/3/1 16:37 2012/3/1 17:38
∗∗
moving 3770 27.5 2012/3/1 16:37 2012/3/1 17:39
∗∗
normal charging 3770 27.5 2012/3/1 23:00 2012/3/1 23:01
Prefecture registered, ∗∗
normal charging 3770 27 2012/3/1 23:00 2012/3/1 23:02
Type of business normal charging … … … … …
∗∗
(electricity company or not, normal charging 3770 99.5 2012/3/1 23:00 2012/3/2 03:53
∗∗
only for commercial BEVs) normal charging 3770 100 2012/3/1 23:00 2012/3/2 03:54
∗∗
moving 3770 100 2012/3/2 08:23 2012/3/2 08:24
∗∗
moving 3770 99 2012/3/2 08:23 2012/3/2 08:25
moving … … … … …
∗∗
moving 3793 81.5 2012/3/2 08:23 2012/3/2 09:30
∗∗
moving 3794 81 2012/3/2 08:23 2012/3/2 09:30
∗∗
moving 3795 81 2012/3/2 14:44 2012/3/2 14:45
∗∗
moving 3796 80 2012/3/2 14:44 2012/3/2 14:46
moving … … … … …
∗∗
moving 3813 66 2012/3/2 14:44 2012/3/2 16:42
∗∗
moving 3813 65 2012/3/2 14:44 2012/3/2 16:23
∗∗
fast charging 3813 65 2012/3/2 16:23 2012/3/2 16:24
∗∗
fast charging 3813 68 2012/3/2 16:23 2012/3/2 16:25
fast charging … … … … …
∗∗
fast charging 3813 96.5 2012/3/2 16:23 2012/3/2 16:46
∗∗
fast charging 3813 97 2012/3/2 16:23 2012/3/2 16:47
JID: TRB
Table 2
Event-driven data.
Event Starting of Event Ending of Event
ODO SOC GPS Date Time ODO SOC GPS Date Time
∗∗ ∗∗
Traveling 3727 47 2012/3/1 08:13 3748 32.5 2012/3/1 09:19
∗∗ ∗∗
Traveling 3749 42.5 2012/3/1 16:37 3770 27.5 2012/3/1 17:39
∗∗ ∗∗
Normal charging 3770 27.5 2012/3/1 23:00 3770 100 2012/3/2 03:54
∗∗ ∗∗
Traveling 3770 100 2012/3/2 08:23 3794 81 2012/3/2 09:30
∗∗ ∗∗
Traveling 3795 81 2012/3/2 14:44 3813 65 2012/3/2 16:23
∗∗ ∗∗
Fast charging 3813 65 2012/3/2 16:23 3813 97 2012/3/2 16:47
Raw Probe Data
Aggregate Event
Yes Moving?
No
Legend
Raw Travelling Raw Charging Database

Data Elimination
Event Data Event Data
Process
Ini_SOC<0% OR Diff_SOC<1% OR
Yes Yes
End_SOC>100% OR Diff_GPS>50m OR
Diff_ODO=0km? Diff_DateTime<2min?
Decision
No No
Clean Event-driven Data
Fig. 2. Probe BEV data pre-processing procedure.
Table 3
Number of charging events for private and commercial
BEVs.
BEV type First year Second year Total
Private 3,909 34,892 38,801

Commercial 3,882 42,557 46,439
Total 7,791 77,449 85,240
After applying the probe BEV data pre-processing procedure for the data of the Japan, we have totally 85,240 charging
events shown in Table 3.
It should be pointed out that most of the charging events were from the second year of the project. This is because the
record for the private BEVs started at the late of the first year, in October 2011. In addition, only a small fraction of BEVs
participated in the project at the very beginning, and most of them joined in the second year of the project. For instance,
only 16 probe BEVs were recorded in the first month, then it grew to be about 250 at the middle stage of project and nearly
500 probe BEVs were observed when the project came to an end. The distinct data size between the first and second years
of the project can be readily seen from Table 3. Besides, BEV users might take time to adapt to the limited driving range and
long charging time of BEVs before their driving and charging behavior became stable. In light of the preceding issues, only
the data from the second year of project are used for this study. Moreover, the socio-economic and demographic information
of these BEV users as well as the details of charging stations, e.g., the charging rate for members or non-members and the
JID: TRB
Table 4
The descriptive statistics of metrics for driving and charging behavior at the vehicle level.
Metrics Private Vehicles (n= 234) Commercial Vehicles (n= 249)
Min. Median Mean Max. Min. Median Mean Max.
Driving Behavior Total Distance (km) 247 5842 7105 31,995 105 4793 5387 22,791
Total Driving Day (days) 45 288 271.80 362 22 196 195.50 359
Aver Distance Per Day 3.25 21.60 25.03 91.94 4.57 25.94 27.02 110.79
Driven (km/day)
Aver Distance Between 12.44 42.40 44.14 109.72 8.58 25.61 27.94 85.00
Charge (km)
Charging Behavior Normal_Home/ 0 86.50 117.48 553 0 145 150.9 547
Normal_Company Num.
Fast_Public Num. 0 1 18.42 317 0 0 5.763 304
Normal_Public Num. 0 1 11.94 252 0 1 14.24 314
Total Charge Num. 5 114 147.80 630 6 163 170.90 709
Aver Charge Num. Per Day 0.05 0.43 0.52 1.81 0.15 0.90 0.86 2.45
Driven
Aver Nor- 0.00 0.35 0.41 1.58 0.00 0.82 0.77 1.86
mal_Home/Normal_Company
Num. Per Day Driven
Aver Fast_Public Num. Per 0.00 0.01 0.06 1.02 0.00 0.00 0.02 1.05
Day Driven
Aver Normal_Public Num. 0.00 0.00 0.05 0.80 0.00 0.01 0.07 1.17
Per Day Driven
Aver Initial SOC (%) 16.03 48.52 48.97 82.23 24.52 67.80 65.20 89.95
number of electrical outlets provided in each stations, are not available. Therefore, the explanatory variables related to the
aforementioned information will not be incorporated by this study.
2.2. Observed driving & charging behavior in Japan
To have a better understanding of BEV users’ behavior, 12 metrics regarding the driving and charging behavior at the
vehicle level and their descriptive statistics for the even-driven data of the Japan project are shown in Table 4. The histogram
and ECDF (empirical cumulative distribution function) of four metrics, including the average travel distance per day driven,
the average distance between two consecutive charging events, the average charging number per day driven as well as the
average initial SOC, are depicted in Fig. 3.
Note that among the 12 metrics, the average travel distance per day driven or between consecutive charging events
for each BEV are frequently reported by the existing studies on travel behavior (Pearre et al., 2011; Smart et al., 2013).
The former is an indicator for travel demand, while the latter can reflect the charging demand to some extent. It can be
seen from Table 4 that private BEVs in Japan are driven 25.03 km per day driven on average with a median value of 21.60,
while commercial BEVs have a slightly higher travel distance with the mean and median being 27.02 and 25.94 km per day
driven respectively. This means that commercial BEV users have a higher demand to travel by their BEVs than private BEVs
averagely and the underlying distributions are both left-skewed rather than normally distributed according to Fig. 3(a). The
ECDF and histogram show that 85%/91% of the private/commercial BEVs average 40 km or less per day driven, with only
three vehicles averaging more than 75 km/day. The vehicle with the largest value of average distance per day driven is a
commercial BEV which is driven 110.79 km/day.
As for the average travel distance between consecutive charging events, it shows that private BEVs have a higher mean
value than commercial ones, indicating that private BEVs are generally driven a longer distance before charging and thus
requiring less chances to charge their BEVs. This speculation can also be made from the ECDF in Fig. 3(b) which shows that
only 45% of the private BEVs are charged after driven 40 km or less on average, while the percentage of commercial BEVs
reaches up to 88%. Additionally, it is apparent from the histogram that there exists considerable variation in the average
distance between charging from vehicle to vehicle and the variation of private BEVs is wider than that of commercial ones.
Moreover, although the commercial BEVs have a higher travel demand than private one on average, the relatively smaller
number of charging events per private vehicle can be readily seen from the total charging number in Table 4. The dis-
tribution of average charging number per day driven of private BEVs is apparently left-skewed while the average charging
number for commercial BEVs is approximately normally distributed. The ECDF in Fig. 3(c) shows that only 8% of private BEVs
are charged more than once per day driven, whereas the percentage of commercial BEVs is 30%. This reveals the more seri-
ously risk-adverse nature of commercial BEVs which are charged more frequently than necessarily. Besides, Fig. 3(d) shows
that the distribution of average initial SOC before charging for commercial BEVs is right-skewed and the ECDF indicates that
56% of private BEVs would not be charged unless their SOC falls below 50% averagely, while more than 85% of commercial
BEVs perform charging even if the SOC are above 50%. The higher mean value of average initial SOC of commercial BEVs
before charging further reinforce our speculation. Furthermore, we can see from Table 4 that the minimum and the median
values for the number of charging events at public stations per vehicle are either 0 or 1. It means that more than half of the
JID: TRB
Fig. 3. Distribution of four metrics at the vehicle level.
drivers of BEV have never tried or only tried once the normal charging or fast charging at public charging stations during
the study period.
3. Mixed logit model building
In order to capture the joint choice behavior of BEV users for the charging mode and location, we define three alterna-
tives in the charging choice set: (i) normal charging at home (for private BEVs)/company premise (for commercial BEVs),
(ii) normal charging at public charging stations and (iii) fast charging at public charging stations. Fig. 4 depicts the total
Fig. 4. Number of times each charging alternatives being chosen.
JID: TRB
number of normal charging at home/company, at public stations and fast charging at public points in the second year of the
japan project for the private and commercial probe BEVs, respectively. It shows that for both private and commercial BEVs,
the normal charging at home/company evidently predominate the other two alternatives in terms of number of times being
chosen. The second choice for private BEVs is fast charging at public charging stations. As for commercial BEVs, they prefer
normal charging more than fast charging at public charging stations. Considering the different charging behavior between
private and commercial BEVs, they will be examined separately.
It is reasonable to assume BEV users follow the utility-maximizing principle in the choice of the above three alternatives.
According to the data profile presented in Section 2, the sequence of choices from each individual BEV users in different
choice situations (at different time) were observed, resulting in a panel data set. There may exist taste variation among
the sampled BEV users in an unobserved manner. For instance, the risk-adverse BEV users are more likely to charge their
BEVs at home in case of getting stagnant in the middle of trip with no charging stations in the vicinity. Some habitual BEV
users prefer to stay with the choices they made previously because they have formed a habit to do so. While the variety-
seeking BEV users may prefer to try different types of charging at various locations. Moreover, there may exist correlation
between some alternatives if they share common unobserved attributes. For example, in some BEV users’ opinion, normal
charging at home is more similar to normal charging at public stations, rather than fast charging. Also, it is possible that
normal charging at public places is thought to be more similar to fast charging at public stations, since both of them do
not have the convenience of charging at home. The mixed logit model, as one of the most promising discrete choice models
(DCMs), is ideal for panel dataset to accommodate the correlation over alternatives as well as the unobserved preference
heterogeneity among sampled BEV users. This approach has been used in many other transportations related topics such as
airport, airline and parking type choice analysis, etc. (Hess and Polak, 2004; 2005; Hess et al., 2007).
3.1. Explanatory variables identification
The independent variables to be incorporated in the mixed logit model are specified based on logical expectation and
some existing studies. To the best of our knowledge, the ten variables which may have implications on the choice of charging
mode and location are identified as follows:
Variable 1: An indicator for longer driving range. Fully charged driving range, or battery capacity is expected to have close
relationship with the fast charging behavior. It is expected that with large-capacity batteries, BEV users would
have more freedom to choose the home charging rather than be forced to charging at public charging stations.
Variable 2: An indicator for midnight from 23:00 to 7:00. Previous studies have demonstrated that the probability of home
charging at night is relatively high (Jabeen et al., 2013; Sun et al., 2015a; Zoepf et al., 2013). In this study, the
midnight indicator serves as a dummy variable for “time period with lower electricity rate” because a low-price
electricity tariff begins at 23:00 in Japan. This parameter, however, suffers from endogeneity issue which will
be discussed at length in next subsection.
Variable 3: A working day indicator. The typical travel day of BEVs is supposed to be different on weekday and weekend.
Variable 4: An indicator for BEVs registered and driven in Tokyo and Kanagawa, abbreviated as T/K indicator. As illustrated
by Sun et al. (2015b), Tokyo and Kanagawa have the highest density of charging stations, more than 55 stations
per 10 0 0 km2 , while the rest are less than 30 stations per 1,0 0 0 km2 . The density is essentially a good indicator
for the accessibility of public charging stations, which is expected to have influence on charging behavior. How-
ever, caution should be taken when interpreting the results as other characteristics of Tokyo/Kanagawa such as
the population, area, development, etc. may also affect the charging behaviors of BEV users.
Variable 5: An indicator for BEVs from Electricity Company (only applicable for commercial BEVs). BEVs operated by Elec-
tricity Company may have different charging behavior considering that they are generally more familiar with
the electricity tariffs and deployment of public charging stations.
Variable 6: Initial SOC before charging, measured in units of percentage (%). As suggested by existing studies (Sun et al.,
2015a; Zoepf et al., 2013), charging BEVs is to some extent a kind of demand-based behavior, meaning that
the lower the SOC, the higher the possibility of charging. Additionally, as an acknowledged proxy variable for
“driving range anxiety” (Nilsson, 2011), effect of the initial SOC should be great on the charging behavior.
Variable 7: The vehicle-kilometres of travel (VKT) on former travel day. Here the former travel day is defined as the day
when the former trip before charging is initiated. Some drivers may charge their BEVs because of the high
consumption of electricity on the former travel day rather than their preparation for future use.
Variable 8: The total travel duration on next travel day, measured in units of days. Here the next travel day is defined
as the day when the next trip after charging is initiated. Since drivers may have a more clear sense of future
travel duration, travel duration rather than travel distance is incorporated in the model.1
Variable 9: The day interval between the start of next trip and the start of current charging event, measured in units of
days. Drivers may estimate the total available time for charging when they initiate charging for their BEVs. If
time permits, they will perform normal charging; otherwise fast charging should be their best choice.
1
Empirical analysis has been conducted and it has been figured out that the travel duration does provide better fit.
JID: TRB
Variable 10: The total number of fast charging events before current charging event is initiated, which serves as a good in-
dicator for BEV users’ familiarity with charging stations/technology. Also, drivers may stay with the choices
they previously made because there may exist inertia or they have formed a habit to do so. This ap-
proach was first employed by Adamowicz (1994) to capture the dynamic aspects of choice behavior for
recreation.
The descriptive statistics of these selected explanatory variables at the event level in terms of the three alternatives
are summarized in Table 5. It shows that for private BEVs, about 30% of all the charging events are performed at mid-
night and the percent of charging at midnight differs significantly across the three alternatives. Specifically, nearly 36% of
the normal charging events at home are performed at midnight, while less than 5% of the fast charging events at public
charging stations are performed at that time. Additionally, over half of fast charging events happen in Tokyo or Kanagawa,
whereas the percentage of normal charging at the two prefectures is less than 20%. This is consistent with our expec-
tation that the fast charging is more likely to be performed by BEV users in Tokyo and Kanagawa given their relatively
higher density of fast charging stations. Moreover, the average initial SOC before charging is approximately 52% and the fast
charging has a slightly smaller initial SOC compared with normal charging on average. Furthermore, the average number of
fast charging events before the current fast charging is initiated is nearly 80, which is almost four times of that for nor-
mal charging at public charging stations and ten times for normal charging at home. The overwhelming domination of the
fast charging over the other two alternatives in term of the average number of the past fast charging events implies that
the more fast charging events the BEV drivers have performed, the more likely they will perform the fast charging next
time.
For commercial BEVs, it can be seen that about 75% of the charging events are performed by BEVs with a longer driving
range (i.e., 180 km) and the percentage for normal charging at company is nearly four times of that for fast charging at
public charging stations. This is quite different from private BEVs in a sense that only 25% of the charging events are per-
formed by private BEVs with a longer driving range and the percentages for fast charging and normal charging at home are
comparable. The high proportion of charging events by commercial BEVs with longer driving range should be attributed to
the large number of commercial BEVs equipped with a large-capacity battery (about 79%). Analogously, over 55% of charg-
ing events are performed by BEVs of Electricity Company and this is due to the high proportion of BEVs from Electricity
Company (about 55%). In addition, less than 10% of charging events of commercial BEVs happened at midnight or weekend,
while the percentage for private BEVs is nearly 30%. The average initial SOC of commercial BEVs before their charging is
about 67%, exceeding that of private BEVs (52%). It indicates that commercial BEVs are generally more risk-adverse than
private ones. This implication is further verified by the longer buffer time between the current charging and next trip of
commercial BEVs. Moreover, compared with private BEVs, the stronger domination of the fast charging over the other two
alternatives in term of the average number of the past fast charging events for commercial BEVs implies the more con-
centrated pattern of fast charging at the vehicle level. In other words, almost all the fast charging events are performed
by a fixed and small group of commercial BEVs. Those who have ever performed fast charging are more likely to visit
fast charging stations constantly, while those who only performed fast charging a few times or have never tried it at all
may never perform fast charging in the future. In fact, all the fast charging events are performed by no more than 30% of
commercial BEVs while the percentage of private BEVs who have never tried fast charging during the study period is less
than 40%.
3.2. Endogenous variable and its instrument
Although several empirical studies have stressed the effect of midnight indicator on the charging behavior of BEV users,
few have taken into account the endogeneity of it. Among the 10 identified explanatory variables, the midnight indicator
may suffer from serious endogeneity issue due to the existence of reciprocal causal relationship. More specifically, it seems
that the probability of charging at home would be higher if it is midnight as revealed by existing studies (Jabeen et al.,
2013; Sun et al., 2015a; Zoepf et al., 2013). In turn, if the BEV drivers have chosen to perform the home charging, then they
are more likely to charge their BEVs at midnight. The reciprocal causal relationship of private BEVs seems more apparent
than commercial ones.
Initially, the identification of midnight indicator aims to gauge the influence of relatively lower electricity rate on the
choice of charging mode and location considering that a low-price electricity tariff begins at 23:00 in Japan. Because the
dataset have no information on the explicit electricity rate or charging cost for each charging event, we hope that the
midnight indicator, as a dummy variable for “time period with lower electricity rate”, can somehow serve as a remedy
for the unavailable data on the electricity rate. However, the midnight indicator can reflect some unobserved attributes of
charging, e.g., the comfort and convenience of charging at home before falling asleep for private BEVs. The mix of above
two effects may be the main induction for the endogeneity issue, which often occurs when the explanatory variables are
correlated with the unobserved factors. Therefore, on the surface, this variable is an indicator for time clock, however, it
is a poor proxy variable for lower electricity rate since it is correlated with some unobserved attributes and results in
endogeneity.
10
JID: TRB
Table 5
The descriptive statistics of explanatory variables.
Variable Private Vehicles (n= 234) Commercial Vehicles (n= 249)
Normal_Home Normal_Public Fast_Public Total Normal_Company Normal_Public Fast_Public Total
M. Xu et al. / Transportation Research Part B 000 (2017) 1–19

(n= 27,725) (n= 2819) (n= 4348) (n= 34,892) (n= 37,575) (n= 3547) (n= 1435) (n= 42,557)
ARTICLE IN PRESS
Long driving range 25.78 38.74 16.93 25.73 78.44 51.31 21.11 74.25
(%)
Midnight (%) 35.77 18.38 4.69 30.49 6.79 7.73 10.17 6.98
Working day (%) 72.80 69.49 69.50 72.12 92.74 84.92 79.51 91.64
Dense stations (%) 14.00 17.88 50.99 18.92 17.02 38.85 53.03 20.05
Electricity company N.A. N.A. N.A. N.A. 60.76 21.31 10.80 55.79
(%)
Initial SOC (%) 52.777 53.990 46.447 52.086 68.701 60.446 51.189 67.422
(19.761) (21.711) (18.002) (19.833) (18.936) (20.308) (19.549) (19.448)
VKT on former 25.318 23.827 31.459 25.963 20.232 18.093 27.493 20.299
travel day (km)
(19.086) (19.892) (23.652) (19.889) (15.883) (14.423) (21.874) (16.073)
Duration on next 0.043 0.040 0.036 0.042 0.042 0.038 0.042 0.042
travel day (days)
(0.032) (0.030) (0.027) (0.032) (0.032) (0.034) (0.034) (0.033)
Day interval_Next 0.874 0.691 0.068 0.759 1.392 1.032 0.052 1.317
trip (days)
(1.082) (1.656) (0.772) (1.139) (2.148) (3.191) (0.455) (2.235)
Number of fast 8.647 20.872 79.820 18.504 3.177 3.650 111.910 6.883
charging events
(23.363) (34.317) (77.870) (42.785) (19.906) (19.835) (89.350) (32.197)
Value in brackets are the standard deviation (SD) of parameters.

“N.A.” denotes variables not considered in the model.
[m3Gsc;March 16, 2017;19:29]

JID: TRB
The standard way to deal with the endogenous variables is to find at least one suitable instrumental variable for mid-
night indicator to facilitate the application of some specific approaches, e.g., BLP2 (Berry et al., 1995), control function ap-
proach and full maximum likelihood approach (Petrin & Train, 2010; Train, 2009), to correct the endogeneity in the DCM.
Nevertheless, caution should be taken in the selection of an appropriate set of instruments for applying these approaches
especially considering the facts that (i) it is controversial issue itself3 ; (ii) the use of those approaches will possibly lead
to a weak instruments problem4 ; (iii) it is impossible to theoretically evaluate or verify the effectiveness of the specified
instruments.
Specially, the instruments in our context should meet two requirements: (i) be correlated with the midnight indicator;
ii) be not correlated with the unobserved attributes, i.e., should not have direct influence on the choices being made. The
first requirement can be easily verified by the correlation test. However, the second requirement is somehow vague and
it can only be maintained by economic behavior or introspection. It is a pity that for the choice of instrumental variables
in our context, we have no many literature to refer to because i) most studies focus on the endogeneity of the price for
alternative products in the market and the available methods to handle endogeneity in DCM are all fairly case-specific; (ii)
all the existing studies on the charging behavior modelling of BEV users ignore the endogeneity issue although some of
them suffer from the same problem, i.e., include the time indicators (morning, lunchtime, midnight, etc.) as an explanatory
variable (Jabeen et al., 2013; Sun et al., 2015a; Zoepf et al., 2013). From our understanding, the day interval between the
end of former trip and the start of charging event (measured in units of days), referred to as “Day interval_Former trip” in
this study, may serve as a potential instrumental variable for “Midnight indicator” . On the one hand, the two variables are
linearly correlated with each other for both private and commercial BEVs.5 On the other hand, the day interval associated
with the former trip may not directly affect the choice of charging mode and location because a driver seldom make decision
by recalling or according to the ending time of last trip when he/she is thinking about where and how to charge their
BEVs. Given the identified instrument, the control function approach will be used to correct the bias resulted from the
endogeneity. Details can be found in Section 3.3.
3.3. Mixed logit model specification
Let J denote the choice set we defined in Section 2, which includes (i) normal charging at home (for private
BEVs)/company premise (for commercial BEVs), (ii) normal charging at public charging stations and (iii) fast charging at
public charging stations. With the pre-defined explanatory variables, the utility that individual n obtains from choosing
alternative j ∈ J in choice situation t is specified by

Un jt = V xn jt , xñ jt , βn j + εn jt (1)
where xnjt is a vector of exogenous explanatory variables specified above for alternative j faced by individual n in choice
situation t, and xñ jt is the endogenous explanatory variable (i.e., midnight indicator) for alternative j faced by individual n
in choice situation t. Both xnjt and xñ jt represent the observed attributes of alternatives which varies across individuals and
choice situations. βnj denotes a vector of coefficients of these explanatory variables for individual n representing his/her
taste. ε njt represents the unobserved error term.
Note that the endogeneity arises due to the correlation between xñ jt and ε njt . It is assumed in the control function
correction that xñ jt can be written as a function of some exogenous explanatory variables, (at least one) observed instrument
and an error term, i.e.,

xñ jt = W zn jt , γ + μn jt (2)
where znjt denotes the vector of selected exogenous explanatory variables and observed instrument.
To apply the control function approach, the parametric form of ε njt should be specified, i.e., we posit that ε njt can be
decomposed into two parts as follows,
εn jt = εn1 jt + εn2 jt (3)
where εn1 jt is assumed to be jointly normal with μnjt for each alternative j and iid over j to account for the correlation of
xñ jt and ε njt , whereas εn2 jt follows the iid extreme value. Furthermore, a key requirement of control function approach is
that εn1 jt can be represented as a function of μnjt and a residual which is independent of μnjt (thus not correlated with xñ jt ),
namely,

εn1 jt = CF μn jt + εñ1 jt (4)
2
Initials of the authors’ name, i.e., Berry et al. (1995).
3
A meaningfully observable or measurable instrumental variable is not always available. The instrumental variables identified in many applications are
controversial.
4
A poor instrumental variable may not be better than having no instrumental variable at all.
5
Pearson correlation between “Midnight indicator” and “Day interval_Former trip” is 0.138 (p=0.0 0 0) and 0.087 (p=0.0 0 0) for private and commercial
vehicles respectively.
JID: TRB
where εñ1 jt follows iid normal with zero mean. For simplicity, we assume CF(μnjt ) is linear function with respect to μnjt .
Therefore, we have

Un jt = V xn jt , xñ jt , βn j + λ j μn jt + σ j ηn jt + εn2 jt (5)
where ηnjt follows iid standard normal.
Hence, the choice probability for consumer choosing alternative i is given by

exp V xnit , xñit , βni + λi μnit + σi ηnit
Pnit = f β n , ηn d β n d ηn (6)
j exp V xn jt , xñ jt , βn j + λ j μn jt + σ j ηn jt
To complete the mixed logit model, we need to specify the way in which the explanatory variables enter the utility
function, i.e., the explicit function form of V (xn jt , xñ jt , βn j ). A linear specification is generally used in most setting, however
it is not appropriate for some variables which may cause the utility to change nonlinearly. In this study, for instance, the
initial SOC before charging may affect the utility in a nonlinear way since the utility would decrease much faster when the
SOC approaches zero. In addition, the total number of fast charging events before current charging event, as a measurement
for the familiarity with the fast charging stations, may also suffer from the nonlinearity problem due to the potential satu-
ration effect in the experience curve. Therefore, different categories of the above two variables should be tested via dummy
variables and the most appropriate categories are chosen. As for the choice of distribution, β is generally assumed to follow
normal or lognormal in real applications, and other distributions like triangular, uniform can also be applied (Ben-Akiva and
Bolduc, 1996; Revelt and Train, 1998, 20 0 0). The appropriate choice for the distribution of β depends on the concrete prob-
lem and may need further justification through trial-and-error in empirical analysis. In this study, normal and lognormal
distribution are tested for every coefficient independently before the distribution is finalized.6
Given the model specification, the unknown coefficients thus can be estimated by maximum simulated likelihood method
with control function approach where Modified Latin Hypercube Sampling (MLHS), initially proposed by Hess et al. (2006),
is adopted due to its increasing popularity for the ease of implementation and alleviation of undesirable correlation for
high dimension sequences. For the detailed elaboration for the approaches, readers may refer to the Chapters 6, 10 and
13 of Train (2009). For the identification of proposed mixed logit model, the alternative of normal charging at public is
normalized to be zero to facilitate results interpretation in the next section.
4. Results and discussions
Simulations are performed in Matlab and run on a personal computer with Intel (R) Core (TM) Duo 3.4 GHz CPU using
500 MLHS draws for each sampled BEV users (Train, 20 06; 20 09). Logically, all the coefficients for the aforementioned
explanatory variables can take either sign, because they are individual-specific which do not vary over alternatives, and only
the relative coefficient to the reference alternative can be estimated.
The nonlinearity testing for private BEVs shows that the estimated coefficients of all the additional dummy variables
for initial SOC and number of fast charging are insignificant, indicating that both variables do not exhibit significant non-
linearity. In addition, the variations of the LL for both parameters are small, i.e., roughly no more than 20. Therefore we do
not include the additional dummy variables in the final model for private BEVs. For the commercial BEVs, the estimates for
most of the additional dummy variables of the initial SOC larger than 30% are insignificant, hence only dummy variables for
intervals [0, 10], [10, 20] and [20, 30] will be incorporated in the final model. The testing for the number of fast charging
events indicates that the inclusion of dummy variables for five intervals, i.e., [0, 50], [50, 100], [100, 150], [150, 200], [200,
250] are sufficient to capture the nonlinearity of this variable. Therefore, we include the dummy variables for all the above
intervals in the final model for commercial BEVs. In addition, considering the empirical difficulty to estimate a mixed logit
model with so many random variables7 from large-sample panel-data, we do not allow correlations among the random
coefficients. However, by employing mixed logit model specification with random coefficients as detailed in Section 3.3, the
correlation among alternatives has been captured implicitly.
As can be seen from Table 6 and 7, most of the parameters included in the model are statistically significant and their
signs are plausible. The parameters identified to be random for private vehicles (at least for one alternative) include: indica-
tor for longer driving range, midnight, T/K indicator, initial SOC, the total travel duration on next travel day, the day interval
between the start of next trip and current charging as well as the unobserved alternative-specific constant terms. While
for commercial vehicles, all the included parameters are identified to be random variable, except working day indicator, the
additional dummy variables of initial SOC, VKT on former travel day as well as the total number of fast charging events
before current charging event is initiated. Among all these random parameters, the initial SOC of private BEVs, the travel
duration on next travel day and the day interval between the start of next trip and current charging for fast charging at
6
Distribution is finalized based on two criteria: i) fix the coefficient if standard deviation (sd) is insignificant; ii) choose Lognormal if it provides at least
θ larger log-likelihood value (LL) than Normal considering the fast convergence, better estimation for the mean and sd of the latter (Hess et al., 2005). In
this study, we set θ = 100.
7
There are 11 and 17 random coefficients in the model for private and commercial BEVs respectively. Allowing correlation among these random coeffi-
cients would call for 55 and 136 more variables to be estimated, which is computationally intractable.
JID: TRB
Table 6
Estimation results for private BEVs.
Variable Alternative Estimates Without Correction Estimates With Correction
mean t-statistic sd t-statistic p< 0 mean t-statistic sd t-statistic p< 0
Long driving range Normal_Home -1.491 (−17.366) 3.135 (36.397) 0.68 -7.980 (−10.223) 2.494 (−8.477) 1.00

Fast_Public 1.208 (11.348) 5.382 (1.3398)
ARTICLE IN PRESS
Midnight Normal_Home 2.427 (18.448) 4.618 (24.242) 0.30 70.196 (16.920) 7.225 (4.640) 0.00
Fast_Public -1.886 (−11.849) -79.941 (−4.444)
Working day Normal_Home 0.038 (0.489) -0.053 (−0.108)
Fast_Public 0.167 (2.008) 0.063 (0.033)
T/K indicator Normal_Home -0.181 (−1.757) 0.352 (−4.194) 0.70 -2.308 (−3.279) 4.836 (−5.187) 0.68
Fast_Public 1.955 (18.936) -0.448 (−0.123)
Initial SOC Normal_Home 1.010 (4.006) 0.062 (47.887) 0.108 (−18.038) 0.012 (0.916)
Fast_Public -0.019 (−8.647) -0.163 (−3.551)
VKT on former travel day Normal_Home -0.002 (−1.220) -0.045 (−3.122)
Fast_Public 0.026 (12.771) 0.094 (1.854)
Duration on next travel day Normal_Home 6.521 (4.792) 15.706 (16.326) 0.34 -50.569 (−5.613) 33.179 (−4.534) 0.94
Fast_Public 10.112 (7.077) 74.285 (1.227)
Days interval_Next Trip Normal_Home 1.521 (28.468) 1.929 (31.924) 0.22 0.560 (2.611) 0.794 (2.574) 0.24
Fast_Public -2.683 (−43.481) 1.427 (−36.071) 0.97 -0.190 (−0.345) 0.200 (−0.052) 0.83
Number of fast charging Normal_Home -0.003 (−2.875) -0.064 (−10.395)
Fast_Public 0.003 (3.796) 0.060 (4.995)
Alternative-specific constant Normal_Home 1.325 (8.586) 0.515 (−16.186) 0.01 -13.549 (−10.038) 2.609 (4.739) 1.00
Fast_Public -0.113 (−0.666) 2.199 (45.696) 0.52 8.187 (2.110) 0.112 (0.008) 0.00
Residual Normal_Home -63.421 (−18.046)
Fast_Public 79.110 (4.9969)
Error term Normal_Home 12.131 (9.638)
Fast_Public 0.292 (0.011)
Observations 34,892 34,892
LL -7154.4 -846.3
Note: value in brackets are the t-statistics and value in bold and italics is significant at 5%.
[m3Gsc;March 16, 2017;19:29]

13
14
JID: TRB
Table 7
Estimation results for commercial BEVs.
Variable Alternative Estimates Without Correction Estimates With Correction
mean t-statistic sd t-statistic p< 0 mean t-statistic sd t-statistic p< 0
Long driving range Normal_Company 0.518 (6.627) 1.054 (16.784) 0.31 4.412 (44.963) 3.292 (42.722) 0.09
Fast_Public -0.593 (−2.516) 4.322 (21.494) 0.55 -5.925 (−17.283) 0.389 (−3.459) 1.00
Midnight Normal_Company -1.484 (−9.999) 2.809 (13.771) 0.70 -82.026 (−102.548) 13.963 (51.163) 1.00
Fast_Public -1.958 (−7.940) 3.703 (16.628) 0.70 -136.460 (−18.704) 11.674 (−14.545) 1.00
Working day Normal_Company 0.173 (2.060) 5.002 (46.443)
Fast_Public 0.246 (1.637) 7.225 (17.950)
T/K indicator Normal_Company -0.806 (−10.72) 1.006 (16.530) 0.79 0.010 (0.109) 0.546 (−13.025) 0.49
Fast_Public 2.804 (12.370) 1.036 (11.891) 0.00 0.586 (2.487) 2.042 (−15.266) 0.39
Electricity Company Normal_Company 0.545 (5.308) 1.010 (8.685)
Fast_Public 0.523 (1.911) 1.059 (−5.891) 0.31 -1.276 (−4.056) 4.725 (11.893) 0.61

Initial SOC Normal_Company 0.015 (7.829) 0.091 (48.364) 0.43 0.016 (6.893) 0.072 (47.187) 0.41
ARTICLE IN PRESS
Fast_Public -0.004 (−0.902) 0.054 (24.761) 0.53 -0.033 (−6.772) 0.048 (18.837) 0.75
Initial SOC [0,10] Normal_Company -0.085 (−0.923) -0.167 (−1.024)
Fast_Public -0.284 (−2.394) -0.535 (−3.137)
Fast_Public -0.100 (−4.795) -0.148 (−5.880)
Fast_Public -0.023 (−2.362) -0.036 (−3.405)
VKT on former travel day Normal_Company 0.012 (5.740) 1.960 (99.084)
Fast_Public 0.026 (7.578) 0.210 (20.538)
Duration on next travel day Normal_Company 5.770 (5.320) 23.324 (23.747) 53.463 (146.184) 6.831 (10.601)
Fast_Public 13.873 (6.206) 5.387 (4.486) 80.145 (80.846) 4.142 (2.925)
Days interval_Next Trip Normal_Company 0.541 (21.557) 0.868 (32.516) 0.27 0.064 (3.178) 0.908 (29.538) 0.47
Fast_Public -14.212 (23.109) 26.475 (14.111) -1.314 (34.079) 0.527 (19.711)
Number of fast charging Normal_Company 0.005 (4.119) 0.024 (14.960)
Fast_Public 0.009 (7.303) 2.954 (53.736)
Number of fast charging [0,50] Normal_Company -0.018 (−2.913) 0.003 (0.352)
Fast_Public 0.032 (3.760) 0.023 (2.109)
Number of fast charging [50,100] Normal_Company 0.020 (4.326) 0.024 (5.143)
Fast_Public 0.031 (6.455) 0.027 (5.250)
Fast_Public 0.019 (5.352) 0.017 (4.631)
Fast_Public 0.018 (4.192) 0.015 (3.434)
Number of fast charging [200, 250] Normal_Company 0.002 (0.854) 0.006 (1.995)
Fast_Public 0.005 (1.919) 0.006 (2.092)
Alternative-specific constant Normal_Company 0.163 (1.018) 0.039 (0.960) 0.00 4.990 (25.308) 0.849 (−21.741) 0.00
[m3Gsc;March 16, 2017;19:29]

Fast_Public -2.791 (−7.696) 2.005 (−20.106) 0.92 5.538 (10.610) 5.882 (34.351) 0.17
Residual Normal_Company 88.197 (101.344)
Fast_Public 131.060 (18.014)
Error term Normal_Company 27.227 (69.503)
Fast_Public 8.338 (25.250)
Observations 42,557 42,557
LL −6634.3 −5720.3
Note: value in brackets are the t-statistics and value in bold and italics is significant at 5%.
JID: TRB
public stations of commercial BEVs are found to follow lognormal distribution. Only the percent less than 0 of the normal
distribution is reported in the Tables.
In addition, the model with correction offers significant improvements in model fit over the model without correction for
private BEVs, increasing the LL value from −7154.4 to −846.3, while the improvement for commercial BEVs from −6634.3
to −5720.3 are not that obvious. This might reflect the fact that endogeneity issue is more serious for private BEVs. More-
over, the increased magnitude of some of the significant estimates should also be attributed to the endogeneity correc-
tion. Such parameters include the indicator for longer driving range and midnight, the travel duration on next travel day
and alternative-specific constant for both the private and commercial BEVs. This result might reveal that the unobserved
attributes correlated with midnight indicator are also correlated with the aforementioned variables. The sign of all the sig-
nificant estimates are consistent with and without correction except the coefficient of the duration on next travel day for
home charging of private BEVs and the alternative-specific constant. For the corrected models, all the residuals and error
terms enter significantly except the error component for fast charging of private BEVs. A positive residual suggests that the
correspondent alternative possesses un-included desirable attributes and vice versa.
The detailed elaborations and implications of estimates are elaborated below. They should be interpreted with respect to
the reference alternative, i.e., the normal charging at public stations. Before the estimates are interpreted and discussed, it is
important to note that those results are only indicative of early BEV adopters and they may not necessarily be generalized
to the entire population.
4.1. Implication of useable battery capacity of BEVs (Variable 1)
Variable 1 (an indicator for longer driving range)

For private BEVs, the negative estimate for normal charging at home indicates that on an average, the probability of
normal charging at home is lower when BEVs have longer driving range. This might due to the fact that over 40% of the
public stations are free to users and the charging efficiency of public stations is often higher than that of home charging.8
In addition, the estimated mean after correction is more than five times of that without correction. This perhaps reflects
that the unobserved attributes correlated with midnight indicator are also correlated with longer driving range indicator.
Quite the opposite, for commercial BEVs, the effect of longer driving range on their choice for normal charging at company
and fast charging at public stations are significantly positive and negative respectively. It implies that commercial BEVs with
longer driving range are more willing to perform normal charging at company and less likely to perform fast charging at
public stations.
The above findings have far-reaching implications for the deployment of public charging stations. If BEVs with high-
storage battery come to market in the near future, the demand for public charging stations would become very uncertain
considering the increasing charging demand from private BEVs and the decreasing charging demand from commercial BEVs.
Moreover, both the battery life and available capacity may deteriorate with time and usage. The life cycle analysis for cost
and benefit of public charging stations should thus be conducted to determine the optimal density and deployment of
public charging stations in a long term, considering the uncertainty of battery development, level of battery usage and the
consequent effects on the demand for public charging stations.
4.2. Implication of charging time and day (Variables 2–3)
Variable 2 (An indicator for midnight from 23:00 to 7:00)

Positive coefficient of midnight indicator for home charging makes it clear that private BEV users, on an average, are
more likely to initiate their charging at midnight. The negative coefficient for fast charging at public stations is expected, as
it hinders with our natural sleep patterns. The result reinforces earlier findings that BEV users show a strong preference for
charging overnight at home (Jabeen et al. 2013; Sun et al. 2015a; Zoepf et al. 2013). Perhaps the relatively low electricity
tariffs at midnight serves as an effective incentive for them to delay their charging from early evening to off-peak hours of
electricity grid. Additionally, the midnight indicator may be also associated with some undesirable attributes of the alter-
natives. For example, some users may have to postpone their sleeping till midnight in order to enjoy the lower electricity
tariffs. The existence of these undesirable attributes can also be inferred from the increased magnitude of the estimates
after correction. The sd of this coefficient is highly significant, indicating that the effect of this attribute does indeed vary in
the population.
The average effect of midnight indicator on the charging behavior of commercial BEVs are not totally analogous to that of
private ones. Particularly, the coefficients of midnight indicator for normal charging at company and fast charging at public
stations are both significantly negative. This implies that commercial BEV users are less able or willing to perform normal
charging at company or fast charging at midnight. The relatively larger coefficient in magnitude for fast charging compared
with normal charging at company reveals drivers’ stronger aversion to the former. The result reveals that commercial BEV
users show strong preference for normal charging overnight at public charging stations. Perhaps they prefer to charge the
commercial vehicles overnight at home or some public stations not far from their residences to achieve time-saving for the
business affairs on the next day.
8
The normal charging at public stations can be operated at Level 1 or Level 2, while most of the normal charging at home is limited to Level 1 only.
JID: TRB
Variable 3 (An working day indicator)

For private BEVs, the coefficient of working day indicator is found insignificant after correction for both alternatives,
implying that there is not much difference in the preference for the three alternatives between working days and weekends.
However, for commercial BEV users, the significantly positive coefficients of working day indicator for normal charging at
company and fast charging after correction demonstrate that they have a clear preference for normal charging at company
and fast charging at public stations on working days. Additionally, the relatively larger coefficient for fast charging compared
with normal charging at company indicates drivers’ stronger inclination to the former.
4.3. Implication of density of public charging stations (Variable 4)
Variable 4 (T/K indicator)

The sign of estimated coefficient for T/K indicator after correction interprets that this variable has significantly neg-
ative effect on the choice for normal charging at home. The results are in line with our expectation that BEV users at
Tokyo/Kanagawa averagely favour charging at public stations partly due to their convenient access to the public charging
stations. Additionally, other characteristics of the two prefectures may also play a positive role. For instance, since both pre-
fectures are well developed, the home charging may not be applicable to every resident, especially those settling in high-rise
apartment buildings. Nonetheless, there exists significant preference heterogeneity for this attribute and there are still 32%
of private BEV users who do not favour normal charging at public stations even driving at Tokyo/Kanagawa. It is possible
that these drivers do not gain easy access to public stations although the coverage of them on the whole is extensive.
The effect of T/K indicator on the charging behavior of commercial BEVs are different from that of private BEVs in a sense
that the coefficient for fast charging is found to be significantly positive for the former while insignificant for the latter. It
suggests that on an average, commercial BEV users at Tokyo/Kanagawa are willing to perform fast charging at public stations.
Moreover, the results after correction show that significant preference heterogeneity exists for T/K indicator and it serve as
a negative inducement for nearly 40% of the commercial BEV users for fast charging at public stations.
4.4. Implication of business type for commercial BEVs (Variable 5)
Variable 5 (An indicator for BEVs from Electricity Company)

The Electricity Company indicator serves as a significant positive inducement for normal charging at company. It suggests
that the probability of normal charging at company is higher when the commercial BEVs are operated by Electricity Com-
pany. This is consistent with the preceding observation that BEV users from Electricity Company are more likely to perform
normal charging at their company. The relatively lower price or free electricity exclusively for their staff/company may be
one of the reasons for this behavior. As for the fast charging at public stations, the significantly negative estimate indicates
that the Electricity Company indicator has negative effect on fast charging after correction. As expected, BEV users from
Electricity Company may be more familiar with the technology and deployment of fast charging stations. It is also possible
that operators of public charging stations may have intercompany transactions with some Electricity Company and free fast
charging service is provided for members from Electricity Company.
4.5. Implication of “range anxiety” (Variable 6)
Variable 6 (Initial SOC before charging)

The significantly positive effect of initial SOC on normal charging at home implies that charging at home may not nec-
essarily be driven by demand. The result is reasonable considering that some drivers may form a habit to charge their BEVs
after daily trips even though the remaining electricity is sufficient for the next trip. This behavior also reveals the risk-
adverse nature of private BEVs which are charged more frequently than necessarily. The significant negative effect of initial
SOC on fast charging at public stations reflects demand-driven aspect of fast charging behaviour.
When the initial SOC falls into the base interval, its effects on the charging choice of commercial BEVs are analogous
to that of private BEVs. Specially, the corrected differential effect of initial SOC on normal charging at company is positive,
whereas its effect on fast charging at public stations is found to be negative. However, unlike private BEVs, the initial SOC
exhibits apparent non-linearity in the utility function of commercial BEVs, especially for the fast charging at public stations.
As expected, with the initial SOC approaching zero, the coefficients of fast charging are increasing significantly in magnitude,
indicating the anabatic willingness of commercial BEV users to perform fast charging. In addition, the coefficient of fast
charging for interval [0, 10%] is larger in magnitude than that of private BEVs. This finding captures the more serious and
diverse risk-adverse behavior on the part of commercial BEV drivers.
4.6. Implication of travel pattern (Variables 7–9)
Variable 7 (The VKT on the former travel day)

The effect of VKT on the former travel day on charging choice of mode and location is expected to be in inverse rela-
tion to that of initial SOC. Both of them reflect the demand-driven aspect of charging behavior. It can be found that after
correction, VKT on the former travel day has negative effect on normal charging at home. For commercial BEVs, both the
JID: TRB
choices for normal charging at home and fast charging at public charging stations are positively affected by VKT on the
former travel day. High consumption of electricity on the former travel day will inevitably drive commercial BEV users to
perform fast charging in the middle of the trip or normal charging at company after the trips in a day. Part of the reasons
for these behaviors can be found in the influence analysis of initial SOC.
Variable 8 (The total travel duration on the next travel day)
The results of empirical tests reveal that travel duration does provide better fit than travel distance. Similar to VKT on the
former travel day, the coefficient of travel duration on the next travel day for home charging is corrected to be negative for
private vehicles. This result implies that the long travel duration on the next travel day significantly reduces the differential
likelihood of normal charging at home. Also, the travel duration has a positive effect on normal charging at company and
fast charging at public stations for commercial BEVs. We speculate that this finding may again be capturing the risk-adverse
behavior of drivers. Although both the VKT on the former travel day and duration on the next travel day have positive effect
on charging, reflecting the risk-adverse behavior of drivers, we caution that the specific motivations for charging in these
two cases are distinct. On one hand, for the VKT on the former travel day, drivers, without knowing future travel demand,
may prefer to charge their BEVs merely to make up for the electricity loss suffered on the former travel day. On the other
hand, duration on the next travel day serves as a good indicator for future demand. If drivers have good estimation for
future demand or set out for a well-planned journey, then they would charge BEVs to meet their need in the future.
Variable 9 (The day interval between the start of next trip and the start of current charging event)
The distance/duration of the former/next trip depicts the amount of electricity offset or charging demand, while the day
interval between the start of the next trip and the start of current charging event, instead, reflect the degree of emergence
to charge. Compared with normal charging at public stations, the positive effect of the day-interval on home charging is
statistically significant for private BEVs, and taste variation exists among the whole population. It reveals that some drivers
prefer to charge their BEVs immediately after arriving home from work to make up for the electricity loss, while others may
wait till the start of next trip to perform charging for future use. On the contrary, the effect of day-intervals on fast charging
at public stations is found to be insignificant after correction.
The effect of aforementioned day-interval on normal charging at company for the commercial BEVs is similar to that
of private BEVs, while its effect on fast charging at public stations fails to be so. To be more precise, the estimate of fast
charging for commercial BEVs is significantly negative for the whole population and preference heterogeneity exists. The
negative effect of this attribute on fast charging at public stations is mainly due to the fact, that most fast charging events
are initiated in the middle of a trip for emergency use, unless the fast charging stations are deployed readily at the origin
or destination of a trip. Note that this difference partly implies that commercial BEVs on an average show a more risk-
adverse attitude to future trips and they prefer fast charging, perhaps for time-saving, if a new trip is coming soon.
4.7. Implication of familiarity to public charging stations (Variables 10)
Variable 10 (The total number of fast charging events before current charging is initiated)
The number of fast charging events before current charging is found to significantly reduce the likelihood of normal
charging at home and fast charging at public stations for all the private BEVs. Note that such conclusion holds with respect
to the normal charging at public stations. It indicates that private BEVs with more fast charging events in the past are more
likely to perform fast charging and normal charging at public stations rather than home charging. This phenomenon is nat-
ural since the more familiar the private BEVs with the fast charging stations, the more trustful they are for this technology
and service. Special incentives can be provided for new comers to encourage them to perform charging at public stations in
their initial stage of utilization, e.g., the first 10 free fast charging trials.
For the commercial BEVs, the effects of the number of fast charging events on both the normal charging at company
and fast charging at public stations are found to be significantly positive. Additionally, almost all the positive estimates of
the additional dummy variables are statistically significant. This suggests that commercial BEVs are more willing to perform
normal charging at company and fast charging at public stations if they have more fast charging experiences in the past.
Besides, the magnitude of the estimate for fast charging is much larger than that for normal charging at company, especially
in the base interval, implying the greater effect of this attribute on fast charging at public stations.
By further exploring the dummy variables, it can be seen that the values of the coefficients for the dummy variables
generally follow a descending order, suggesting that the effect of this attribute is decreasing with the growing number of
fast charging events. The result is consistent with the aforementioned saturation effect in the experience curve for the fast
charging at public stations. However, unlike private BEVs, the familiarity of commercial BEVs with fast charging at public
stations does not necessarily lead to the popularity of normal charging at public stations. Also, the magnitude of coefficient
for fast charging of commercial BEVs are much larger than that of private BEVs. These distinctions may arise from the
speculation that commercial BEVs are more concerned with the charging efficiency rather than charging cost.
5. Conclusions and future research
The choice behavior for charging mode and location of BEV users with revealed preference data in Japan has been suc-
cessfully investigated by means of the mixed logit modelling method. The model parameters estimation and results inter-
pretations/discussions are conducted for private and commercial BEVs respectively. Note that no research has been carried
JID: TRB
out for the choice behavioral analyses of charging mode and location using revealed preference data, which validates the
novelty of this study. In addition, we have proposed a tangible procedure to perform data pre-processing in order to facili-
tate the application of DCM. An appropriate instrument, i.e., the day interval between the end of former trip and the start
of charging event, is identified to correct the endogeneity issue in the model and significant improvement of model fit has
been achieved. Those contributions have demonstrated the significance of our study.
The results indicate that the effect of parameters on the choice of private and commercial BEVs for charging mode
and location are quite different. For instance, both the initial SOC and number of fast charging events enter the utility
of commercial BEVs nonlinearly, while the nonlinearity for private BEVs is not obvious. Also, the taste variation exists for
many attributes. A prominent example is the great preference heterogeneity among commercial BEVs for the choice of fast
charging conditional on T/K indicator.
Although the estimation results reported herein is specific to the early BEV adopters in Japan, the findings are illustrative
of issues that may arise in other countries with mass demonstration of vehicle electrification. Some implications can be
explored from these findings. First, better coordination between BEV charging and electricity grid system can be achieved
by incentives like lower electricity tariff at off-peak hours. Those practices have been proven effective to achieve desirable
effect of peak-load shifting by attracting BEV users to perform charging at midnight.
Second, for the operation of public stations, on one hand, encouraging the adoption of BEVs among government or com-
panies can better address the low utilization problem of fast charging stations because they are more concerned with the
charging efficiency and are generally more risk-adverse. On the other hand, some drivers may not have easy access to fast
charging stations although the coverage of public charging stations on the whole is extensive. Hence by improving the
deployment of public charging stations, more BEV users can be induced to charge BEVs at public stations to sustain its op-
eration. Also, operator of public charging stations may seek to provide better service or special incentives for new comers
to encourage them to perform fast charging at their initial stage of utilization.
Third, the life-cycle cost and benefit analyses for the public charging stations should be conducted to determine the
optimal density and deployment of public charging stations in a long term, where demand uncertainty for public charging
stations as well as the dynamic risk-adverse behavior of BEV users are accommodated. Moreover, the strategy to deploy
public stations should also take into account the local environment/development. For instance, for those settling in high-
rise apartment buildings in some well-developed cities, home charging may not be available and they have to rely solely on
public charging stations.
More explanatory variables may be included in the future research. For examples, the charging price/electricity tariff is
expected to have significant effect on the joint choice for charging location and mode. With the charging price/electricity
tariff parameter incorporated in the discrete choice model, the willingness to pay for normal charging at home/company and
fast charging at public charging stations can be obtained, which may provide implications for the price-setting of electricity
stakeholders and fast charging stations. Besides, stated preference experiment may be carried out to overcome the limitation
of current analysis. For instance, the effect of low SOC on the charging choice of commercial BEVs cannot be fully captured
due to little invariant of this attribute in reality. With the help of stated preference, more reliable results can be interpreted
combining the outcomes from both stated and revealed preference experiment.
Acknowledgments
We are indebted to the guest editor and three anonymous reviewers for their thoughtful comments and suggestions that
have helped substantially improve this work. We also thank the Japan Automobile Research Institute (JARI) for providing the
GPS dataset. This study was partly supported by the research project “Studying Autonomous Vehicles Policies with Urban
Planning of Toa Payoh in Singapore” funded by L2NIC of Singapore. The third and fourth authors would like to appreciate the
support of the Grant-in-Aid for Scientific Research (No. 25289164) from the Ministry of Education, Culture, Sports, Science
and Technology of Japan as well as the Japan Society for the Promotion of Science, and National Natural Science Foundation
of China (Grant No. 51378091) as well as Fundamental Research Funds for the Central Universities of China (Grants No.
DUT12ZD203).
References
Adamowicz, W.L., 1994. Habit formation and variety seeking in a discrete choice model of recreation demand. J. Agric. Resour. Econ. 19–31.
Arslan, O., Karaşan, O.E., 2016. A Benders decomposition approach for the charging station location problem with plug-in hybrid electric vehicles. Transp.
Res. Part B 93, 670–695.
Ben-Akiva, M., Bolduc, D., 1996. Multinomial probit with a logit kernel and a general parametric specification of the covariance structure. Department of
Civil and Environmental Engineering Working paper. Massachusetts Institute of Technology.
Berry, S., Levinsohn, J., Pakes, A., 1995. Automobile prices in market equilibrium. Econometrica 841–890.
Boyacı, B., Zografos, K.G., Geroliminis, N., 2017. An integrated optimization-simulation framework for vehicle and personnel relocations of electric carsharing
systems with reservations. Transp. Res. Part B 95, 214–237.
Brownstone, D., Bunch, D.S., Train, K., 20 0 0. Joint mixed logit models of stated and revealed preferences for alternative-fuel vehicles. Transp. Res. Part B 34
(5), 315–338.
Clement-Nyns, K., Haesen, E., Driesen, J., 2010. The impact of charging plug-in hybrid electric vehicles on a residential distribution grid. Power Syst., IEEE
Trans. 25 (1), 371–380.
Dagsvik, J.K., Wennemo, T., Wetterwald, D.G., Aaberge, R., 2002. Potential demand for alternative fuel vehicles. Transp. Res. Part B 36 (4), 361–384.
Daina, N., 2013. Electric Vehicle Market: Stated Valuation of the Charging Operation Oxford.
Fetene, G.M., Hirte, G., Kaplan, S., Prato, C.G., Tscharaktschiew, S., 2016. The economics of workplace charging. Transp. Res. Part B 88, 93–118.
JID: TRB
Ghamami, M., Zockaie, A., Nie, Y.M., 2016. A general corridor model for designing plug-in electric vehicle charging infrastructure to support intercity travel.
Transp. Res. Part C. 68, 389–402.
Glerum, A., Stankovikj, L., Thémans, M., Bierlaire, M., 2013. Forecasting the demand for electric vehicles: accounting for attitudes and perceptions. Transp.
Sci. 48 (4), 483–499.
Hackbarth, A., Madlener, R., 2013. Consumer preferences for alternative fuel vehicles: A discrete choice analysis. Transp. Res. Part D 25, 5–17.
Hadley, S.W., 2006. Impact of plug-in hybrid vehicles on the electric grid. No. ORNL/TM-2006/554. October.
Hagerty, J.R., Ramsey, M., 2011. Charging stations multiple but electric cars are few. Wall Street Journal.
He, S.Y., Kuo, Y.H., Wu, D., 2016. Incorporating institutional and spatial factors in the selection of the optimal locations of public electric vehicle charging
facilities: A case study of Beijing, China. Transp. Res. Part C 67, 131–148.
He, F., Wu, D., Yin, Y., Guan, Y., 2013. Optimal deployment of public charging stations for plug-in hybrid electric vehicles. Transp. Res. Part B 47, 87–101.
He, F., Yin, Y., Lawphongpanich, S., 2014. Network equilibrium models with battery electric vehicles. Transp. Res. Part B 67, 306–319.
He, F., Yin, Y., Zhou, J., 2015. Deploying public charging stations for electric vehicles on urban road networks. Transp. Res. Part C 60, 227–240.
Hess, S., Adler, T., Polak, J.W., 2007. Modelling airport and airline choice behaviour with the use of stated preference survey data. Transp. Res. Part E 43 (3),
221–233.
Hess, S., Bierlaire, M., Polak, J.W., 2005. Estimation of value of travel-time savings using mixed logit models. Transp. Res. Part 39 (2), 221–236.
Hess, S., Polak, J.W., 2004. Mixed Logit estimation of parking type choice. In: 83rd Annual Meeting of the Transportation Research Board. Washington, DC,
pp. 561–582.
Hess, S., Polak, J.W., 2005. Mixed logit modelling of airport choice in multi-airport regions. J. Air Transp. Manage. 11 (2), 59–68.
Hess, S., Train, K.E., Polak, J.W., 2006. On the use of a modified latin hypercube sampling (MLHS) method in the estimation of a mixed logit model for
vehicle choice. Transp. Res. Part B 40 (2), 147–163.
Hof, J., Schneider, M., Goeke, D., 2017. Solving the battery swap station location-routing problem with capacitated electric vehicles using an AVNS algorithm
for vehicle-routing problems with intermediate stops. Transp. Res. Part B 97, 102–112.
Jabeen, F., Olaru, D., Smith, B., Braunl, T., Speidel, S., 2013. Electric vehicle battery charging behaviour: findings from a driver survey. Australasian Transport
Research Forum (ATRF), 36th, 2013.
Jensen, A.F., Cherchi, E., Mabit, S.L., 2013. On the stability of preferences and attitudes before and after experiencing an electric vehicle. Transp. Res. Part D
25, 24–32.
Jiang, N., Xie, C., Waller, S., 2012. Path-constrained traffic assignment. Transp. Res. Record 2283, 25–33.
Li, S., Huang, Y., Mason, S.J., 2016a. A multi-period optimization model for the deployment of public electric vehicle charging stations on network. Transp.
Res. Part C 65, 128–143.
Li, X., Ma, J., Cui, J., Ghiasi, A., Zhou, F., 2016b. Design framework of large-scale one-way electric vehicle sharing systems: A continuum approximation
model. Transp. Res. Part B 88, 21–45.
Liao, C.S., Lu, S.H., Shen, Z.J.M., 2016. The electric vehicle touring problem. Transp. Res. Part B 86, 163–180.
Liu, H., Wang, D.Z., 2017. Locating multiple types of charging facilities for battery electric vehicles. Transp. Res. Part B. (in press).
Mak, H., Rong, Y., Shen, Z.M., 2013. Infrastructure planning for electric vehicles with battery swapping. Manage. Sci. 59 (7), 1557–1575.
Nie, Y.M., Ghamami, M., 2013. A corridor-centric approach to planning electric vehicle charging infrastructure. Transp. Res. Part B 57, 172–190.
Nie, Y.M., Ghamami, M., Zockaie, A., Xiao, F., 2016. Optimization of incentive polices for plug-in electric vehicles. Transp. Res. Part B 84, 103–123.
Nilsson, M., 2011. Electric vehicles: the phenomenon of range anxiety. ELVIRE, Gif-sur-Yvette Cedex, France, ELVIRE Consortium FP7CICT-2009C4-249105.
Pearre, N.S., Kempton, W., Guensler, R.L., Elango, V.V., 2011. Electric vehicles: How much range is required for a day’s driving. Transp. Res. Part C 19 (6),
1171–1184.
Petrin, A., Train, K., 2010. A control function approach to endogeneity in consumer choice models. J. Marketing Res. 47 (1), 3–13.
Revelt, D., Train, K., 1998. Mixed logit with repeated choices: households’ choices of appliance efficiency level. Rev. Econ. Stat. 80 (4), 647–657.
Revelt, D., Train, K., 20 0 0. Customer-specific taste parameters and mixed logit: households’ choice of electricity supplier. Department of Economics, UCB.
Russo, E., 2015. Public electric-car charging stations sit idle most of time, The Seattle Times.
Sathaye, N., Kelley, S., 2013. An approach for the optimal planning of electric vehicle infrastructure for highway corridors. Transp. Res. Part E 59, 15–33.
Smart, J., Powell, W., Schey, S., 2013. Extended Range Electric Vehicle Driving and Charging Behavior Observed Early in the EV Project. Society of Automotive
Engineers (SAE) Technical Paper, 01-1441.
Sun, X.-H., Yamamoto, T., Morikawa, T., 2015a. Charge timing choice behavior of battery electric vehicle users. Transp. Res. Part D 37, 97–107.
Sun, X.-H., Yamamoto, T., Morikawa, T., 2015b. Stochastic frontier analysis of excess access to mid-trip battery electric vehicle fast charging. Transp. Res.
Part D 34, 83–94.
Train, K.E.. Mixed logit estimation by maximum simulated likelihood https://eml.berkeley.edu/∼train/software.html. (accessed 16.08.2016).
Train, K.E., 2009. Discrete Choice Methods With Simulation. Cambridge University Press.
Trigg, T., Telleen, P., Boyd, R., Cuenot, F., D’Ambrosio, D., Gaghen, R., Gagné, J., Hardcastle, A., Houssin, D., Jones, A., 2013. Global EV Outlook: Understanding
the Electric Vehicle Landscape to 2020. International Energy Agency, pp. 1–40.
Waraich, R.A., Galus, M.D., Dobler, C., Balmer, M., Andersson, G., Axhausen, K.W., 2013. Plug-in hybrid electric vehicles and smart grids: Investigations based
on a microsimulation. Transp. Res. Part C 28, 74–86.
Xu, M., Meng, Q., Liu, K., 2017. Network user equilibrium problems for the mixed battery electric vehicles and gasoline vehicles subject to battery swapping
stations and road grade constraints. Transp. Res. Part B 99, 138–166.
Zoepf, S., MacKenzie, D., Keith, D., Chernicoff, W., 2013. Charging choices and fuel displacement in a large-scale plug-in hybrid electric vehicle demonstra-
tion. Transp. Res. Board 92.

Kendaraan Listrik 2

Uploaded by

Copyright:

Available Formats

Kendaraan Listrik 2

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Kendaraan Listrik 2

Uploaded by

Copyright:

Available Formats

JID: TRB

ARTICLE IN PRESS [m3Gsc;March 16, 2017;19:29]

Transportation Research Part B 0 0 0 (2017) 1–19

Contents lists available at ScienceDirect

Transportation Research Part B

Joint charging mode and location choice model for battery

2 M. Xu et al. / Transportation Research Part B 000 (2017) 1–19

M. Xu et al. / Transportation Research Part B 000 (2017) 1–19 3

Fig. 1. Non-residential charging stations by slow/fast in 2012 (Trigg et al., 2013).

4 M. Xu et al. / Transportation Research Part B 000 (2017) 1–19

2.1. Probe BEV data pre-processing procedure

M. Xu et al. / Transportation Research Part B 000 (2017) 1–19 5

Event Starting of Event Ending of Event

Raw Probe Data

Raw Travelling Raw Charging Database

Fig. 2. Probe BEV data pre-processing procedure.

BEV type First year Second year Total

Private 3,909 34,892 38,801

6 M. Xu et al. / Transportation Research Part B 000 (2017) 1–19

Metrics Private Vehicles (n= 234) Commercial Vehicles (n= 249)

Min. Median Mean Max. Min. Median Mean Max.

2.2. Observed driving & charging behavior in Japan

M. Xu et al. / Transportation Research Part B 000 (2017) 1–19 7

Fig. 3. Distribution of four metrics at the vehicle level.

3. Mixed logit model building

Fig. 4. Number of times each charging alternatives being chosen.

8 M. Xu et al. / Transportation Research Part B 000 (2017) 1–19

3.1. Explanatory variables identiﬁcation

M. Xu et al. / Transportation Research Part B 000 (2017) 1–19 9

3.2. Endogenous variable and its instrument

Variable Private Vehicles (n= 234) Commercial Vehicles (n= 249)

Normal_Home Normal_Public Fast_Public Total Normal_Company Normal_Public Fast_Public Total

M. Xu et al. / Transportation Research Part B 000 (2017) 1–19

Value in brackets are the standard deviation (SD) of parameters.

[m3Gsc;March 16, 2017;19:29]

M. Xu et al. / Transportation Research Part B 000 (2017) 1–19 11

3.3. Mixed logit model speciﬁcation

εn jt = εn1 jt + εn2 jt (3)

12 M. Xu et al. / Transportation Research Part B 000 (2017) 1–19

4. Results and discussions

Variable Alternative Estimates Without Correction Estimates With Correction

mean t-statistic sd t-statistic p< 0 mean t-statistic sd t-statistic p< 0

M. Xu et al. / Transportation Research Part B 000 (2017) 1–19

[m3Gsc;March 16, 2017;19:29]

Variable Alternative Estimates Without Correction Estimates With Correction

mean t-statistic sd t-statistic p< 0 mean t-statistic sd t-statistic p< 0

M. Xu et al. / Transportation Research Part B 000 (2017) 1–19

[m3Gsc;March 16, 2017;19:29]

M. Xu et al. / Transportation Research Part B 000 (2017) 1–19 15

4.1. Implication of useable battery capacity of BEVs (Variable 1)

Variable 1 (an indicator for longer driving range)

4.2. Implication of charging time and day (Variables 2–3)

Variable 2 (An indicator for midnight from 23:00 to 7:00)

16 M. Xu et al. / Transportation Research Part B 000 (2017) 1–19

Variable 3 (An working day indicator)

4.3. Implication of density of public charging stations (Variable 4)

Variable 4 (T/K indicator)

4.4. Implication of business type for commercial BEVs (Variable 5)

Variable 5 (An indicator for BEVs from Electricity Company)

4.5. Implication of “range anxiety” (Variable 6)