Professional Documents
Culture Documents
Taxi Trajectory Data Based Fast-Charging Facility Planning For Urban Electric
Taxi Trajectory Data Based Fast-Charging Facility Planning For Urban Electric
Applied Energy
journal homepage: www.elsevier.com/locate/apenergy
Taxi trajectory data based fast-charging facility planning for urban electric
taxi systems
Hua Wang, De Zhao, Yutong Cai, Qiang Meng ∗, Ghim Ping Ong
Department of Civil & Environmental Engineering, National University of Singapore, 117576, Singapore
Keywords: This study develops a taxi trajectory data based fast-charging facility planning model for an urban taxi system
Electric taxi by considering battery degradation and vehicle heterogeneity in driving range. The developed model comprises
Taxi trajectory three functional modules: (i) fast-charging location determination, (ii) fast-charging facility deployment (FCFD)
Fast-charging facility deployment
and (iii) FCFD solution tuning. Under the FCFD module, charging demand prediction considering battery
Battery degradation
degradation and vehicle heterogeneity, charging demand allocation and charger configuration optimization
Vehicle heterogeneity in driving range
Multi-objective planning
are executed sequentially. The FCFD solution is tuned by an effective backward elimination method to find
a more economic and practical planning solution where the minimal number of chargers at each station
can be specified. A case study in Singapore is thoroughly conducted, and insightful policy implications are
revealed: policy-makers could use the tuning mechanism to significantly save investment and reduce total
waiting time for charging; overlooking battery degradation and vehicle heterogeneity will yield a biased electric
taxi charging facility planning.
1. Introduction forecasted that more than 30 million (24%) passenger vehicles in the
US will be electrified by 2040 [3]. Of course, EVs have been gaining
In the new century, electric vehicles (EVs) has attracted an increas- an increasing acceptance in taxi industry and is also paving a way to
ing attention from car consumers, manufacturers and transport sectors achieve the goal of high electrification and automation. For example,
because of potential merits of energy saving, zero exhaust emission and Tesla’s products are becoming popular with taxis in European countries
lower maintenance cost. Thanking to continuous national-wide policy
such as Netherlands, Sweden and Norway, and Tesla Model 3 is the first
support such as purchase subsidy and license incentive, the EV industry
full-electric vehicles to be approved for use as New York City yellow
has experienced an explosive development, where the top three EV
cab in 2019. In Singapore, a similar development tendency appears.
markets (China, the US and Norway) contribute to about 1.5 million EV
sales in 2018. The report ‘‘Global EV Market Outlook 2019’’ released ComfortDelgro (a local taxi magnate), Grab Taxi (a new IT ride-hailing
by Frost & Sullivan indicated that the global EV fleet in 2018 hit operator) and the HDT Taxi (a taxi firm who fully utilizes EVs) are
5.1 million, up 2 million from the last year and nearly doubling the looking to a rapid rollout for more new green vehicles. For example,
quantity of new sales [1]. According to the market forecasting made by the HDT Taxi has deployed more than one thousand electric-powered
Electric Power Research Institute [2], consumers have more purchase cabs.
choices in EV brands and EV models in near future, as about 132 new It is worth noting that the average travel mileage of an taxi is often
models will be projected by 2022. ‘‘Global EV Outlook 2019’’ also over 300 km per day and as a result heavy emission pollutants will be
optimistically predicted in its EV30@30 scenario of future development generated given a large taxi fleet in a city, e.g., over 23,000 operating
that EV sales around the world will reach 43 million and the stock taxis in Singapore. In predictable future, decision-makers from urban
number will greater than 250 million by 2030. planning and transport authorities would have to consider an electrifi-
In context of a prosperous development of EV industry and the
cation upgrading of urban taxi system. Although vehicle electrification
Paris Agreement of controlling climate change in this century, quite a
has been viewed as a sound solution to control exhaust emissions and
few governments worldwide make electrification plans to reduce car
alleviate fossil fuel reliance [4–6], it still needs to overcome several key
exhaust emissions, one of which is to replace gasoline cars by EVs.
For example, the Energy Information Administration (EIA) of the US challenges regarding battery technology and charging infrastructure
∗ Corresponding author.
E-mail addresses: hwang191901@gmail.com (H. Wang), ceezde@nus.edu.sg (D. Zhao), e0133218@u.nus.edu (Y.T. Cai), ceemq@nus.edu.sg (Q. Meng),
ceeongr@nus.edu.sg (G.P. Ong).
https://doi.org/10.1016/j.apenergy.2021.116515
Received 6 November 2020; Received in revised form 6 January 2021; Accepted 13 January 2021
Available online 22 January 2021
0306-2619/© 2021 Elsevier Ltd. All rights reserved.
H. Wang et al. Applied Energy 286 (2021) 116515
planning, both of which have lagged to match the cracking pace set analysis approach to find the optimal EV charging locations based on
with the prosperous EV market. Decision-makers need to make tangible actual data of conventional private car travels within the urban area of
charging facility deployment schemes that addresses a set of issues in Rome. He et al. [27] performed a case study on Beijing road network
the running of vehicle electrification for taxi systems. For instance, to test three CFD models (the set covering model, maximal covering
how to make a high-quality forecasting of charging demand of electric location model and p-median model). Wang et al. [34] proposed an
taxis (ETs) and is it possible to do a time-dependent charging demand easy-to-perform four-step CFD method based on macroscopic planning
estimation based on existing taxi trajectories? Based on the estimated statistics regarding residential carparks, private EVs and ETs and val-
charging demand, we need to figure out the locations of charging idated the method through a case study in Singapore. Vazifeh et al.
stations and the respective number of chargers to be deployed at each [24] put forward a data-driven optimization approach for EV charging
station. In addition, how to balance different planning objectives when location problem, in which the geographical area is divided into equal
deploying charging facilities? This is defined as the charging facility grids and the number of trips stopping in a cell is counted based
deployment (CFD) problem (or called charging location problem) in on EV individuals’ travel records. Baresch and Moser [29] examined
the literature. Our research focus/scope in this paper is restricted to the distribution of future charges across various types of charging
urban ET system, while in the following we still make a comprehensive stations in Austria so as to assess the need for charging infrastructure,
literature review to elaborate the theoretical gaps. i.e. its number, design and cost-effectiveness. Grote et al. [30] proposed
a GIS based analytical method by using routinely available census
1.1. Literature review and parking data to identify suitable street-side charging points, and
made an application in Southampton, UK. Jochem et al. [31] used an
In the literature, there are two prevailing research paths for the CFD arc-cover-path-cover flow-refueling location model to determine the
problem, namely optimization models and data-driven methods. As for minimal number of fast charging stations along European highway
the optimization research line, Nozick [7] is the first one to formulate network based on the dataset of passenger car trips in Europe. Dong
the CFD problem with coverage constraints as a mixed integer linear et al. [33] developed a two-step method to optimize charging locations
programming model. In spirit of this seminal work, a large number by combining spatial statistics and maximal coverage location model,
of optimization models are formulated to solve the CFD problem, and explored the deployment of charging points in London. Huang et al.
which are further categorized into two mathematical programming [32] put forward a novel GIS assisted charging infrastructure planning
groups. On one hand, dozens of bi-level programming models are model for high-density cities where the life-cycle cost of charging
developed for the CFD problem so as to model EV users’ routing and stations subject to a pre-set coverage rate was minimized, and made
charging behaviors [8–11]. At the upper level, policy-makers/planners an empirical validation in Hong Kong.
aim to maximize charging network’s performance (e.g., travel cost In addition, decision-making approaches are employed for solving
minimization or service coverage rate maximization) by determining EV charging infrastructure planning problems. For example, Guo and
optimal charging locations and/or charger configuration. At the lower Zhao [35] proposed a multi-criteria fuzzy TOPSIS (Technique for Order
level, a user equilibrium (UE) model is always applied to depict users’ of Preference by Similarity to Ideal Solution) method to determine the
travel decisions, such as route and charging choices [12]. On the other optimal EV charging locations from sustainability perspective.
hand, quite a few scholars focus on developing single-level optimization
models for the CFD problem [13–19]. For this research stream, integer 1.2. Research motivation and the position of our work
linear programming models are proposed to pursue optimal coverage of
charging infrastructure and/or maximize total EV travel mileage with We below make some remarks on the above literature review,
a budget constraint. In some cases, such models could be simplified to highlight the important gaps that are imperative to be filled in the CDF
a shortest path procedure that assigns EVs to desirable charging points. problem for ET systems, and clarify the position of our work.
The optimization models have also been extended to address a We first look at research path of optimization models. Although
series of complicated traffic scenarios, for example, the CFD prob- there are a large number of optimization models for the CFD problem,
lem with stochastic demand [18,19], wireless charging infrastruc- they would be inapplicable for the ET charging infrastructure planning
ture planning [9], the CFD problem with particular EV types (EV due to the following reasons. First, almost all bi-level optimization
buses/taxis) [15,20], the sizing optimization of workplace charging models in the literature are based on the rush-hour commuting pat-
stations considering probabilistic reactive power support for plug-in tern during morning or evening peak and rely on the equilibrium
hybrid EVs [21] and the CFD problem with driving range anxiety [19]. assumption; while ET fleet is operating in a different manner — pickup-
One of the main challenges that restricts the adoption of optimization delivery pattern. In fact, as pointed out by Xu et al. [36], Muratori et al.
models for solving CFD problems in real world is the efficiency of [37] and Baresch and Moser [29], the majority of commuting driving
solution algorithms. Both global solution algorithm and meta-heuristics of private EV users (over 85%) could be well catered for home-based
are investigated in the literature. For example, Arslan and Karaşan [16] charging. However, ETs, particularly the two-shift ones, will replenish
proposed a Benders decomposition algorithm to solve the CFD problem depleted energy during operational duration and thus behave in an en-
in California road network. Lee and Han [22] put forward a Benders- route charging manner. This implies that charging demand predication
and-Price algorithm with column generation procedure to solve the mechanisms for two patterns are totally different. Besides, bi-level opti-
CFD model. Genetic algorithm is also widely used for solving the CFD mization models are rarely applied for real large-size networks because
optimization models [17,18,23,24]. of heavy computational burden (see more details in Section 3.2). This
Recently, data-driven methods have received a rising attention and also restricts applying them for solving CFD problems for ET systems
extensive empirical investigations are made [15,20,24–33]. For exam- in practice. Second, the single-level optimization models in previous
ple, public charging station locations for ET systems were studied based studies highly rely on two assumptions: all EVs on roads are under con-
on historic taxi trajectories in Beijing [25] and Vienna, Austria [15]. In trol by a central operator and no queue exists at charging point. Even
addition, EV bus charging station planning in Stockholm is analyzed though ET fleets belong to limited taxi firms, and ET drivers can receive
by Xylia et al. [20]. Zhang et al. [14] developed a set covering model booking requests and suggested charging strategies from the control
to choose the minimum set of the candidate charging locations to centers of respective firms, they might still make own decisions based
serve possible routes as many as possible, where vehicle travels were on their personal perception on road/traffic condition and operational
extracted from the 2000 California Household Travel Survey and the experience on battery state. The system optimum derived from a full
shortest paths between the known geographic positions were deter- cooperation might be hardly achieved due to self-motivated manners
mined by the ArcGIS platform. Andrenacci et al. [28] applied a cluster of ET drivers.
2
H. Wang et al. Applied Energy 286 (2021) 116515
We then move to another research branch, data-driven methods. charger number with an objective to minimize the weighted sum of
These methods mainly focused on determining suitable charging lo- total waiting time and total cost of constructing charging stations.
cations by statistical and/or cluster-similar analyses based on parking, This study makes the following substantial contributions to the
trip and trajectory data. They seldom concern on the second key issue literature of the CFD problem.
of the CFD problem — charging capacity/charger configuration at
(1) To the best of our knowledge, our work is the first attempt to
each station, but simply assume equal/uncapacitated charging service
integrate data-driven analysis of charging demand estimation
capacity for each station [24,27].
and optimization model together for solving the CFD problem
More importantly, an undeniable truth is that the models/methods
for ET systems. It could fill the gap and well address the above
in both research branches fail to characterize complex charging de-
mentioned limitations in the literature.
cisions since charging demand forecasting has not received adequate
(2) The proposed method could balance multiple charging system
attention (it is either assumed to be known and given, or improp-
objectives in ET charging infrastructure planning, including
erly derived based on travel demand between given origin–destination
charging service coverage rate, level of charging service (mea-
pairs). On one hand, without an accurate temporal–spatial distribution
sured by total waiting time for charging) and total construction
of charging demand, it is almost impossible to account for queuing
cost. As a result, a more tangible and reasonable FCFD scheme
effect and turnaround time at each charging point. As a result, the exact
could be obtained.
charger number and level of charging service of the CFD scheme cannot
(3) We incorporate the two most important influential factors –
be well determined and assessed. On the other hand, missing a proper
battery degradation and vehicle heterogeneity – into charging
charging demand prediction procedure, it is hard to consider the impact
demand estimation. This would help to design an unbiased FCFD
of practical influential factors, such as battery degradation, charging
plan for ET systems.
preference and vehicle difference, and thus a biased CFD plan might be
(4) We propose a refinement mechanism to iteratively improve the
obtained. It is easy to understand that the amount of charging demand
ET–FCFD scheme and thus an more economic and reasonable
will be definitely underestimated if battery degradation is neglected,
plan could be ensured.
since the actual travel mileage of a full charge is shorter than ideal
(5) Last but not least, we validate the developed method by ap-
state of a new battery. Additionally, vehicle heterogeneity should not
plying it to a Singapore case study. Some insightful findings
be ignored since more than hundred EV brands exist in the actual
are unveiled: (i) the refinement mechanism could help to save
market and they exhibit distinct vehicle particulars, especially in driv-
investment and enhance system performance in reducing to-
ing range. There are two driving forces behind vehicle heterogeneity
tal waiting time for charging; and (ii) overlooking the two
in ET market: (i) taxi operators periodically deploy new batches of
influential factors will lead to a biased planning of charging
EVs of various brands/generations to upgrade their own fleet, and (ii)
infrastructure and the system objectives will be optimistically
within individual ET firm, a certain proportion of drivers operate with
overestimated, up to 19% of overestimation of reducing the total
their own EVs of diverse models. The importance of considering these
waiting time.
factors has been highlighted in the literature, for example, for the EV
battery life predication [38,39] and energy consumption analysis [39– The structure of the remaining paper is stated below. The problem
41]. For example, Wang et al. [40] reported that more than 10% is explained in detail in Section 2, followed by a conceptual ET–FCFD
energy consumption estimation bias occurs if battery degradation is not optimization model in Section 3 and our research methodology in
considered. In reality, it is also worth noting that the availability of ETs’ Section 4. In Section 5, a case study in Singapore is carried out. The
real-time trajectories creates a nice chance to forecast time-dependent last section concludes the paper.
charging demand.
To fill the above research gaps, we in this study are devoting 2. Problem description
to proposing a data-driven optimization method for solving the CFD
problem for ET systems that combines the merits of two research In this paper, we focus on the fast-charging facility deployment
paths. Moreover, we in this paper focus on balancing multiple sys- problem for an electric taxi system (called ET–FCFD problem hereafter)
tem objectives of ET charging infrastructure planning and understand- by taking into account multi-dimensional considerations, including
ing/quantifying the impact of battery degradation and vehicle hetero- multiple objectives and impacts of battery degradation and vehicle het-
geneity in driving range. erogeneity (multi-type taxis with different driving ranges).1 In general,
ET drivers would prefer using fast chargers to recharge the batteries so
1.3. Objectives and contributions as to save more time for their business, such as passenger searching,
passenger pick up and delivery. Although this study focuses on fast-
The objective of this research is to find the optimal fast charging charging facility planning, the proposed method could be adapted to a
facility deployment plan that determines suitable charging station lo- general case with multi-type chargers.
cations and specifies the best charger configuration (the number of fast To make a tangible ET–FCFD plan, several important and practical
chargers) for each station to support operation of urban electric taxi factors shall be taken into account. In this study, we incorporate vehicle
(ET) system. heterogeneities in terms of battery degradation and driving range into
To achieve this goal, we propose a three-module electric taxi fast- the ET–FCFD problem. Both factors exert crucial impacts on ETs’ charg-
charging facility deployment (ET–FCFD) method that addresses multi- ing demand distribution on the network. First, battery degradation is
ple charging system objectives and incorporates the impact of battery
degradation and vehicle heterogeneity in driving range. The proposed
1
method consists of a module of charging point selection with coverage We in this study focus on the fast-charging facility planning exclusively for
rate requirement, a three-step FCFD module and a refinement module. ET system for the following considerations. Different from fuel/petrol stations,
the ET charging infrastructure is seldom open for other types of EVs, such as
For the second core module, we develop an ET trajectory based charg-
private EVs and electric buses. Typically, ET firms operate with their own EV
ing demand estimation by taking into account battery degradation and
fleets and establish exclusive charging infrastructure. To develop an economic
vehicle heterogeneity, and propose an allocation model to determine charging infrastructure network, each ET operator would maintain a high
the temporal–spatial distribution of charging demand and thus get the utilization rate of its infrastructure and thus would not be likely to share its
charging request arrival pattern for each station. With the assigned charging services with other types of EVs, especially with the consideration of
demand, we formulate an optimization model of determining optimal infrastructure maintenance.
3
H. Wang et al. Applied Energy 286 (2021) 116515
an indispensable factor. ETs in a taxi transport system would exhibit charging requests of ETs.2 We can formulate a generalized optimization
obvious diversity in served life, including new cars and different-year model for the ET–FCFD problem as follows.
vehicles, and their battery degradation levels are of course different. ( 𝑚 )
⎧ ⎫
On one hand, ETs’ charging frequencies are influenced by vehicles’ ⎪max𝑦𝑚𝑖 𝑓coverRate 𝑦𝑖 , ( , ) ⎪
⎪min 𝑚 𝑓 𝑚 ⎪
battery degradation levels. It is easy to understand that, to fulfill the (ET–FCFD model) constructCost (𝑦𝑖 ) (1)
⎨ 𝑦𝑖
( ( ))⎬
same travel mileage, a new ET needs less times of re-charging and an ⎪min 𝑚 𝑚 𝑓waitTime 𝑦 , 𝑖 𝑥 , 𝑦
( ) 𝑚 𝑚 𝑚 ⎪
ET with a terrible degraded battery has to re-charge more frequently. ⎪ 𝑥𝑞,𝑖 ,𝑦𝑖 𝑖 𝑞,𝑖 𝑖 ⎪
⎩ ⎭
On the other hand, an ET’s temporal–spatial distribution of charging
s.t.
demand depends on its actual driving range which is closely related to
its battery status. 𝑦𝑚
𝑖 ∈ {0, 1}, ∀ 𝑖 ∈ , 𝑚 ∈ ∶= {1, 2, … , 𝑀},
Second, the disparity of vehicles’ driving ranges is also important. (2)
∑
As reported by EPRI in 2019, customers can find an EV in almost 𝑦𝑚
𝑖 ≥ 𝑦min , ∀ 𝑖 ∈ , (3)
all vehicle classes, with about 41 new models available now and 132 𝑚∈
projected by 2022. This unveils a fact that a variety of ET models will ∑ ∑
𝑥𝑚 𝑚
𝑞,𝑖 𝑦𝑖 = 1, ∀ 𝑞 ∈ , (4)
exist on the road network, which possess unique driving ranges. For 𝑖∈ 𝑚∈
instance, a 2019 Nissan Leaf has a 40 kWh battery and a range of 151 ∑ ∑
𝑥𝑚
𝑞,𝑖 = 1, ∀ 𝑞 ∈ , (5)
miles, while the long range version of Tesla model reaches a range 𝑖∈ 𝑚∈
of 310 miles. As pointed out by Wang et al. [41], a biased result of
𝑥𝑚
𝑞,𝑖 ∈ {0, 1}, ∀ 𝑞 ∈ , 𝑖 ∈ , 𝑚 ∈ , (6)
charging demand estimation would be obtained if a mean driving range
is used. where the binary variable 𝑦𝑚
is the decision variable for charging
𝑖
To sum up, because the optimization of ET–FCFD is based on ET facility planning that 𝑦𝑚 𝑖 = 1 if the 𝑚th fast charger at the 𝑖th charging
charging demand distribution, the above two influential factors should station will be built, 𝑦𝑚
𝑖 = 0 otherwise.∑ The total number of fast chargers
not be neglected. In order to account the two factors, we in this paper deployed at charging station 𝑖 is 𝑦𝑖 = 𝑚∈ 𝑦𝑚 𝑖 . Constraint (3) specifies
assume ETs’ battery degradation levels follow a known probability a practical requirement of minimal number of chargers (𝑦min ) to be
distribution and consider a probabilistic grouping in driving range deployed at charging station 𝑖 so as to enjoy economics of scale and
based on ET/EV market penetration forecasting. ensure a more cost-efficient and reasonable plan where system resource
For an ET system in a city, we in this study focus on solving three such as land and electrical grids could be well utilized.
problems for its fast-charging facility deployment: (i) time-dependent It is worth noting that the developed optimization model simultane-
charging demand estimation based on taxi trajectories, (ii) locations to ously determines optimal ET–FCFD plan (charging location, number of
build fast charging stations for ETs and (iii) number of fast chargers fast chargers at each station) and charging demand assignment scheme.
to be deployed at each station, by considering the impacts of battery Binary variables 𝑥𝑚 𝑞,𝑖 are defined for decision-making of charging re-
degradation and vehicle heterogeneity in driving range. Different from quest assignments that 𝑥𝑚 𝑞,𝑖 = 1 if a charging request 𝑞 ∈ will be
fulfilled at the 𝑚th charger at charging station 𝑖, 𝑥𝑚 𝑞,𝑖 = 0 otherwise.
previous studies, we here will address multiple system objectives for
Constraints (4) and (5) guarantee that each charging request 𝑞 will be
the ET–FCFD problem, including the overall coverage rate of charging
served only one time.
services, level of charging service (measured by the total waiting time
As shown in Eq. (1), we consider three optimization objectives. The
at charging station) and total construction cost.
first one is to maximize the coverage rate of charging facilities with a
given service radius for an urban transportation network denoted by
3. A conceptual optimization model ( , ) where is the set of nodes and is the set of directed links
(road segments). As for the second system optimization objective, we
We in this section present a conceptual optimization model for aim to minimize the total construction cost of charging infrastructure,
the ET–FCFD problem, and discuss technical intractability for apply- including equipment cost, installation cost and maintenance cost over
ing the optimization model to solve large-scale ET–FCFD problems whole charger lifetime. The last objective function is to enhance the
in real world. Although this conceptual optimization model has little level of charging service of ET charging system, which is measured by
value for solving real ET–FCFD problems (see detailed reasons in the total waiting time here. The waiting time at a charging station not
Section 3.2), we develop it to illustrate the modeling components such only depends on charger configuration 𝑦𝑚 𝑖 at the station
( but) also relies
as decision variables, system objectives and practical constraints of on arrival pattern of real-time charging requests 𝑖 𝑥𝑚 𝑚 (see more
𝑞,𝑖 , 𝑦𝑖
the ET–FCFD problem, which will facilitate the understanding of the details in Section 4.2.2).
logic and working principle of our proposed research methodology in
Section 4. 3.2. Challenges and application intractability of the ET-FCFD optimization
model
3.1. A conceptual ET-FCFD optimization model
Theoretically speaking, we could find the best solution for the ET–
FCFD problem by solving the formulated integer optimization (IP)
The ET–FCFD problem is to satisfy charging demand of ETs and
model (1)–(6). However, it is technically impossible to extend to large-
maximize the charging system performance from various perspectives,
scale problems in real life due to several intractable challenges. First,
including coverage rate of charging service network, level of charging
there exists a computational intractability issue for the large-scale IP
service measured by total waiting time and charging infrastructure
optimization model. Take a realistic planning scenario in Singapore
construction cost (equipment cost, installation cost and life-time main-
as an example. Consider 150 of candidate charging sites, the maximal
tenance cost) by determining optimal number of fast chargers at each
charging point. Let 𝑀 be a sufficiently large integer to denote the
maximum number of chargers that is feasible to deploy at a charging 2
Each charging request has both time and location attributes. The former
station subject to physical conditions (e.g., land-use, parking space determines the arrival sequence of charging requests at each charging station.
and electrical grid impact). Let be the set of candidate sites to The latter influences the actual SoC of the charging request when arriving at
build charging stations and be the set of spatio-temporal-dependent its desirable charging station.
4
H. Wang et al. Applied Energy 286 (2021) 116515
number of chargers to be deployed at each site 𝑀 = 30 and 6000 ETs ET–FCFD problem) of ET charging infrastructure could be satisfied.3
in urban transport system in 2025 that each re-charges twice everyday. The specific procedure is given below:
It is estimated that there are over 150 × 30 × 6000 × 2 = 54 millions of
Step 1: Prepare temporal and spatial distribution of charging demand
binary decision variables. Even the state-of-art commercial IP solvers,
(more information of charging demand estimation will be
such as CPLEX and GUROBI, are unable to solve the problem within
elaborated in next section), and location information of possi-
an acceptable time. One may seek help from using some decomposition
ble charging sites (depots, petrol stations and public carparks
techniques to make the problem solvable. However, the following extra
etc.) and set 𝑘 = 1 (the value could be empirically set).
challenges hinder such movements.
Step 2: Perform a K-means clustering analysis with 𝑘 clusters for the
Second, as discrete charging activities arrive in random at charging
given charging demand distribution. The K-means clustering is
points, each charging station should be treated as a queuing system
a method of vector quantization that aims to partition all road
with deterministic arrival rate and deterministic service rate with sev-
segments into 𝑘 clusters in which each road segment belongs
eral chargers (D/D/S system). In this sense, it is technically challenging
to derive a continuous closed-form function of the waiting time at each to the cluster with the nearest mean (cluster centers), serving
charger. Therefore, in the optimization model, it is almost impossi- as a prototype of the cluster.
ble to define an analytical objective function containing waiting time Step 3: Adjust the cluster centers to the locations included in the set of
(𝑓waitTime (⋅)) that is derived upon discrete customer arrival rates and possible charging sites according to a proximity rule; i.e., each
first-come-first-served working principle. This implies the optimization cluster center is re-aligned to its nearest available charging
model makes little sense in practice. site.
Third, charging requests are also heterogeneous due to vehicle Step 4: Given a charging service radius (e.g., 1 km), examine the cov-
heterogeneity in terms of battery degradation level and driving range. erage rate of the obtained cluster centers. If overall coverage
So they may have charging tasks which require different service times rate meets the pre-determined threshold (e.g., 85%), then stop
at desirable charging stations. This gives rise to an extra difficulty of and get the candidate charging sites; otherwise, let 𝑘 = 𝑘 + 1
quantitatively measure queuing effect for objective function 𝑓waitTime (⋅). and go to Step 2.
To deal with the different customer types, we need to differentiate the We here use a geospatial coverage rate (i.e., to maximize percent
charging demand assigned to each charger (not each charging station). of network area covered) for the following reason. A geospatial cov-
Last but not least, addressing multiple objectives in the proposed erage rate provides a system metric from the planning perspective
optimization model (1)–(6) also increase computational intractability which grants charging-infrastructure-to-road-infrastructure matching.
and difficulty in developing effective algorithms to find out optimal The adoption of this approach will not lead to an ET–FCFD scheme with
solutions on Pareto frontier. inferior charging request coverage. Once there are sparse taxi activities
We so far explain the main challenges and intractability to ap- as well as low charging requests around a candidate charging site,
ply the optimization model to solve large-scale ET–FCFD problems our developed model (Module 2 in next subsection) will assign fewer
in real world. We thus devote to proposing a tangible method for chargers at the particular charging station. In addition, we propose a
the problem in next section that integrates optimization of multiple refinement mechanism (Module 3) to remove those candidate charging
system objectives, charging demand estimation and allocation, and sites with very few chargers (e.g., < 2) to ensure a more economic ET–
FCFD optimization together. FCFD scheme. Therefore, our proposed method could ensure there is
no established charging station with very low charging demand.
4. Our research methodology
4.2. A three-step fast charging facility deployment (FCFD) approach
In this section, we propose a tailored research methodology for
solving large-scale ET–FCFD problems. Fig. 1 presents our research In the second module, we are going to optimize the configuration
methodology that comprises three fundamental modules. The first mod- of charging infrastructure (i.e., the number of fast chargers) for each
ule is to filter out candidate charging sites for ET–FCFD, where a charging site/station. As shown in Fig. 1, we in this module develop
pre-determined planning goal of spatial coverage rate (the 1st objective a three-step FCFD approach by considering the impacts of battery
𝑓coverRate (⋅)) is satisfied. In the second module, we propose a three- degradation and vehicle heterogeneity in driving range, and address-
step FCFD approach to optimize the number of fast chargers at each ing two planning objectives. The proposed approach consists of three
charging site by taking into the impact of battery degradation and components, namely charging demand estimation, charging demand
vehicle heterogeneity in driving range, and balance the other two allocation model and an optimization model for charging capacity
optimization objectives (waiting time and construction cost). Module determination. We in the following explain each of them in details.
3 aims to refine the charging locations in line with the practical
planning requirement of Constraint (3). Note that the proposed method 4.2.1. Charging demand estimation
decomposes the ET–FCFD problem into several subproblems to make its Charging demand distribution provides essential input data for
curse of dimensionality tractable. Although a global optimal solution the modeling of charging facility deployment. Similar to travel trip/
may not be ensured, it is indeed capable to find satisfactory solutions demand estimation, survey tools such as questionnaire interview and
for real large-scale ET–FCFD problems. on-site video could be applied to obtain ET charging demand, including
daily charging frequency and charging preference in charging duration
4.1. Module 1: Selection of candidate charging sites and charging time period etc. But such survey is always labor intensive,
time consuming and also very costly, and in most cases only a biased
In general, a suitable charging site should have adequate parking sample can be obtained.
space and good accessibility. At urban/traffic planning level, an effi- For a city with relatively stable population, mature economic en-
cient and favorable ET charging network should ensure a high coverage tities and developed road infrastructure, travel pattern within urban
rate of charging service provision. In urban area, candidate charging
sites could be taxi depots, petrol stations and public carparks. If there
is a very large number of possible charging sites, we can also apply a 3
Of course, some facility location models such as maximal-coverage
clustering analysis to facilitate the selection of candidate charging sites location model could be used to determine the candidate charging sites,
where a pre-determined service coverage rate (the 1st objective of the particularly when there are not too many potential charging points in a city.
5
H. Wang et al. Applied Energy 286 (2021) 116515
6
H. Wang et al. Applied Energy 286 (2021) 116515
s.t. where 𝑡𝜅 is the arrival time and 𝑡′𝜅 is the time that the charging server
∑ is ready to accept the 𝜅th charging request. Therefore, the total waiting
𝑧𝑞,𝑖 = 1, ∀ 𝑞 ∈ 𝑗 , 𝑗 ∈ , (9)
𝑖∈ time at the 𝑗th sample of battery degradation is
∑ 𝑖,𝑗
𝑧𝑞,𝑖 ∈ {0, 1}, ∀ 𝑞 ∈ 𝑗 , 𝑗 ∈ , (10) 𝑇𝑊 𝑇 = 𝑤queue , ∀ 𝑗 ∈ , (12)
𝑖∈
where 𝑧𝑞,𝑖 are binary decision variables, where 𝑧𝑞,𝑖 = 1 if the charging
request 𝑞 is assigned to be served at charging station 𝑖, 0 otherwise. The and the ATWT is then given by
objective function (8) is to minimize the total travel distance that ETs 1 ∑ ∑ 𝑖,𝑗
move to desirable charging stations. Constraints (9) require that each 𝐴𝑇 𝑊 𝑇 = 𝑤 , (13)
|𝐽 | 𝑗∈ 𝑖∈ 𝑞
charging request will be served once at only one charging station.
The above IP model could be relaxed to an equivalent linear pro- where 𝑤𝑖,𝑗
queue is the average queuing time at charging station 𝑖 at the
gramming if the minimum element in the travel distance vector 𝐬𝑞 ∶= sample 𝑗.
(𝑠𝑞,𝑖 , 𝑖 ∈ ) is unique. If so, the IP model can be simplified into a shortest
path assignment problem. (2) Charging capacity optimization model
Since there would be a certain distance from the place of an in- We now aim to determine an optimal decision of the number of fast
stantaneous charging request to desirable charging station, we need to chargers to be deployed at each charging station so as to find a trade-off
calibrate the actual time and SoC of an ET arriving at the charging between ATWT minimization and total construction cost minimization
station. The actual time of charging request 𝑞 arriving at charging under a given planning period (e.g., 5 years). Note that finite charging
station 𝑖 can be calculated by requests will be fulfilled at each charging station. Therefore, there
would be no queue if a sufficiently large number chargers are provided.
𝑡arr
𝑞,𝑖 = 𝑡𝑞 + 𝑠𝑞,𝑖 ∕𝑣,
̄ ∀ 𝑞, 𝑖,
In turn, building more chargers means more investment. Thus, we here
where 𝑡𝑞 is the time when the charging request 𝑞 is generated and 𝑣̄ is formulate a weighted sum optimization model to seek such trade-off
the average velocity of ETs on road network. for each charging station:
The SoC when ET 𝑣 (mapping to charging request 𝑞) arrives at the 𝑇plan
charging station 𝑖 could be updated by (Charging capacity model) min 𝛼𝐶𝐶(𝑦𝑖 ) + (1 − 𝛼) 𝜏𝐴𝑇 𝑊 𝑇 (𝑦𝑖 )
𝑦𝑖 𝑇
( )
SoC𝑞,𝑖 = 𝜗 − 𝑠𝑞,𝑖 ∕ 𝑅𝑘𝑣 (1 − 𝜌𝑣,𝑗 ) , ∀ 𝑞, 𝑖, (14)
s.t.
and thus the amount of energy (measured in SoC) to be replenished is
𝜗̄ − SoC𝑞,𝑖 . 𝑦𝑖 ∈ N ∶= {0, 1, 2, …}, (15)
𝑘
Let 𝑞𝑣,𝑡,𝑖,𝑗 be the charging request of ET 𝑣, 𝑣 ∈ 𝑘 at time 𝑡 to 𝑦𝑖 ≤ 𝑦̄𝑖 , (16)
be fulfilled by charging station 𝑖 at the 𝑗th sample of battery degra-
dation. Let 𝑖,𝑗 be the set of charging demand allocated to charg- where 𝑦𝑖 is the decision variable (the number of fast chargers at charg-
ing
( station 𝑖 at the 𝑗th sample ) of battery degradation, i.e., 𝑖,𝑗 ∶= ing station 𝑖). It is defined as a nonnegative integer by Constraint (15).
𝑘
𝑞𝑣,𝑡,𝑖,𝑗 , ∀ 𝑘 ∈ , 𝑡 ∈ , 𝑣 ∈ 𝑘 . We can thus obtain the arrival pattern Constraint (16) specifies a cap of the number of chargers (𝑦̄𝑖 ) that can
(arrival sequence in time series) of charging requests at charging station be deployed at charging station 𝑖. Coefficient 𝛼, 𝛼 ∈ [0, 1] is a weighted
𝑖. Fig. 2 gives an illustrative example of arrival pattern of charging parameter and 𝑇plan is the given period of charging infrastructure plan-
requests at charging station 𝑖. ning (e.g., 5 years) in line with agencies’ planning roadmap. Coefficient
𝜏 is the value of time ($/min) which is used to transform the second
4.2.3. Charging capacity optimization model objective of the ATWT into an equivalent monetary item. By this way,
We in the third step propose an optimization model to determine the the two objective components could be measured in the same unit and
optimal number of fast chargers (charging capacity) for each charging time dimension. Function 𝐶𝐶(⋅) is used to calculate the investment of
station. The optimization objective includes the average total waiting establishing a fast charging station,
time (ATWT) accumulated during a given planning period and total
construction cost. 𝐶𝐶(𝑦𝑖 ) = 𝑦𝑖 ∗ 𝐸𝐼𝐶 + 𝑛year ∗ 𝐴𝑂𝐶 ∗ 𝑦𝑖 , (17)
(1) 𝐷∕𝐷∕𝑠 queuing system with multiple chargers where 𝐸𝐼𝐶 is the construction cost (equipment and installation cost)
We first look at the calculation of the ATWT. The charging service to build a fast charger and 𝐴𝑂𝐶 is the annual operational cost (for
process at each charging station is characterized by a 𝐷∕𝐷∕𝑠 queuing maintenance and infrastructure repair) for one charger; 𝑛year is the
system with multiple chargers (servers), where the first-come-first- planning year.
served principle is applied. Without loss of generality, the charging Note that the maximal number of chargers 𝑦̄𝑖 is an finite integer
requests 𝑖,𝑗 can be sorted in the order of arriving time. To facilitate and can be set as the number of parking lots at charging station 𝑖.
the narrative, the subscripts 𝑖 and 𝑗 are dropped hereafter, and the Since there exist limit feasible solutions subject to nonnegative integer
following analysis focuses on queuing process at charging station 𝑖 at condition and the bound of 𝑦̄𝑖 , it is not difficult to solve the above
the 𝑗th sample of battery degradation unless otherwise stated. developed model (14) by using the enumeration method.
Let 𝐴(𝑡) be the accumulative arrivals of ET charging requests by time It is worth pointing out that, in Module 2, once the charging demand
𝑡, 𝐷(𝑡) be the cumulative number to enter for charging service by time 𝑡,
assigned to each candidate charging site is determined, the optimiza-
and 𝐷∗ (𝑡) be the cumulative departures to have left the charging servers
tion model is applied to minimize the AWAT and construction cost by
by time 𝑡. As shown in Fig. 3, the total waiting time in queue is the area
determining an optimal charger number at each station. That means,
bounded by 𝐴(𝑡) and 𝐷(𝑡) and the total waiting time in system is the
the network-level charging infrastructure planning problem is decom-
area bounded by the two curves of 𝐴(𝑡) and 𝐷∗ (𝑡). The average time in
posed into hundreds of station-based sub-problems of charger number
charging queue for ET drivers 𝚥 + 1 to 𝚥 + 𝑛 inclusive is given by
optimization with given charging requests. Using this approach, the
1 ∑ ′
𝜅=𝚥+𝑛
computational intractability mentioned in Section 3 is greatly reduced
𝑤queue ≡ (𝑡 − 𝑡𝜅 ), (11)
𝑛 𝜅=𝚥+1 𝜅 and the ET–FCFD problem becomes solvable.
7
H. Wang et al. Applied Energy 286 (2021) 116515
Fig. 3. Graphical representation of cumulative arrivals and departures from a charging queue.
4.3. Refinement of ET charging facility deployment Step 1: Examine if the charger numbers of all selected charging sta-
tions (denoted by a set of 𝛱) satisfy the pre-determined
After implementing Module 2 with given candidate charging sites, requirement of minimal charger number; if yes, stop and
we can obtain the FCFD result for the ET system. However, some report the CFD result; otherwise, filter out the unsuitable ones
charging stations may only have very few chargers, such as one charger. (denoted by a set of 𝛱 ′ ) and go to Step 2.
Such situation means an unattractive scale of economics because mul- Step 2: Iteratively refine the charging locations based on backward
tiple chargers often share a unified electrical grid system and the elimination and coverage rate examination.
same-purpose land use. Almost all stakeholders including ET operators,
government agencies, urban planning sectors and ET drivers would like Step 2.1: Choose one element from 𝜋 ∈ 𝛱 ′ and remove it from
to enjoy the scale of economics of a certain degree of integrate charging the set of 𝛱, i.e., let 𝛱 ∶= 𝛱∖{𝜋}.
sites so as to save land use and construction cost, reduce management Step 2.2: Given the re-defined set of charging sites 𝛱 and
challenges and maintenance cost, and enhance convenience and level examine the coverage rate by using Step 4 in Module
of charging service.4 1. If the coverage rate is satisfactory, then do Module
To make a more reasonable and economic planning of charging 2 again and go to Step 1; otherwise, let 𝛱 ′ ∶=
infrastructure, we propose the Module 3 to refine the FCFD via a 𝛱 ′ ∖{𝜋} and choose another element from 𝜋 ∈ 𝛱 ′
backward elimination method to ensure a minimum number of fast and go to Step 2.1 .
chargers (e.g., 2 chargers) at every station. The specific procedure for Step 2.3: Stop criterion: if 𝛱 ′ = 𝜙, stop and report the result
the refined module is elaborated below: of refined FCFD scheme.
8
H. Wang et al. Applied Energy 286 (2021) 116515
9
H. Wang et al. Applied Energy 286 (2021) 116515
Table 1
The parameter setting for the case study.
Parameter Value/Setting
Taxi fleet 20,434 operating taxis in the dataset
ET fleet 6130 ≈ 30% (ET penetration rate) × 20, 434
Candidate charging sites 185 petrol stations across the island
Refinement principle At least 2 chargers for each charging station
ET groups Group No. Range (km) Battery size (kWh) Market share (%)
Group 1 143 25.5 16
Group 2 183 28 22
Group 3 200 33.2 25
Group 4 243 40 12
Group 5 354 50 8
Group 6 383 60 6
Group 7 415 64 5
Group 8 465 75 4
Group 9 500 80 2
Other parameters: Planning period 𝑛year = 5 years
Sample size 1000
Fast charging rate 1.2C
Weighted parameter 𝛼 = 0.3
Value of time 𝜏 = 0.5 S$/min
Unit construction costa 50,672 S$/charger
Unit annual operational costa 617 S$/charger
Required coverage rate (service radius of 2 km) 85%
a The unit costs are obtained from a technical report of Land Transport Authority in Singapore [42].
Table 2 more than 32 charging sites with only one charger under initial scheme
The optimized charging facility deployment results.
(solution without refinement) will be removed from the consideration
System indicators Solution without Refined solution of charging infrastructure planning. We here also compare the coverage
refinement
rate of charging service before and after implementing refinement. The
Deployment No. of stations 183 151
coverage rate with service radius of 2 km slightly drops from 88.7% to
indicators Total charger No. 808 786
Maximal charger 30 30 88.5% if the refinement technique is applied. We then focus on the two
No. at one station performance indicators. Through the refinement,
( the ATWT (measured)
Performance Coverage rate 88.7% 88.5% in daily waiting time) is reduced by 7.95% 7.95% = 297.808−274.135 and
indicators ATWT (h/day) 297.808 274.135
( 297.808 )
43.508−42.324
TCC (Million S$) 43.508 42.324
the TCC is saved by about 1.2 million S$ 2.72% = 43.508
. In
summary, the refinement helps to save investment and improve system
performance in reducing the ATWT. This is because that the refinement
yields a more aggregate FCFD scheme in which the utilization rate of
island are considered as the candidate fast charging points for ETs. To the charging infrastructure would be enhanced, compared to the initial
give a straightforward presentation, we plot the spatial distribution of scheme obtained by solving Module 2 once.
charging demand allocation in Fig. 6.
5.4. The importance of incorporating vehicle heterogeneity and battery
5.3. Fast charging facility deployment scheme for the ET system degradation
10
H. Wang et al. Applied Energy 286 (2021) 116515
( )
786−718
the MEG-NBD case where multi-type ETs are addressed, 149 charging about 8.65% = . Accordingly, the ATWT and TCC are underes-
786 ( )
stations will be built and 718 fast chargers will be deployed across timated by 13.30% 13.30% = 274.135−237.666 and 8.65%
the island. As a result, the ATWT is 237.666 hours per day and TCC ( ) 274.135
42.324−38.662
is 38.662 million S$. If the vehicle heterogeneity is not taken into 8.65% = 42.324
. A similar trend can be obtained by the compar-
ison analysis between the SEC-NBD case and SEG-BD case. It implies
account, the number of charging stations and the total number of fast
that the FCFD plan will be also inaccurately determined if the bat-
chargers will, respectively, decrease to 142 and 681, and the ATWT
tery degradation is not considered. This is because that the charging
and TCC decline by 7.43% and 5.15% respectively. The lesson we learn
demand would be greatly underestimated without considering bat-
here is that overlooking vehicle heterogeneity will lead to a biased
tery degradation. As a result, fewer chargers and less investment are
planning of charging infrastructure and the system performance will required.
be optimistically overestimated. The same conclusion could be drawn We also gauge the gap if both factors are not addressed by compar-
and validated based on a comparison analysis between the SEG-BD case ing the MEG-BD case and SEG-NBD ( case. For the) FCFD results, the num-
and the MEG-BD case, although the degrees of overestimation in terms ber of stations is reduced by 6% 6% = 151−142 and the total number of
of the ATWT and TCC may be slight different, about 5.20% and 5.85% (151 )
chargers is underestimated by 13.36% 13.36% = 786−681 786
. Regarding
respectively.
the system performance,
( ) the estimation errors are more than 19.74%
We then move to explore the influence of battery degradation by
19.74% = 274.135−220.009 of the ATWT and 13.17%
comparing the results between the MEG-NBD case and MEG-BD case. ( 274.135 )
42.234−36.67
Under the MEG-BD case, 151 charging stations and 786 chargers will be 13.17% = 42.234
of the TCC. The absolute estimation error of
deployed, and the ATWT and TCC are respectively 274.135 hours/day the total construction cost even achieves 5.654 million S$. Based
and 42.324 million S$. When the battery degradation is neglected, on the above discussions, it can be concluded that a biased charging
the total number of faster chargers deployed sharply decreases to 718 infrastructure planning will be made if either factor is overlooked and
11
H. Wang et al. Applied Energy 286 (2021) 116515
Fig. 6. Spatial distribution of charging demand across 185 charging points (petrol stations).
12
H. Wang et al. Applied Energy 286 (2021) 116515
Table 3
Comparison of charging facility deployment results (refined solutions) among the four cases.
Tested case SEG-NBD case MEG-NBD case SEG-BD case MEG-BD case
No. of stations 142 149 147 151
System indicator Total charger No. 681 718 740 786
Maximal charger No. at one station 31 27 29 30
ATWT (h/day) 220.009 237.666 259.879 274.135
TCC (Million S$) 36.670 38.662 39.847 42.324
the system performance will be remarkably overestimated especially Declaration of competing interest
when both factors are not taken into account.
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared to
6. Concluding remarks
influence the work reported in this paper.
13
H. Wang et al. Applied Energy 286 (2021) 116515
[19] Yang W. A user-choice model for locating congested fast charging stations. Transp [31] Jochem P, Szimba E, Reuter-Oppermann M. How many fast-charging stations do
Res E 2018;110:189–213. we need along European highways?. Transp Res D 2019;73:120–9.
[20] Xylia M, Leduc S, Patrizio P, Kraxner F, Silveira S. Locating charging [32] Huang P, Ma Z, Xiao L, Sun Y. Geographic information system-assisted optimal
infrastructure for electric buses in stockholm. Transp Res C 2017;78:183–200. design of renewable powered electric vehicle charging stations in high-density
[21] Hung DQ, Dong ZY, Trinh H. Determining the size of PHEV charging stations cities. Appl Energy 2019;255:113855.
powered by commercial grid-integrated PV systems considering reactive power [33] Dong G, Ma J, Wei R, Haycox J. Electric vehicle charging point placement
support. Appl Energy 2016;183:160–9. optimisation by exploiting spatial statistics and maximal coverage location
[22] Lee C, Han J. Benders-and-price approach for electric vehicle charging models. Transp Res D 2019;67:77–88.
station location problem under probabilistic travel range. Transp Res B [34] Wang H, Zhao D, Meng Q, Ong GP, Lee D-H. A four-step method for electric-
2017;106:130–52. vehicle charging facility deployment in a dense city: An empirical study in
[23] Efthymiou D, Chrysostomou K, Morfoulaki M, Aifantopoulou G. Electric vehicles Singapore. Transp Res A 2019;119:224–37.
charging infrastructure location: a genetic algorithm approach. Eur Transp Res [35] Guo S, Zhao H. Optimal site selection of electric vehicle charging station
Rev 2017;9(2):27. by using fuzzy TOPSIS based on sustainability perspective. Appl Energy
[24] Vazifeh MM, Zhang H, Santi P, Ratti C. Optimizing the deployment of elec- 2015;158:390–402.
tric vehicle charging stations using pervasive mobility data. Transp Res A [36] Xu M, Meng Q, Liu K, Yamamoto T. Joint charging mode and location choice
2019;121:75–91. model for battery electric vehicle users. Transp Res B 2017;103:68–86.
[25] Shahraki N, Cai H, Turkay M, Xu M. Optimal locations of electric public charging [37] Muratori M, Elgqvist E, Cutler D, Eichman J, Salisbury S, Fuller Z, Smart J.
stations using real world vehicle travel patterns. Transp Res D 2015;41:165–76. Technology solutions to mitigate electricity cost for electric vehicle DC fast
[26] Kara EC, Macdonald JS, Black D, Bérges M, Hug G, Kiliccote S. Estimating charging. Appl Energy 2019;242:415–23.
the benefits of electric vehicle smart charging at non-residential locations: A [38] Pelletier S, Jabali O, Laporte G, Veneroni M. Battery degradation and behaviour
data-driven approach. Appl Energy 2015;155:515–25. for electric vehicles: Review and numerical analyses of several models. Transp
[27] He SY, Kuo Y-H, Wu D. Incorporating institutional and spatial factors in the Res B 2017;103:158–87.
selection of the optimal locations of public electric vehicle charging facilities: A [39] Yang F, Xie Y, Deng Y, Yuan C. Predictive modeling of battery degradation and
case study of Beijing, China. Transp Res C 2016;67:131–48. greenhouse gas emissions from US state-level electric vehicle operation. Nature
[28] Andrenacci N, Ragona R, Valenti G. A demand-side approach to the optimal Commun 2018;9(1):2429.
deployment of electric vehicle charging stations in metropolitan areas. Appl [40] Wang H, Zhao D, Cai Y, Meng Q, Ong GP. A trajectory-based energy consumption
Energy 2016;182:39–46. estimation method considering battery degradation for an urban electric vehicle
[29] Baresch M, Moser S. Allocation of e-car charging: Assessing the utilization of network. Transp Res D 2019;74:142–53.
charging infrastructures by location. Transp Res A 2019;124:388–95. [41] Wang H, Zhao D, Meng Q, Ong GP, Lee D-H. Network-level energy consumption
[30] Grote M, Preston J, Cherrett T, Tuck N. Locating residential on-street elec- estimation for electric vehicles considering vehicle and user heterogeneity.
tric vehicle charging infrastructure: A practical methodology. Transp Res D Transp Res A 2020;132:30–46.
2019;74:15–27. [42] ERI@N. E-MOBILITY. Tech. rep., Energy Research Institute @ Nanyang
Technological University (ERI@N); 2015.
14