Transportation Research Part D 130 (2024) 104178

Contents lists available at ScienceDirect

Transportation Research Part D

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

Allocate electric vehicles’ public charging stations with charging

demand uncertainty
Ting Wu a , Emily Fainman b ,∗, Yasmina Maïzi c , Jia Shu d , Yongzhen Li e
a Department of Mathematics, Nanjing University, 22 Hankou Road, Nanjing, 210093, Jiangsu, China
b McCoy College of Business, Texas State University, 601 University Drive, San Marcos, 78666, TX, USA
c
École des Sciences de la Gestion, Université du Québec à Montréal, 315 St Catherine St E, Montreal, H2X 3X2, Quebec, Canada
d
School of Management and Economics, University of Electronic Science and Technology of China, 2006 Xiyuan
Avenue, Chengdu, 611731, Sichuan, China
e
Department of Management Science and Engineering, School of Economics and Management, Southeast University, 2 Sipai
Road, Nanjing, 210096, Jiangsu, China

ARTICLE INFO ABSTRACT

Keywords: Nowadays, electric vehicles (EVs) have developed rapidly due to their great advantages in
Robust optimization fuel savings, reduced greenhouse gas emissions, and low air pollution. However, charging
Allocation problem imbalances and shortages have limited the prevalence of EVs. We optimize the siting and sizing
Electric vehicle charging infrastructure
of public charging stations under uncertain urban en-route recharging demand. Considering
Discrete-event simulation
charging congestion, we introduce the Erlang’s loss formula to represent the service level
requirement of each charging station. First, we determine the optimal locations and sizes
of those stations by a robust optimization model, where we figure out its deterministic
dual following linear approximation of the loss rate in queueing models. We further use a
discrete-event simulation to relax the assumption of time-independent charging demand in the
optimization approach and model real-time traffic based on real-world traffic and power grid
information in Nanjing, China. Thus, we verify the robust optimum outperforming deterministic
and stochastic optimums.

1. Introduction

The global transition to electric vehicles (EVs) represents a pivotal shift toward sustainable transportation, promising reduced
emissions and decreased dependence on fossil fuels. However, despite significant advancements in battery technology and the
proliferation of financial incentives, EV adoption remains in its nascent stages, characterized by a limited number of users worldwide.
Amidst discussions surrounding factors influencing EV prevalence, such as government subsidies and technological innovations, the
significance of public charging infrastructure often remains underestimated. For instance, recent studies by Hall and Lutsey (2017)
and Islam et al. (2018) shed light on the pivotal role of public charging infrastructure in shaping EV awareness and instilling
confidence in potential users. Currently, two primary en-route charging options exist: battery swap stations (BSS) – allowing users
to exchange exhausted batteries with fully charged ones – and public EV charging infrastructure. While BSS confronts challenges
such as battery standardization, infrastructure preparedness, and financial investment requirement (Setiawan et al., 2023), several
leading EV manufacturers to abstain from adopting this option for now,1 the need for accessible and reliable public EV charging
infrastructure becomes increasingly imperative.

∗ Corresponding author.
E-mail address: [email protected] (E. Fainman).
1 See https://www.teslarati.com/tesla-shuts-down-battery-swap-program-for-superchargers/ (retrieved on 2019-02-26).

https://doi.org/10.1016/j.trd.2024.104178
Received 9 August 2023; Received in revised form 13 February 2024; Accepted 19 March 2024
Available online 6 April 2024
1361-9209/© 2024 Elsevier Ltd. All rights reserved.
T. Wu et al. Transportation Research Part D 130 (2024) 104178

To address the diverse needs of EV users, charging infrastructure is typically categorized into three tiers: level 1, level 2, and
direct current fast chargers (DCFCs) (Union of Concerned Scientists, 2018). While level-1 chargers primarily cater to household use,
level-2 and DCFCs are deployed in public spaces, with DCFCs offering faster charging options yet 10–15 times more expensive than
level-2 chargers (Habib et al., 2017). Capable of replenishing EV batteries under 30 minutes, DCFCs are particularly sought after for
en-route charging, serving long-distance drivers and individuals without overnight or workplace charging access. Drivers prioritize
convenient access to DCFCs to minimize charging time during their journeys. Moreover, public DCFCs have garnered increasing
national policy and financial support in major Chinese cities.2 Therefore, we focus on optimizing the deployment of DCFCs, stations,
and associated electricity grids for direct plug-in charging equipment in public spaces such as parking lots of shopping malls, railway
stations, and streets.
The substantial installation costs of DCFCs pose challenges for widespread deployment, highlighting the importance of optimizing
their allocation at strategic locations to maximize charging infrastructure utility. Utilization rates of current chargers vary
significantly among Chinese cities, with some areas reaching up to 70%, while others exhibit extremely low rates in the same region.3
This underscores the need for appropriate allocation. A critical challenge in optimizing DCFC deployment is accurately estimating
charging demand within specific regions. Without clear projections of future EV driver populations or comprehensive insights into
charging patterns and user behavior, determining the appropriate number and locations of DCFCs is daunting. To address this gap,
we propose a model for a robust network of public EV charging stations with DCFCs. Our model considers uncertain charging
demand estimated from real-world data on traffic flow density, as well as time-varying traffic patterns influenced by weather and
seasons.
Our study, despite of using data from China, can apply to any type of EV which can be externally charged, including pure EVs
and plug-in hybrid vehicles. In particular, we aim to minimize the total charging associated cost by allocating optimal numbers
of DCFCs at the optimal locations to fulfill a high level of charging demand. The associated cost under consideration includes
upfront investment of power transmission lines and DCFCs, as well as the value of travel times to the nearest charging stations
throughout the decision period. Because travelers’ charging demands are random and can vary across time, we first capture the
uncertain charging demands with a robust optimization model. To account for potential waiting times at fully occupied charging
stations, we model the utilization rate of a charging station – the percentage of times chargers are utilized – using a 𝑀∕𝐺∕𝑚∕𝑚
queueing model, where each charging station has a specific number of chargers, and charging times follow a general distribution.
However, as the resulting utilization rate follows the Erlang-B formula or Erlang’s loss formula (Shortle et al., 2018), which renders
the optimization problem intractable, we establish a linear relationship between a queue’s arrival rate and the number of servers.
This approach enables the formulation of a linear robust optimization model. After analyzing the deterministic dual of the robust
optimization problem, we are able to determine the number of chargers at certain candidate stations. Next, we verify the robust
optimum in realistic traffic and power networks with a comprehensive simulation approach. In the simulation model, we relax the
assumption of time-independent charging demand in the optimization approach and model real-time traffic based on real-world
traffic and power grid information in Nanjing, China. We contribute to the EV charging siting and sizing literature with (1) a
robust optimization framework to accommodate uncertain urban en-route recharging demand, (2) a linear approximation of Erlang-B
formula, which allows to reformulate the robust optimization problem for solvers, and (3) a comprehensive simulation model of
realistic time-dependent charging demand to examine the effects of our proposed public charging stations.
The remainder of this paper is structured as follows. In Section 2, we review relevant literature with this study. In Section 3,
we introduce the robust model, followed by the approximated linearization of loss rate constraint and its equivalent reformulation.
In Section 4, we present a case study, including the comparison with different optimization models, sensitive analysis, and the
simulation based on real data. Finally, we conclude this paper by discussing limitations and future research directions in Section 5.

2. Relevant literature

Our work is closely relevant with literature focused on estimating EV charging demand and allocating EV charging facilities.
Specifically, we draw connections to studies using robust optimization, queueing models, and simulation approaches to address
siting and sizing of EV charging infrastructure.
It is challenging to quantify EV users’ charging demand or gauge its distribution, before EVs are widely prevalent. Several studies
consider influential geographic and macro economic factors in estimating charging demand in specific regions, such as Arias and Bae
(2016), He et al. (2016), and Wang et al. (2019). Other studies charging demand estimation by stratifying user types (Berkelmans
et al., 2018), clustering users based on time periods (Wolbertus et al., 2018), applying Markov transition matrix based on historical
data (Arias et al., 2017), and leveraging informatics (Yi et al., 2022). Other researches take different approaches by gauging charging
demand directly in specific settings, such as parking lots (Cai et al., 2014) and highways (Bae and Kwasinski, 2011). In this study,
we directly estimate expected value and variability of en-route recharging demand based on traffic flows in the network. This is
because we study the charging infrastructure and its operations in a specific transportation network, macro-level factor models and
their resulting demand in one region do not apply to this work.
The problem of locating EV charging stations has been widely studied in the literature. We refer to Shen et al. (2019), Kchaou-
Boujelben (2021), and Kchaou-Boujelben (2021) for extensive reviews on the planning of EV charging infrastructure. Recent works

2 See http://www.gov.cn/zhengce/zhengceku/2020-04/23/content_5505502.htm (retrieved on 2021-03-15).

3 See http://www.nrdc.cn/Public/uploads/2019-04-20/5cbb125a31059.pdf (retrieved on 2021-03-15).

2
T. Wu et al. Transportation Research Part D 130 (2024) 104178

further study the planning of charging stations under various emerging applications, such as the power-grid expansion (Quddus et al.,
2021), locating charging stations for the shared EV (Brandstätter et al., 2020), or a charging system with valet charging service (Li
et al., 2021). Different from many EV charging facility allocation problems deal with long-distance, inter-city trips (e.g. Kınay et al.,
2021; Li et al., 2022; Yılmaz and Yagmahan, 2022), electric taxi fleets (e.g. Cilio and Babacan, 2021), and existing bus systems
(e.g. Azadeh et al., 2022), this paper focuses on locating charging stations within a city region, explicitly considers uncertain charging
demands, and carefully captures the charging time based on queue theory.
Unlike conventional gas stations, EV charging stations subject to congestion due to limited capacity and substantial charging
times. Despite queueing models can capture the randomness in charging demand and the charging process, only a few works
consider charging congestion in locating EV charging stations using queueing models. Notably, several works (e.g., Jung et al.,
2014; Ghamami et al., 2016) model the queueing delay as nonlinear waiting time cost components in objective functions and apply
heuristics to solve the corresponding mixed integer nonlinear model. Others (e.g., Yang et al., 2017; Xie et al., 2018; Yang, 2018)
model charging congestion as service level constraints, which are then linearized approximately based on the queueing theory.
Besides diffusion and other traditional approximation methods for 𝑀∕𝐺∕𝑚∕𝑚 queues (e.g., Hokstad, 1978; Kimura, 1983; Breuer,
2008), recent works propose reflected Brownian motion (Yao, 1985; Khazaei et al., 2011) and a new steady-state probability
approximation method based on Laplace transform (Khazaei et al., 2011, 2012). Despite these methods are sufficiently accurate
with huge numbers of servers and overdispersed service times (the coefficient of variation of service time is greater than 1), they
are intractable for our robust optimization model. In this study, we contribute to the queueing theory with an approximately linear
relationship between loss rates and numbers of servers.
The exogenous and uncertain charging demand set our work relevant with robust optimization, despite a very limited number of
works study the EV charging station location problems based on robust optimization. While Hosseini and Sarder (2019) develop a
robust scenario-based model to handle uncertain EV driving range and traffic flows, works on robust queueing theory (e.g. Bertsimas
et al., 2011; Bandi et al., 2015; Whitt and You, 2018) model the uncertainty in the arrivals and services via polynomial uncertainty
sets, rather than assuming specific distributions. Under uncertain charging demand, other works employ robust models to optimize
the infrastructure planning of EV BSS (Mak et al., 2013), to simultaneously allocate charging stations and route EVs (Schiffer and
Walther, 2018), to make location-inventory decisions with electricity prices (Sun et al., 2019), and to minimize the worst-case
total travel cost and unused capacity penalty (Pourgholamali et al., 2023). We follow the framework of Bertsimas and Sim (2004)
and Bertsimas and Thiele (2006) to construct and solve the deterministic dual of the original robust optimization.
Our work also relates to simulation studies of traffic flows in transportation networks. Although recent efforts develop highly
computational and efficient simulation models for urban transportation systems (e.g., Basile et al., 2012; Thomin et al., 2013;
Shanmukhappa et al., 2018), the application of simulations to EVs remains relatively limited. Notable studies include Mallig
et al. (2016), who simulate EV charging demands in Greater Stuttgart under various market penetration scenarios, and Li et al.
(2016), utilizing Beijing taxi trajectory data for a time series-based simulation model to determine charging station locations.
Additionally, Marmaras et al. (2017) introduces a comprehensive simulation analyzing EV driver behaviors in transportation and
power networks. Some works employ simulation-based optimization models, capturing EV driver travel patterns, charging activity,
and evaluating performance measures for location decisions (e.g. Xi et al., 2013; Dong et al., 2014). In contrast to existing simulation
works of EVs, our simulation approach considers time-dependent traffic flows of an urban network and the user impatience,
which leads to demand loss at fully occupied charging stations. The simulation model validates optimal decisions obtained from
a robust optimization model by relaxing the time-independent assumption. Interested readers can refer to our recent conference
proceeding (Maïzi et al., 2019) for extensive scenarios and further discussion on simulation-based approaches.

3. Robust optimization model

In Fig. 1, we illustrate the structure and organizational framework of both the methods (Section 3) and the case study (Section 4).
In this section, we formulate a robust optimization problem to determine the siting and sizing of public charging stations, so
that a certain percentage of EV charging demand on local roads can be satisfied (Section 3.1). Following the proposition of an
approximately linear relationship between loss rates and numbers of servers in a 𝑀∕𝐺∕𝑚∕𝑚 queue (Section 3.2), we can derive a
deterministic equivalent reformulation of the robust optimization problem (Section 3.3). Therefore, we can solve the reformulated
mixed integer linear program with commercial solvers. Finally, we formulate a stochastic model as the benchmark to the robust
optimization model (Section 3.4). In addition, we present notations of our models in Table 1. Estimating parameters and sources
from which their values are derived will be presented in Section 4.2

3.1. Model formulation

We estimate the siting and sizing of public EV fast-charging stations by modeling each charging station as an 𝑀∕𝐺∕𝑚∕𝑚
queue. We extend the time-independent charging demand at charging stations to time-dependent arrivals in our simulation model
(Section 4.5). Here we focus on aggregated charging demand in a specific region; thus, we consider homogeneous EV drivers who
are assumed to follow the shortest-path principal and go to their nearest charging station for charging. Heterogeneous EV drivers
and their different charging choices are beyond the scope of this study.
Consider the problem of locating charging stations in a set 𝐼 of potential sites and setting up chargers in each station to satisfy
charging demand in a set 𝐽 of demand nodes. For each site 𝑖 ∈ 𝐼 and demand node 𝑗 ∈ 𝐽 , let 𝑌𝑖 ∈ {0, 1} be the decision of whether
a charging station is located in site 𝑖, 𝑚𝑖 ∈ Z+ be the number of chargers set up in station 𝑖, and 𝑋𝑖𝑗 ∈ {0, 1} be the decision of

3
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Fig. 1. The organization of the methods (Section 3) and the case study (Section 4).

whether station 𝑖 serves demand node 𝑗. Denote 𝑿, 𝒀 , and 𝒎 as the vectors of 𝑋𝑖𝑗 , 𝑌𝑖 , and 𝑚𝑖 , respectively. The problem aims to
optimize (𝑿, 𝒀 , 𝒎) to minimize the total cost and satisfy charging demand under certain service level requirement. For this purpose,
given (𝑿, 𝒀 , 𝒎), the charging system in each station 𝑖 is modeled by an 𝑀∕𝐺∕𝑚𝑖 ∕𝑚𝑖 queue. To be more specific, for each station 𝑖,
∑
customers (EV drivers) arrive following a Poisson process with arrival rate 𝜆𝑖 ∶= 𝑗∈𝐽 𝜔𝑗 𝑋𝑖𝑗 , where 𝜔𝑗 denotes the average charging
demand rate in demand node 𝑗. Station 𝑖 has 𝑚𝑖 servers (chargers). Because EVs may have varying depleted batteries at the time of
charging, we model this uncertain state-of-charge by considering servers’ service (charging) time following a generally-distributed
stochastic process with service rate 𝜇. When an EV driver arrives at a station, the EV gets charged immediately if any of chargers
is available, while the arriving drivers who find the system full leave immediately. In other words, waiting is not considered and
the capacity of station 𝑖 is 𝑚𝑖 . In this paper, we consider the service level requirement on the loss rate of charging demand. Denote
𝑓 (𝜆𝑖 , 𝜇, 𝑚𝑖 ) as the loss rate of the 𝑀∕𝐺∕𝑚𝑖 ∕𝑚𝑖 queue system, which is exactly the fraction of leaving customers who find the system
full. According to Shortle et al. (2018), the loss rate 𝑓 (𝜆𝑖 , 𝜇, 𝑚𝑖 ) is given by
(𝜆𝑖 ∕𝜇)𝑚𝑖
𝑚𝑖 !
𝑓 (𝜆𝑖 , 𝜇, 𝑚𝑖 ) ∶= ∑ . (1)
𝑚𝑖 (𝜆𝑖 ∕𝜇)𝑘
𝑘=0 𝑘!

Then, the service level requires the loss rate to be close to a preset value 𝛼, that is,

|𝑓 (𝜆𝑖 , 𝜇, 𝑚𝑖 ) − 𝛼| ≤ 𝜖

where 𝜖 is small but allows 𝑚𝑖 ∈ Z+ is feasible on the set of non-negative integers.

Suppose the average demand rate 𝜔𝑗 is exactly known, then this charging station location problem can be modeled as
∑ ∑ ∑ ∑
min 𝑐𝑚𝑖 + 𝑠𝑖 𝑌𝑖 + 𝜔𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗 (2a)
𝑖∈𝐼 𝑖∈𝐼 𝑗∈𝐽 𝑖∈𝐼

s.t. |𝑓 (𝜆𝑖 , 𝜇, 𝑚𝑖 ) − 𝛼| ≤ 𝜖, ∀𝑖 ∈ 𝐼, (2b)

∑
𝑋𝑖𝑗 = 1, ∀𝑗 ∈ 𝐽 , (2c)
𝑖∈𝐼

𝑋𝑖𝑗 ≤ 𝑌𝑖 , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 , (2d)
𝑋𝑖𝑗 , 𝑌𝑖 ∈ {0, 1}, 𝑚𝑖 ∈ Z+ , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 , (2e)
∑
The objective function (2a) minimizes the sum of three charging associated costs, namely the cost of setting up chargers 𝑖∈𝐼 𝑐𝑚𝑖 ,
∑
the fixed cost of setting up charging stations 𝑖∈𝐼 𝑠𝑖 𝑌𝑖 , and the monetary value for total travel time of driving to charging stations
∑ ∑
𝑗∈𝐽 𝜔𝑗 𝑖∈𝐼 𝑑𝑖𝑗 𝑋𝑖𝑗 during the time horizon of study. Constraint (2b) specifies the loss rate restriction, where the loss rate 𝑓 (𝜆𝑖 , 𝜇, 𝑚𝑖 )
is given as in (1). Constraint (2c) requires that each demand node be served by one station. Constraint (2d) states that only the
open stations can serve demand nodes.

4
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Table 1
Model notations.
General parameters
𝐼 Set of potential sites to set up charging stations.
𝐽 Set of demand nodes.
𝑠𝑖 Fixed cost of setting up a charging station in site 𝑖.
𝑐 Unit cost of setting up a charger in stations.
𝑀𝑖 Maximal number of chargers that can be deployed in station 𝑖.
𝑑𝑖𝑗 Time cost of traveling from demand node 𝑗 to station 𝑖 to charge a vehicle.
𝜔𝑗 Mean charging demand rate from node 𝑗.
𝜇 Average service rate of a charger.
𝛼 Service level requirement, where 𝛼 represents the maximal loss rate.
𝛽 Coefficient determined by the loss rate 𝛼.
Parameters in the queueing model (Section 3.2)
𝜆𝑖 Arrival rate of EVs needing charging at site 𝑖.
𝑓 (⋅) Loss rate function of a 𝑀∕𝐺∕𝑚∕𝑚 queue.
𝜆 Arrival rate of a 𝑀∕𝐺∕𝑚∕𝑚 queue.
𝑚 Number of servers of a 𝑀∕𝐺∕𝑚∕𝑚 queue.
Parameters in the robust optimization model (Section 3.3)
𝜔̃𝑗 The actual charging demand at node 𝑗.
𝜔̄𝑗 Nominal value of 𝜔̃𝑗 .
𝜔̂𝑗 Deviation in the support of 𝜔̃𝑗 .
𝝎
̃ = (𝜔̃ 𝑗 ∀𝑗 ∈ 𝐽 ), the vector of 𝜔̃ 𝑗 .
𝛺 ̃ the uncertainty set.
= {𝝎}
𝛤 Coefficient of uncertainty level.
𝑧𝑖𝑗 Auxiliary variables to represent the level of uncertainty.
𝑍 Auxiliary variable transforming a robust optimization to a deterministic optimization.
𝑢𝑗 Auxiliary variables transforming a robust optimization to a deterministic optimization.
𝑣 Auxiliary variable transforming a robust optimization to a deterministic optimization.
𝑝𝑖 Auxiliary variables transforming a robust optimization to a deterministic optimization.
𝑞𝑖𝑗 Auxiliary variables transforming a robust optimization to a deterministic optimization.
Parameters in the stochastic optimization (Section 3.4)
𝑆 Set of samples.
𝑝𝑠 Probability of sample 𝑠.
𝑷 Joint distribution of (𝜔̃ 𝑗 ∀𝑗 ∈ 𝐽 ).
𝜔𝑠𝑗 Charging demand at site 𝑗 in sample 𝑠.
Decision variables
𝑋𝑖𝑗 ∈ {0, 1}. 𝑋𝑖𝑗 = 1 if vehicles from demand node 𝑗 ∈ 𝐽 is charged in station
𝑖 ∈ 𝐼 and 𝑋𝑖𝑗 = 0 otherwise.
𝑌𝑖 ∈ {0, 1}. 𝑌𝑖 = 1 if a charging station is set up in potential location 𝑖 ∈ 𝐼
and 𝑌𝑖 = 0 otherwise.
𝑚𝑖 ∈ Z+ . The number of chargers in station 𝑖 ∈ 𝐼.

3.2. Linear approximation of loss rate constraint

Due to constraint (2b), model (2) is a complex nonlinear program and infeasible to solve (Tijms et al., 1981). To this end, we
have the following proposition for the 𝑀∕𝐺∕𝑚𝑖 ∕𝑚𝑖 queue, which is self-contained with the subscript 𝑖 dropped. Thus, we show that
constraint (2b) can be approximately linearized, which leads to tractable linear approximations of model (2).

Proposition 3.1. In the 𝑀∕𝐺∕𝑚∕𝑚 queue with arrival rate 𝜆 and service rate 𝜇, if the loss rate 𝑓 (𝜆, 𝜇, 𝑚) defined in (1) is constant and
𝑚 is sufficiently large, the number of servers 𝑚 is approximately proportional to the arrival rate 𝜆. In other words, if 𝑓 (𝜆, 𝜇, 𝑚) = 𝛼, the
number of servers 𝑚 is the smallest integer that is no less than 𝜆(1 − 𝛼)∕𝜇 as 𝑚 ⟶ ∞, i.e.
𝜆 𝜇
⟶ , as 𝑚 ⟶ ∞. (3)
𝑚 1−𝛼

Proof. According to Harel (1988), 𝑓 (𝜆, 𝜇, 𝑚) has the upper bound 𝑈 and lower bound 𝐿 such that 𝐿 ≤ 𝑓 (𝜆, 𝜇, 𝑚) ≤ 𝑈 , ∀𝜌, where 𝐿
and 𝑈 are
√
𝑚(1 − 𝜌)2 + 2𝜌 − (1 − 𝜌) 4𝑚𝜌 + 𝑚2 (1 − 𝜌)2
𝑈= √ ;
−𝑚𝜌(1 − 𝜌) + 2𝜌 + 𝜌 4𝑚𝜌 + 𝑚2 (1 − 𝜌)2
[ √ ]+
𝑚(1 − 𝜌)2 + 2𝜌 − (1 − 𝜌) 4𝑚𝜌 + 4(𝑚 + 1) + 𝑚2 (1 − 𝜌)2
𝐿= √ .
−𝑚𝜌(1 − 𝜌) + 2𝜌 + 𝜌 4𝑚𝜌 + 4(𝑚 + 1) + 𝑚2 (1 − 𝜌)2

5
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Fig. 2. Loss rate 𝑓 (𝜆, 𝜇, 𝑚) and the cross section with fixed loss rate.

Fig. 3. Linear relationship between ratio 𝜆∕𝜇 and the number of servers in an 𝑀∕𝐺∕𝑚∕𝑚 queue.

Thus, when 𝑚 becomes large,

√ { 𝜌−1
(1 − 𝜌)2 − (1 − 𝜌) (1 − 𝜌)2 𝜌
, if 𝜌 > 1;
lim 𝑈 =
𝑚⟶∞
√ =
−𝜌(1 − 𝜌) + 𝜌 (1 − 𝜌)2 0, if 0 < 𝜌 ≤ 1.
√ { 𝜌−1
(1 − 𝜌)2 − (1 − 𝜌) (1 − 𝜌)2 𝜌
, if 𝜌 > 1;
lim 𝐿 =
𝑚⟶∞
√ =
−𝜌(1 − 𝜌) + 𝜌 (1 − 𝜌)2 0, if 0 < 𝜌 ≤ 1.

Therefore, lim𝑚⟶∞ 𝑓 (𝜆, 𝜇, 𝑚) = (𝜌 − 1)∕𝜌 = 𝛼, we have

𝜆 1
𝜌= ⟶ , as 𝑚 ⟶ ∞.
𝑚𝜇 1−𝛼
Proposition 3.1 suggests that in a time-independent queue with Poisson arrivals and generally-distributed service times of
a certain service rate, the number of servers should change proportionally with the arrival rate to maintain the service level
requirement 𝛼. Please note that the size of the queue can be as small as ten servers to be in compliance with the proposition,
despite that the approximation of 𝜆∕𝑚 converges to the exact ratio of 𝜇 and 1 − 𝛼 as the number of servers increase in the queue.
Now we justify Proposition 3.1 in the following illustrative example.
Example. Consider a charging system following the 𝑀∕𝐺∕𝑚∕𝑚 queue with 𝜆 being the system arrival rate and 𝜇 being the
service rate of each charger. First, we fix 𝜇 and enumerate the loss rate 𝑓 (𝜆, 𝜇, 𝑚), which is a function of 𝜆 and 𝑚 of varied values
as shown in Fig. 2(a). We then plot the cross section with 𝑓 (𝜆, 𝜇, 𝑚) = 90% and 𝜖 = 0.001 as shown in Fig. 2(b), where 𝑅2 = 0.9735
in the linear regression demonstrates a sufficiently good fitness.
Keeping 𝜇 the same, we further plot the scatter chart of 𝑚 and 𝜆∕𝜇 for different loss rate 𝑓 (𝜆, 𝜇, 𝑚) in Fig. 3. Thus, we show the
linear relationship between the ratio 𝜆∕𝜇 and the number of servers 𝑚 in an 𝑀∕𝐺∕𝑚∕𝑚 queue. For reasonably large numbers of
servers and arrival rates, the slope of the line is equal to 1 − 𝛼, as shown in (3).

6
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Therefore, we can re-write the constraint (2b) as

∑
𝜔𝑗 𝑋𝑖𝑗 = 𝛽𝑚𝑖 , ∀𝑖 ∈ 𝐼,
𝑗∈𝐽
1
where 𝛽 ∶= 𝜇 1−𝛼 is the coefficient determined by a specific loss rate 𝛼 in Proposition 3.1. Thus, model (2) can be approximated by
∑ ∑ ∑ ∑
min 𝑐𝑚𝑖 + 𝑠𝑖 𝑌𝑖 + 𝜔𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗
𝑖∈𝐼 𝑖∈𝐼 𝑗∈𝐽 𝑖∈𝐼
∑
s.t. 𝜔𝑗 𝑋𝑖𝑗 = 𝛽𝑚𝑖 , ∀𝑖 ∈ 𝐼,
𝑗∈𝐽
∑ (4)
𝑋𝑖𝑗 ≥ 1, ∀𝑗 ∈ 𝐽 ,
𝑖∈𝐼
𝑋𝑖𝑗 ≤ 𝑌𝑖 , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 ,
𝑋𝑖𝑗 , 𝑌𝑖 ∈ {0, 1}, 𝑚𝑖 ∈ Z+ , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 ,

3.3. Robust model and equivalent reformulation

Section 3.2 implies that model (2) can be approximated by model (4) with satisfactory fitness. Therefore, we next consider
demand uncertainty in model (4), formulate the corresponding robust counterpart model, and derive the equivalent reformulation.
Consider the robust model with demand uncertainty, which is represented as 𝜔̃ 𝑗 and the corresponding vector is denoted by 𝝎.
̃
Suppose the uncertain demand 𝝎̃ lies in the following uncertainty set,
⎧ ||𝜔̃ 𝑗 = 𝜔̄ 𝑗 + 𝜔̂ 𝑗 𝑧𝑗 , ∀𝑗 ∈ 𝐽 , ⎫
⎪ || ⎪
̃ |0 ≤ |𝑧𝑗 | ≤ 1,
𝛺 ∶= ⎨𝝎 ∀𝑗 ∈ 𝐽 , ⎬
⎪ ||∑ ⎪
⎩ | 𝑗∈𝐽 |𝑧𝑗 | ≤ 𝛤 ⎭
where 𝜔̄ 𝑗 and [𝜔̄ 𝑗 − 𝜔̂ 𝑗 , 𝜔̄ 𝑗 + 𝜔̂ 𝑗 ] represent the nominal value and the support of 𝜔̃ 𝑗 , respectively. And 𝛤 indicates the level of
uncertainty for the charging demand. A large 𝛤 value allows more deviation in the charging demand 𝜔, compared with a small 𝛤 .
We study the impacts of 𝛤 on the resulting robust optimums and values of objective function in sensitivity analysis (Section 4.4).
Then the robust counterpart of model (4) is given by
∑ ∑ ∑ ∑
min 𝑐 𝑖 𝑚𝑖 + 𝑠𝑖 𝑌𝑖 + max 𝜔̃ 𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗 (5a)
𝝎∈𝛺
̃
𝑖∈𝐼 𝑖∈𝐼 𝑗∈𝐽 𝑖∈𝐼
∑
s.t. max 𝜔̃ 𝑗 𝑋𝑖𝑗 = 𝛽𝑚𝑖 , ∀𝑖 ∈ 𝐼, (5b)
𝝎∈𝛺
̃
𝑗∈𝐽
∑
𝑋𝑖𝑗 ≥ 1, ∀𝑗 ∈ 𝐽 , (5c)
𝑖∈𝐼

𝑋𝑖𝑗 ≤ 𝑌𝑖 , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 , (5d)
𝑋𝑖𝑗 , 𝑌𝑖 ∈ {0, 1}, 𝑚𝑖 ∈ Z+ , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 , (5e)
where the objective minimizes the cost of location and charger deployment and the worst-case travel time cost. The first and second
constraints guarantee the feasibility for the worst-case scenario. In this model, the location, the charger deployment, and the demand
allocation decisions are all made prior to any realization of uncertain data.
We next have the following theorem which shows that model (5) can be equivalently reformulated as a mixed integer linear
program.

Theorem 3.1. Model (5) can equivalently reformulated as

∑ ∑
min 𝑐𝑖 𝑚𝑖 + 𝑠𝑖 𝑌𝑖 + 𝑍
𝑖∈𝐼 𝑖∈𝐼
∑ ∑ ∑
s.t. 𝑍≥ 𝜔̄ 𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗 + 𝑣𝛤 + 𝑢𝑗 ,
𝑗∈𝐽 𝑖∈𝐼 𝑗∈𝐽
∑
𝑣 + 𝑢𝑗 ≥ 𝜔̂ 𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗 , ∀𝑗 ∈ 𝐽 ,
𝑖∈𝐼
∑ ∑
𝛽𝑚𝑖 ≥ 𝜔̄ 𝑗 𝑋𝑖𝑗 + 𝑝𝑖 𝛤 + 𝑞𝑖𝑗 , ∀𝑖 ∈ 𝐼,
𝑗∈𝐽 𝑗∈𝐽 (6)
𝑝𝑖 + 𝑞𝑖𝑗 ≥ 𝜔̂ 𝑗 𝑋𝑖𝑗 , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 ,
∑
𝑋𝑖𝑗 ≥ 1, ∀𝑗 ∈ 𝐽 ,
𝑖∈𝐼
𝑋𝑖𝑗 ≤ 𝑌𝑖 , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 ,
𝑋𝑖𝑗 , 𝑌𝑖 ∈ {0, 1}, 𝑚𝑖 ∈ Z+ , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 ,
𝑍, 𝑢𝑗 , 𝑣, 𝑝𝑖 , 𝑞𝑖𝑗 ≥ 0, ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 .

7
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Proof. First, the objective (5a) can be written as

∑ ∑
min 𝑐𝑖 𝑚𝑖 + 𝑠𝑖 𝑌𝑖 + 𝑍
𝑖∈𝐼 𝑖∈𝐼

where
∑ ∑
𝑍 ≥ max 𝜔̃ 𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗 (7)
𝝎∈𝛺
̃
𝑗∈𝐽 𝑖∈𝐼

Substituting the definition of uncertainty set 𝛺, constraint (7) can be written as

{ }
∑ ∑ |∑
| 𝑗∈𝐽 |𝑧𝑗 | ≤ 𝛤 ,
𝑍 ≥ max (𝜔̄ 𝑗 + 𝜔̂ 𝑗 𝑧𝑗 ) 𝑑𝑖𝑗 𝑋𝑖𝑗 |
𝑧𝑗 ∀𝑗∈𝐽 |0 ≤ |𝑧 | ≤ 1, ∀𝑗 ∈ 𝐽 ,
𝑗∈𝐽 𝑖∈𝐼 | 𝑗
{ }
∑ ∑ ∑ ∑ |∑
| 𝑗∈𝐽 |𝑧𝑗 | ≤ 𝛤 ,
= 𝜔̄ 𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗 + max 𝜔̂ 𝑗 𝑧𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗 |
𝑧𝑗 ,∀𝑗∈𝐽 |0 ≤ |𝑧 | ≤ 1, ∀𝑗 ∈ 𝐽 ,
𝑗∈𝐽 𝑖∈𝐼 𝑗∈𝐽 𝑖∈𝐼 | 𝑗

For the maximization subproblem as in the last term, we have

{ }
∑ ∑ |∑ |𝑧 | ≤ 𝛤 ,
|
max 𝜔̂ 𝑗 𝑧𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗 | 𝑗∈𝐽 𝑗
𝑧𝑗 ,∀𝑗∈𝐽 |0 ≤ |𝑧𝑗 | ≤ 1, ∀𝑗 ∈ 𝐽 ,
𝑗∈𝐽 𝑖∈𝐼 |
⎧ ∑ ⎫
| 𝑗∈𝐽 (𝑧+ + 𝑧− ) ≤ 𝛤 ,
⎪∑ ∑ | 𝑗 𝑗
⎪
+ − |𝑧+ + 𝑧− ≤ 1,
= max ⎨ 𝜔̂ 𝑗 (𝑧𝑗 − 𝑧𝑗 ) 𝑑𝑖𝑗 𝑋𝑖𝑗 | 𝑗 ∀𝑗 ∈ 𝐽 , ⎬
𝑧+ − ,∀𝑗∈𝐽 | + −𝑗
𝑗 𝑗,𝑧 ⎪𝑗∈𝐽 𝑖∈𝐼 |𝑧 , 𝑧 ≥ 0, ∀𝑗 ∈ 𝐽 ⎪
⎩ | 𝑗 𝑗 ⎭
⎧ |𝑣 + 𝑢 ≥ 𝜔̂ ∑ 𝑑 𝑋 ∀𝑗 ∈ 𝐽 , ⎫
⎪ ∑ || 𝑗 𝑗 ∑ 𝑖∈𝐼 𝑖𝑗 𝑖𝑗 ⎪
= min ⎨𝑣𝛤 + 𝑢𝑗 |𝑣 + 𝑢𝑗 ≥ −𝜔̂ 𝑗 𝑖∈𝐼 𝑑𝑖𝑗 𝑋𝑖𝑗 ∀𝑗 ∈ 𝐽 , ⎬
𝑢𝑗 ,𝑣,∀𝑗∈𝐽 |
⎪ 𝑗∈𝐽 |𝑣, 𝑢𝑗 ≥ 0, ∀𝑗 ∈ 𝐽 ⎪
⎩ | ⎭
{ ∑ }
∑ ||𝑣 + 𝑢 ≥ 𝜔̂ ∀𝑗 ∈ 𝐽 ,
= min 𝑣𝛤 + 𝑢𝑗 | 𝑗 𝑗 𝑖∈𝐼 𝑑𝑖𝑗 𝑋𝑖𝑗 ,
𝑢𝑗 ,𝑣,∀𝑗∈𝐽 |𝑣, 𝑢𝑗 ≥ 0, ∀𝑗 ∈ 𝐽
𝑗∈𝐽 |
where 𝑣 and 𝑢𝑗 denote the dual variables of the corresponding constraints, respectively. Thus, constraint (7) can be equivalently
written as
∑ ∑ ∑
𝑍 ≥ 𝑗∈𝐽 𝜔̄ 𝑗 𝑖∈𝐼 𝑑𝑖𝑗 𝑋𝑖𝑗 + 𝑣𝛤 + 𝑢𝑗 ,
∑ 𝑗∈𝐽
𝑣 + 𝑢𝑗 ≥ 𝜔̂ 𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗 , ∀𝑗 ∈ 𝐽 ,
𝑖∈𝐼
𝑣, 𝑢𝑗 ≥ 0, ∀𝑗 ∈ 𝐽 .
Similarly, constraint (5b) can be equivalently written as, for each 𝑖 ∈ 𝐼,
∑
𝛽𝑚𝑖 ≥ max 𝜔̃ 𝑗 𝑋𝑖𝑗
𝝎∈𝛺
̃
𝑗∈𝐽
{ }
∑ |∑
| 𝑗∈𝐽 |𝑧𝑗 | ≤ 𝛤 ,
= max (𝜔̄ 𝑗 + 𝜔̂ 𝑗 𝑧𝑗 )𝑋𝑖𝑗 |
𝑧𝑗 ,∀𝑗∈𝐽 |0 ≤ |𝑧𝑗 | ≤ 1, ∀𝑗 ∈ 𝐽 ,
𝑗∈𝐽 |
{ }
∑ ∑ |∑
| 𝑗∈𝐽 |𝑧𝑗 | ≤ 𝛤 ,
= 𝜔̄ 𝑗 𝑋𝑖𝑗 + max 𝜔̂ 𝑗 𝑧𝑗 𝑋𝑖𝑗 |
𝑧𝑗 ,∀𝑗∈𝐽 |0 ≤ |𝑧𝑗 | ≤ 1, ∀𝑗 ∈ 𝐽 ,
𝑗∈𝐽 𝑗∈𝐽 |
For the maximization subproblem as in the last term, we have
{ }
∑ |∑ |𝑧 | ≤ 𝛤 ,
|
max 𝜔̂ 𝑗 𝑧𝑗 𝑋𝑖𝑗 | 𝑗∈𝐽 𝑗
𝑧𝑗 ,∀𝑗∈𝐽 |0 ≤ |𝑧𝑗 | ≤ 1, ∀𝑗 ∈ 𝐽 ,
𝑗∈𝐽 |
⎧ ∑ ⎫
| 𝑗∈𝐽 (𝑧+ + 𝑧− ) ≤ 𝛤 ,
⎪∑ | 𝑗 𝑗
⎪
+ − |𝑧+ + 𝑧− ≤ 1,
= max ⎨ 𝜔̂ 𝑗 (𝑧𝑗 − 𝑧𝑗 )𝑋𝑖𝑗 | 𝑗 ∀𝑗 ∈ 𝐽 , ⎬
𝑧+ − | + −𝑗
𝑗 ,𝑧𝑗 ,∀𝑗∈𝐽 ⎪𝑗∈𝐽 |𝑧 , 𝑧 ≥ 0, ∀𝑗 ∈ 𝐽 , ⎪
⎩ | 𝑗 𝑗 ⎭
⎧ | ∀𝑗 ∈ 𝐽 , ⎫
⎪ ∑ ||𝑝𝑖 + 𝑞𝑖𝑗 ≥ 𝜔̂ 𝑗 𝑋𝑖𝑗 ⎪
= min ⎨ 𝑝𝑖 𝛤 + 𝑞𝑖𝑗 |𝑝𝑖 + 𝑞𝑖𝑗 ≥ −𝜔̂ 𝑗 𝑋𝑖𝑗 ∀𝑗 ∈ 𝐽 , ⎬
𝑝𝑖 ,𝑞𝑖𝑗 ∀𝑗∈𝐽 |
⎪ 𝑗∈𝐽 |𝑝𝑖 , 𝑞𝑖𝑗 ≥ 0, ∀𝑗 ∈ 𝐽 ⎪
⎩ | ⎭
{ }
∑ ||𝑝 + 𝑞 ≥ 𝜔̂ 𝑋 , ∀𝑗 ∈ 𝐽
= min 𝑝𝑖 𝛤 + 𝑞𝑖𝑗 | 𝑖 𝑖𝑗 𝑗 𝑖𝑗
𝑝𝑖 ,𝑞𝑖𝑗 ∀𝑗∈𝐽 |𝑝𝑖 , 𝑞𝑖𝑗 ≥ 0, ∀𝑗 ∈ 𝐽
𝑗∈𝐽 |

8
T. Wu et al. Transportation Research Part D 130 (2024) 104178

where 𝑝𝑖 and 𝑞𝑖𝑗 represent the dual variables for the corresponding constraints, respectively. Thus, constraint (5b) can be equivalently
written as
∑ ∑
𝛽𝑚𝑖 ≥ 𝜔̄ 𝑗 𝑋𝑖𝑗 + 𝑝𝑖 𝛤 + 𝑞𝑖𝑗 , ∀𝑖 ∈ 𝐼,
𝑗∈𝐽 𝑗∈𝐽

𝑝𝑖 + 𝑞𝑖𝑗 ≥ 𝜔̂ 𝑗 𝑋𝑖𝑗 , ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 ,
𝑝𝑖 , 𝑞𝑖𝑗 ≥ 0, ∀𝑖 ∈ 𝐼, 𝑗 ∈ 𝐽 .
To conclude, model (5) can be equivalently written as in (6), which completes the proof.

Note that model (6) is a mixed integer linear program, which is tractable by many commercial solvers like Gurobi and CPLEX. In
the following section, we solve model (6) by Gurobi and propose a simulation approach to evaluate the performance of the robust
model under a real data set.

3.4. Benchmark: A corresponding stochastic model

Let 𝑷 be the joint distribution of (𝜔̃ 𝑗 ∀𝑗 ∈ 𝐽 ), so we can derive the stochastic model corresponding to model (4) as
[ ]
∑ ∑ ∑ ∑
min 𝑐𝑚𝑖 + 𝑠𝑖 𝑌𝑖 + E𝑷 𝜔̃ 𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗
𝑖∈𝐼 𝑖∈𝐼 𝑗∈𝐽 𝑖∈𝐼
∑ (8)
s.t. 𝜔̃ 𝑗 𝑋𝑖𝑗 = 𝛽𝑚𝑖 , 𝑷 -a.s. ∀𝑖 ∈ 𝐼,
𝑗∈𝐽

(5c)−(5e).
where 𝑷 -a.s. in the first constraint implies that the constraint must be satisfied almost surely under the probability measure 𝑷 . We
next approximate model (8) by the Sample Average Approximation (SAA). For each sample 𝑠 = 1, … , 𝑆, denote 𝑝𝑠 as the probability
∑
of sample such that 𝑆𝑠=1 𝑝𝑠 = 1, and (𝜔𝑠𝑗 ∀𝑗 ∈ 𝐽 ) as the demand in this scenario. Then the SAA reformulation of model (4) is as
follows:
∑ ∑ ∑
𝑆 ∑ ∑
min 𝑐𝑚𝑖 + 𝑠𝑖 𝑌𝑖 + 𝑝𝑠 𝜔𝑠𝑗 𝑑𝑖𝑗 𝑋𝑖𝑗
𝑖∈𝐼 𝑖∈𝐼 𝑠=1 𝑗∈𝐽 𝑖∈𝐼
∑ (9)
s.t. 𝜔𝑠𝑗 𝑋𝑖𝑗 = 𝛽𝑚𝑖 , ∀𝑖 ∈ 𝐼, 𝑠 = 1, … , 𝑆,
𝑗∈𝐽

(5c)−(5e).

4. Case study

In this section, we perform a case study using traffic flow data and power network information in the district of Luhe (also
known as Liuhe). First, we describe our real traffic data in Section 4.1, followed by parameter estimation for the optimization
models in Section 4.2. After presenting optimal allocations of charging stations in Section 4.3, we perform sensitivity analysis of
varying uncertainties and charging demands in Section 4.4. Finally, we compare the optimums of our robust optimization model with
deterministic and stochastic optimization models in a simulation study (Section 4.5). The simulation study also relax the assumption
of time-independent charging demand in the optimization models.

4.1. Data description

Luhe, as shown in Fig. 4, is one metropolitan county of Nanjing, the capital of Jiangsu Province in Southeast China. With a
population of approximately one million inhabitants, it covers an area of 1481 square kilometers, boasting a population density
of 639 people per square kilometer4 —figures comparable to medium-sized U.S. cities such as Austin, TX, Nashville, TN, and
Jacksonville, FL.5 Serving as the northern gateway to Nanjing, the provincial capital, Luhe has historically been a local focal point for
economy, transportation, culture, and tourism. Notably, it forms part of the Nanjing Jiangbei New Area, featuring several burgeoning
industrial parks.6 The historic town of Luhe, nestled within the meandering curve of the Chuhe River, is easily accessible by ferry
and metro. The southeast Lingyan Mountain contributes to the cultural and mythical traditions of the region.7 In its province, EVs
have been experiencing a surge in popularity, constituting about one-quarter of new vehicle registrations in 2022. Impressively, the
province boasts over 9.9 million EVs as of 2022, reflecting a remarkable 96.3% year-to-year increase.8 The City of Nanjing host

4 See https://en.wikipedia.org/wiki/Nanjing#Demographics https://postal-codes.cybo.com/china/luhe-district/ (retrieved on 2023-11-21).

5 See https://en.wikipedia.org/wiki/List_of_United_States_cities_by_population (retrieved on 2023-11-21).
6 See https://english.nanjing.gov.cn/investment/keypark/201602/t20160202_1946071.html (retrieved on 2023-11-21).
7 See https://isocarp.org/app/uploads/2016/01/ISOCARP-AfE-n3.pdf (retrieved on 2023-11-22).
8 See https://www.thenanjinger.com/news/nanjing-news/how-many-cars-are-in-nanjing-and-how-many-are-electric/ (retrieved on 2023-11-22).

9
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Fig. 4. The traffic network. The stars represent road segments of varying demand points.

several major Chinese EV manufacture sites, namely BYTON Auto Industry Park and ChangAn EV’s factories.9 While the city has
been one of the leading cities in China receiving city-level government incentives since 2020,10 Chinese government is planning to
boost EV adoption in suburban areas like Luhe district.11 The district also boasts robust electricity infrastructure exemplified by the
renowned Yangba Power Station.12
In Luhe District, there are over 20 public charging stations, each equipped with at least four DCFCs.13 The largest station boasts
20 DCFCs, all of which are in excellent working condition. These stations are managed by various chained EV charging companies,
including Orange Unicorn, EV POWER, Star Charge, E-Charge by State Grid Corporation of China, among others. Based on interviews
with industry professionals associated with these charging stations or their operating companies, it was revealed that one popular
station, established in 2018, features eight DCFCs. This station consumed over 286,000 kilowatt-hours, achieving a power utilization
rate of 90% (calculated as the ratio of actual charging capacity to rated power capacity). Another station, established in July 2023,
houses four chargers and has serviced more than 8000 EVs. The utilization rate of its chargers surpasses 30%, signaling a significant
demand and suggesting the need for additional chargers or charging stations.14 Remarkably, with the exception of one station, all
charging facilities are situated within 500 meters (approximately 0.3 miles) of the nearest road intersection. Furthermore, seven of
these stations are directly positioned on land adjacent to road intersections.
The traffic flow data covers 46 segments of the first two levels of major roads, which include but are not limited to

1. Ningluo Expressway (G36),

2. Nanjing Ring Expressway (G2501),
3. Jiangbei Avenue Expressway,
4. Shanghai-Shaanxi Expressway (G40),
5. Changshen Expressway (G25),
6. Ninglian Expressway (G205),
7. Jinjiang Road, Changshen Bridge (S353),

9 https://siteselection.com/issues/2019/jul/nanjing-china-a-one-stop-shop-for-new-energy-vehicles.cfm (retrieved on 2023-11-27).

10 https://theicct.org/wp-content/uploads/2023/02/China-NEPC-city-policies_final.pdf (retrieved on 2023-11-27).
11 https://senecaesg.com/insights/china-to-promote-ev-charging-infrastructure-targeting-rural-ev-market/ (retrieved on 2023-11-27).
12 https://www.gem.wiki/Nanjing_Liuhe_District_Yangba_power_station (retrieved on 2023-11-27).
13 https://www.google.com/maps/search/charging+station/@32.3213358,118.7233816,12z/data=!4m8!2m7!3m6!1scharging+station!2sLuhe+District,

+Nanjing,+Jiangsu,+China!3s0x35b5fd7b45f8bd67:0xc176eaae1f573d8a!4m2!1d118.82155!2d32.3222199?authuser=0&entry=ttu (retrieved on 2023-11-27).

14 https://evchargingsummit.com/blog/top-metrics-to-measure-the-performance-of-your-ev-charging-stations/ (retrieved on 2024-01-30).

10
T. Wu et al. Transportation Research Part D 130 (2024) 104178

8. Yanyang Road (S353),

9. G235 National Highway,
10. Eastern Main Line (S4210),
11. Liuhe Bridge (G328),
12. Ninghai Line,
13. Qinglu Line (S356),
14. Emei Road (S247).

And many more segments from other major roads are also included in the dataset. The studied area contains 39 road intersections
in total identified by stars in Fig. 4. In this study, we designate road intersections as potential locations for EV charging stations.
This is grounded in the observation that sites proximate to road intersections are commonly selected for DCFCs due to their inherent
convenience. A pertinent example is the National Electric Vehicle Infrastructure (NEVI) Program, which mandates that EV charging
infrastructure should be situated no more than one mile from interstates or highway intersections15 —a regulation that aligns with
similar guidelines established and implemented in the United States.16 To specify, we designate road intersections as candidate spots
for public EV charging stations, encompassing locations within a radius of 500 meters or 0.3 miles from the road intersections. This
decision is informed by the observation that nearly all existing EV charging stations in the study area are situated within this
proximity to their nearest road intersections. Please note that one limitation associated with placing charging stations near road
intersections is the potential disruption to traffic flow caused by frequent car stops for charging. This has the potential to induce
traffic jams and bottlenecks on highways. To address this concern, additional precautions should be exercised when siting charging
stations near road intersections. This includes optimizing traffic signal timing, avoiding busy intersections with excess capacity,
and strategically clustering charging stations at nearby intersections.17 Despite that they are beyond the scope of this study, these
measures are crucial to ensuring the seamless integration of EV charging infrastructure while mitigating any adverse impact on the
efficiency of traffic flow. We also acknowledge that alternative locations suitable for EV charging stations encompass areas near
transmission lines and substations,18 existing parking lots, as well as traditional fuel or gas stations.19 Our models and methods in
Section 3 are adaptable and can be easily customized to consider diverse sets of candidate sites.
We use the Baidu map coordinate-picking system20 to determine the latitudes and longitudes of these road segments and
intersection, followed by collecting the real-time two-way traffic flow data on these road segments at the Baidu map open platform.21
In particular, we develop a webpage on localhost with JavaScript codes, which allows us to request traffic data every 15 min from
December 2018 to March 2020. The dynamics of traffic flows exhibit temporal variability. Typically, we commonly observe sluggish
traffic during morning and afternoon peak periods on workdays, as well as in the periods preceding and following extended holidays.
While our robust optimization approach (Section 3) adeptly addresses the variation range of the derived charging demand, it fails
to fully encapsulate the real-time dynamics of traffic flows. Consequently, in Section 4.5, we employ a simulation approach to
model the time-dependent charging demand. This approach intricately aligns with the temporal dynamics of traffic flow, providing
a nuanced and temporally sensitive representation of the charging demand scenario.

4.2. Parameter estimation

To this end, we calibrate parameters in this case study based on real data. Table 2 provides a summary of the parameters along
with their corresponding values or the sources from which these values are derived.
We refer the fixed cost 𝑠𝑖 of setting up station 𝑖 to the cost of building power lines to location 𝑖, which is proportional to the
distance of location 𝑖 to the nearest power substation with the unit cost being 20 Chinese Yuan (CNY) per meter. The unit cost 𝑐
of setting up a charger in stations is set to 130,000 CNY.22 For each charger following the general process to serve a customer, the
service rate 𝜇 is set to 2, where the charging time is half an hour on average.23 For the time cost 𝑑𝑖𝑗 of traveling from demand node
𝑗 to station 𝑖 to charge a vehicle, it is proportional to the distance from 𝑖 to 𝑗, where the unit cost per km is calculated according
to the average speed on expressways and the minimum hourly wage in Nanjing.24
Regarding the charging demand, we directly estimate expected value and variability of DC charging demand at point 𝑗 based
on traffic flow in the network. Let 𝑣̄ be the average speed at a certain time on a specific road segment. 𝑣̄ are rarely zero within
15-minute intervals in our data, despite of the fact the traffic may stop moving for a short period of time in practice. Thus, the
expected number of vehicles 𝑘̄ passing this segment in a unit time 𝑇 can be estimated by
𝑇
𝑘̄ = .
𝑣̄

15 See https://blinkcharging.com/understanding-nevi-funding-for-ev-charging-infrastructure-along-americas-highways/ (retrieved on 2023-11-22).

16 See https://www.route-fifty.com/infrastructure/2023/03/race-build-out-ev-charging-stations-some-cities-and-states-have-leg/383559/ (retrieved on 2023-
11-21).
17 See https://energy5.com/examining-the-effects-of-ev-charging-stations-on-traffic-signal-optimization (retrieved on 2023-11-22).
18 See https://calstart.org/electric-highways-study/ (retrieved on 2023-11-27).
19 See https://www.cnbc.com/2023/08/19/how-gas-station-economics-will-change-in-the-ev-charging-future.html (retrieved on 2023-11-21).
20 See http://api.map.baidu.com/lbsapi/getpoint/index.html (retrieved on 2021-12-30).
21 See http://lbsyun.baidu.com (retrieved on 2022-01-15).
22 See https://www.gldjc.com/ for a website to inquiry the construction cost and the material price.(retrieved on 2021-12-30)
23 See https://www.tesla.com for the average charging time.(retrieved on 2021-12-30)
24 See http://jshrss.jiangsu.gov.cn/art/2021/7/28/art_77279_9953183.html for the standard of the minimum hourly wage in Nanjing.(retrieved on 2021-12-30)

11
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Table 2
Notations for parameter estimation.
Parameter Description (Source of ) Value
𝑐 Unit cost of setting up a charger in stations 130,000 CNY
𝜇 Average service rate of a charger 2 vehicles per hour
𝛼 Service level requirement 90%
𝑇 A time period 15 min
𝑝𝑠 Probability of a sample path 0.01
𝑡 Spot time Time-dependent
𝑠𝑖 Fixed cost of setting up a charging station Site-dependent
𝑑𝑖𝑗 Time cost of traveling Footnote (20)
𝐿 Length of a road segment Footnote (20)
𝑛 Number of lanes each way on the segment Footnote (20)
𝑣̄ Average speed on a road segment Footnote (21)
𝑘̄ Expected number of vehicles passing the segment Footnote (21)
𝑣𝑠 Spot speed on a road segment Footnote (21)
𝑣𝑗 Minimal speed on the road segment Footnote (21)
𝑣𝑐 Maximal speed on the road segment Footnote (21)
𝑘𝑗 Number of vehicles associated with 𝑣𝑗 Footnote (21)
𝑘𝑐 Number of vehicles associated with 𝑣𝑐 Footnote (21)
𝑘 Estimated number of vehicles Footnote (21)
𝜔̂𝑗 𝑡 Number of vehicles needing charging on road Footnote (21)
segment 𝑗 at time 𝑡
𝜙 Charging rate, i.e., the proportion of vehicles 1% (0.5%–10% in
needing on-route charging Section 4.4)
𝛤 Uncertainty level 1 (1–14 in Section 4.4)

According to Notley et al. (2009), Jiming et al. (2017), and Funke et al. (2019), the number of vehicles on the road segment affects
the spot speed 𝑣𝑠 , together with minimal and maximal speeds—denoted by 𝑣𝑗 and 𝑣𝑐 , respectively—on a road segment, that is,
(𝑣𝑐 − 𝑣𝑗 ) ln(𝑘∕𝑘𝑐 )
𝑣𝑠 = 𝑣𝑐 − ,
ln(𝑘𝑗 ∕𝑘𝑐 )
where 𝑘𝑗 and 𝑘𝑐 are the number of vehicles associated with the minimal and maximal speeds. On the other hand, spot speed 𝑣𝑠 can
reversely affect the number of vehicles 𝑘, i.e.,
{ }
𝑣 − 𝑣𝑐 𝑘𝑗
𝑘 = 𝑘𝑐 exp ln (10)
𝑣𝑗 − 𝑣𝑐 𝑘𝑐
Suppose the length of the road segment is 𝐿, and there are 𝑛 lanes each way on the road segment. At certain time point 𝑡, we can
derive the number of vehicles 𝑘 on one unit of road segment per unit of time based on (10). Therefore, the total charging demand
at time 𝑡 on that road segment across the time horizon of study 𝑇 is estimated as

𝜔̂𝑗 𝑡 = 𝜙𝑘𝐿𝑛𝑇 . (11)

We take the mean and variance of all 𝜔̂𝑗 𝑡 over time as the expected value and the variability of charging demand in the robust
optimization model. The charging rate 𝜙 refers to the percentage of vehicles requiring on-route charging service among all vehicles
on the roads. Given historical traffic flow in this area, we project that 𝜙 = 1% of these vehicles would be EVs that require on-
route charging service in the near future. This estimation is due to several rationales. First, the private EV adoption in China has
experienced remarkable growth. Although EVs accounted for 4.9% of all passenger vehicles in China, EV sales in China were more
than double that of the US and Canada in 2022.25 Secondly, the state of EV adoption in specific regions is also noteworthy. In
2017, California witnessed an EV adoption rate of approximately 5% according to Bellan (2018). Moreover, by the end of 2018,
EV sales constituted over 10% of all passenger vehicles in Canada (Schmidt, 2019). Finally, we assume almost all EVs traveling
in this inner-city urban transportation network rely only on DCFCs. This is due to the fact that almost all residences in the area
are multi-story apartment buildings, making it impossible for EV users to charge at home. Given escalating EV adoption rates, we
explore the optimal sizing and siting of charging stations under different charging rates in sensitivity analysis (Section 4.4).

4.3. Optimums

In this section, we present the solution of a robust model based on the case study. For comparison, we also find optimums to
the two benchmarks corresponding to model (4)—the deterministic model and the stochastic model (8). The deterministic model
is obtained by replacing each charging demand 𝜔𝑗 in (4) with the corresponding expectation. To solve model (9), we generate 100
scenarios with appropriately generated probability 𝑝𝑠 = 0.01. For each scenario, the random demand 𝜔𝑠𝑗 are generated by a normal

25 See https://en.wikipedia.org/wiki/Electric_car_use_by_country (retrieved on 2023-11-21).

12
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Fig. 5. The optimum to the deterministic, stochastic, and robust optimization problem. The size of the red dots represent the number of chargers in the charging
station, while the stars are the candidate stations without chargers in the optimum. (For interpretation of the references to color in this figure legend, the reader
is referred to the web version of this article.)

Fig. 6. The optimum to the robust optimization problem under varying uncertainties. The size of the red dots represent the number of chargers in the charging
station, while the stars are the candidate stations without chargers in the optimum. (For interpretation of the references to color in this figure legend, the reader
is referred to the web version of this article.)

distribution of the same mean and standard deviation as the historical traffic flow data. Then model (9) is solved to obtain the
stochastic optimums.
We solve the proposed robust optimization model (6) and two benchmark models by Gurobi 9.0.2 in Python 3.8.2. The results
by three models are presented in Fig. 5, where we set 𝛤 = 1 for the robust model. Compared with the deterministic and the stochastic
models, the robust model leads to more allocated chargers in more scattered stations. This is because the robust model is required
to be insensitive to uncertain charging demand, which seeks for more chargers to be allocated at more candidate stations in robust
optimum than the those of deterministic and stochastic models.

4.4. Sensitivity analysis

We perform sensitivity analysis on the robust model with respect to uncertainties, i.e., the values of 𝛤 . In Fig. 6, we present the
solution to model (6) with the open charging stations represented by red dots when 𝛤 = 4, 7, 14, and the size of dots corresponds
to the number of chargers in the selected stations. Charging piles become more widely distributed when the value of 𝛤 increases.
Specifically, when 𝛤 ≤ 7, six charging stations will be established; when 7 ≤ 𝛤 ≤ 14, there would be 8 charging stations needed; and
when 𝛤 ≥ 14, we would require 14 charging stations. Fig. 7 illustrates the change of total costs with respect to 𝛤 . Uncertainty of
demand can raise the total cost, because the total cost grows as the value of 𝛤 increases. Moreover, in the three ranges (1 ≤ 𝛤 ≤ 6,
7 ≤ 𝛤 ≤ 14 and 𝛤 ≥ 14), the total cost and 𝛤 show nearly linear relationship. After reaching the critical value 𝛤 = 14, the number of
charging piles no longer increases, although the total cost slightly increases. This shows that when the uncertainty reaches a certain
level, the robust optimization has reached the worst-case scenario and will not become worse. The growing cost is due to the travel
time cost of EVs to charging stations only, and yet the infrastructure costs on power grid and chargers remain unchanged.

13
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Fig. 7. Total costs with respect to different 𝛤 values.

Fig. 8. The optimum to the robust optimization problem under varying charging rate, i.e., the proportion of traffic flow requiring on-route charging. The size
of the red dots represent the number of chargers in the charging station, while the stars are the candidate stations without chargers in the optimum. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Furthermore, we examine the optimal solution of the robust optimization problem under varying charging demands. With the
proliferation of EV adoption, we note that the optimal sizes of charging stations exhibit an increase, while the optimal siting
of charging stations remains relatively constant (Fig. 8). Additionally, our observations reveal a linear relationship between the
penetration rates of EVs and the total optimal costs (Fig. 9).

4.5. Simulation

In this section, we first briefly introduce our simulation model and discuss our model validation, followed by two scenarios to
verify effectiveness and efficiency of the deterministic, stochastic, and robust optimums mentioned beforehand. In this simulation
study, we consider charging stations with time-dependent charging demand arrivals. That is, we relax the assumption of time-
independent queues in the optimization models, providing insights regarding stations’ allocation and deployment decisions in a
realistic traffic setting.

4.5.1. Modeling framework

To assess the efficacy of the optimal solution within a practical transportation network, we develop a simulation model utilizing
ARENA software version 16.0 (Rockwell). Our simulation is designed as a discrete event model, and it takes into account the presence
of uncertain traffic intensity, which represents the fluctuating number of vehicles per unit length over a given period of time. This
simulation is based on the following assumptions:

• At each road intersection, vehicles choose a specific turning direction, namely, go straight, turn left or right, with probabilities
based on historical data observed in the modeled metropolitan area.

14
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Fig. 9. Total costs with respect to different charging rates.

Table 3
Compare simulated and real traffic flows.
Road sections Traffic flow (vehicles per day) Percentage error
Real data Simulated
Local 16,763 16,168 3.55%
Highway 12,344 12,733 3.15%

• No vehicles stop in the network unless for charging purposes at identified charging stations. In other words, there is no
interruption of the traffic flow.
• Charging stations are independent queueing systems with chargers as servers.
• Vehicles’ charging time follows an independent and identical triangular distribution, with a minimum time of 20 min, a
maximum of 40 min and a mode—the most likely time required to charge a vehicle—of 30 min.
• The traffic is generated at source nodes located along the boundaries of the network under study.

Please note that the source nodes in the simulation models are different from the demand nodes in optimization models
aforementioned. Now we explain how the ‘‘traffic amount’’ at source nodes (i.e., the number of vehicles per unit time from each
source node) are generated in our discrete event simulation model. Each of the fourteen source nodes generates entities (i.e., EVs)
in the network, based on traffic distributions estimated from historical data. Specifically, we first estimate vehicles’ arrivals at each
source node, and find time-dependent Poisson processes that fit their arrivals the best. Then we use ARENA’s Schedule Data sheet
module to incorporate these distributions in our simulation model. Vehicles are routed to different road segments following estimated
probabilities. The movement of entities along each road segment is governed by specific speed distributions. These distributions
are determined based on historical data and are fitted using MATLAB. These speed distributions are stored in an Arena 39 × 39
expression. Each time an entity enters a new road segment, it is assigned a speed following the speed distribution on that segment.
We use ARENA’s process modules to model charging activities at each charging station.
We perform model validation to make sure (1) that our simulation model integrates all specifications of the system under study,
and (2) that it behaves as close as possible to the system under study. Specifically, we compare real traffic flows of the existing
network to those generated by source nodes in the simulation model. Table 3 presents the traffic flows for two randomly selected road
segments within the simulated network, representing one highway and one local road segment. The percentage of error displayed
in Table 3 indicates minor discrepancies between the simulated and actual traffic flows. Consequently, the simulation model is
deemed valid with a 95% level of confidence. Moreover, we use the animation of our model to attest that the simulated network is
a reasonably close approximation of the existing network.
Considering that urban traffic in any modeled area runs in an infinite time horizon, we conduct analysis to determine the warm-
up period for the model. Specifically, we first run 30 replications of the simulation model and analyze output statistics, including
lengths of queues and number of entities in process of all replications, through the ARENA output analyzer. Our analysis shows that
all replications reach a steady state after 5 days. Therefore, we identified a 5-day warm-up period followed by a 30-day simulation
run length as the study period.

4.5.2. Scenarios and results

We use the following key performance indicators (KPIs) to evaluate the optimal solutions resulting from Section 3.

Service rate refers to the percentage of charging demand being fulfilled across all available charging stations. This represents the
ratio of charged EVs to the total number of EVs that require charging at each station. This metric serves as a measure of the
charging network’s efficiency.

15
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Table 4
Summarized simulation results for three optimization problems.
Optimization Utilization (%) Service rate (%)
<50 >75 >80 >85 Average ≥50 Average
Deterministic 77 13 13 13 21.75 31 29.27
Stochastic 82 8 5 5 15.72 26 24.60
Robust 15 77 77 72 85.76 72 70.57

Utilization (in %) is the fraction of time when a charger is charging EVs. This is a measure of effectiveness for the charging
configuration.

In our simulation, we model the charging stations and their capacities resulting from deterministic, stochastic and robust
optimization problems. Simulated EV users passing by a charging station have a fixed probability of charging their vehicles. If
they choose to charge, they are served on a first-come-first-serve (FCFS) basis. In the case where all chargers in the chosen charging
station are occupied, the EV will wait in the pooled queue, if the queue length at the charging station is constrained to be no greater
than the number of available chargers. In simpler terms, we assume that EV drivers are willing to wait if, on average, there is no more
than one vehicle in front of them for each charger. This assumption is reasonable considering clients’ patience and the required time
to charge an EV. When the number of EVs waiting for charging at the chosen station exceeds the number of available chargers, the
EV user will opt to go to the second nearest charging station with available charging spots. To this end, with simulated traffic flow
in the network that is very close to the reality, we are able to compare the efficiency and effectiveness of different configurations
resulting from Section 3.
In Table 4, we present simulation results from three types of optimization. In the Average columns, we report the percentages
of average utilization and average service rate across all proposed charging stations in the three optimums. Charging configuration
resulting from robust optimization realizes an average utilization of over 85%, or four-folds for that of deterministic optimization and
5.5 times better than the stochastic method. The service rate for robust optimum is over 70%, more than double that of deterministic
and almost three times that of the stochastic method.
In addition, we display the proportion of stations whose utilization is less than 50%, over 75%, more than 80% and beyond
85%. In the optimal charging network resulting from robust optimization, 72% of the charging stations have utilization over 85%
and only 15% of the stations are utilized less than 50%. In contrast, 77% and 82% of charging stations are used less than 50% from
the deterministic and stochastic optimizations, respectively. While most of the charging stations from deterministic and stochastic
optimization are not well utilized, the sufficiently utilized stations (utilization rate over 80%) account for 13% in the deterministic
method and only 5% for the stochastic method.
Moreover, we also show the proportion of charging stations having a service level of no less than 50%. In the robust
optimal configuration, over 70% of the charging stations can fulfill at least half of their demand. Whereas fewer than one
third of deterministic optimal stations and 26% of the stochastic ones can serve 50% of their charging demand. Therefore, we
demonstrate that robust optimization outperforms deterministic and stochastic optimization to accommodate uncertain traffic flows
and associated charging demand for public EV charging.

5. Conclusion

This forward-looking study aims to minimize the infrastructure costs of setting up public DCFC stations to meet emerging
EV charging demand in any specific metropolitan area. Because of current extremely low adoption rates of EVs worldwide, it is
impossible to gauge the exact charging demand or its distribution. Motivated by this challenge of unknown and uncertain charging
demand, we develop a robust optimization problem with queue theory, after estimating the first and second moments of charging
demand based on existing traffic flows in the transportation networks. To overcome the intractability of queueing constraints in the
robust optimization, we figure out an approximate yet highly accurate linear relationship between the number of servers and arrival
rates under any specific loss rate. Therefore, we can solve the deterministic dual of the original robust optimization, and hence find
the optimum to the robust optimization problem.
Furthermore, we validate the optimum using a simulation approach, which takes into account more realistic factors, namely
time-varying traffic flows and EV users’ limited patience to wait for available chargers. The simulation approach finds that the
realized utilization of chargers in the charging station layout determined by the robust optimization approach is very close to the
targeted utilization rate.
Although our case study focuses on a specific metropolitan region in Southeast China, both our robust optimization and
simulation approaches can be easily applied to other geographic areas. Nevertheless, our static models examine a single-epoch only.
It is worthwhile for future research to investigate multiple-period decision making problems on public charging station allocation.
Specifically, at each decision epoch, it is valuable to allocate charging stations with newly added chargers that satisfy constantly
growing charging demand over time.

16
T. Wu et al. Transportation Research Part D 130 (2024) 104178

CRediT authorship contribution statement

Ting Wu: Writing – review & editing, Writing – original draft, Supervision, Resources, Methodology, Formal analysis, Data
curation, Conceptualization. Emily Fainman: Writing – review & editing, Writing – original draft, Validation, Software, Project
administration, Methodology, Formal analysis, Conceptualization. Yasmina Maïzi: Validation, Software, Methodology, Formal
analysis. Jia Shu: Validation, Supervision, Project administration. Yongzhen Li: Writing – original draft, Methodology, Formal
analysis.

Acknowledgments

Dr. Ting Wu’s work was supported by the National Natural Science Foundation of China (No. 71971108), the Mathematics
Tianyuan Fund of the National Natural Science Foundation of China (No. 12326321), and the Top Six Talents’ Project of Jiangsu
Province, China (No. XNYQC-001). Dr. Jia Shu’s research was supported by the National Natural Science Foundation of China (No.
72091213 and 71831004) and the Fundamental Research Funds for the Central Universities, China. Dr. Yongzhen Li’s research was
supported by the National Natural Science Foundation of China (No. 72101056) and Southeast University Zhishan Young Scholar
Program, China.

References

Arias, M.B., Bae, S., 2016. Electric vehicle charging demand forecasting model based on big data technologies. Appl. Energy 183, 327–339.
Arias, M.B., Kim, M., Bae, S., 2017. Prediction of electric vehicle charging-power demand in realistic urban traffic networks. Appl. Energy 195, 738–753.
Azadeh, S.S., Vester, J., Maknoon, M., 2022. Electrification of a bus system with fast charging stations: Impact of battery degradation on design decisions. Transp.
Res. C 142, 103807.
Bae, S., Kwasinski, A., 2011. Spatial and temporal model of electric vehicle charging demand. IEEE Trans. Smart Grid 3 (1), 394–403.
Bandi, C., Bertsimas, D., Youssef, N., 2015. Robust queueing theory. Oper. Res. 63 (3), 676–700.
Basile, F., Chiacchio, P., Teta, D., 2012. A hybrid model for real time simulation of urban traffic. Control Eng. Pract. 20 (2), 123–137.
Bellan, R., 2018. The grim state of electric vehicle adoption in the U.S.. https://www.citylab.com/transportation/2018/10/where-americas-charge-towards-
electric-vehicles-stands-today/572857/.
Berkelmans, G., Berkelmans, W., Piersma, N., Van Der Mei, R., Dugundji, E., 2018. Predicting electric vehicle charging demand using mixed generalized extreme
value models with panel effects. Procedia Comput. Sci. 130, 549–556.
Bertsimas, D., Gamarnik, D., Rikun, A.A., 2011. Performance analysis of queueing networks via robust optimization. Oper. Res. 59 (2), 455–466.
Bertsimas, D., Sim, M., 2004. The price of robustness. Oper. Res. 52 (1), 35–53.
Bertsimas, D., Thiele, A., 2006. Robust and data-driven optimization: modern decision making under uncertainty. In: Models, Methods, and Applications for
Innovative Decision Making. INFORMS, pp. 95–122.
Brandstätter, G., Leitner, M., Ljubić, I., 2020. Location of charging stations in electric car sharing systems. Transp. Sci. 54 (5), 1408–1438.
Breuer, L., 2008. Continuity of the M/G/c queue. Queueing Syst. 58 (4), 321–331.
Cai, H., Jia, X., Chiu, A.S., Hu, X., Xu, M., 2014. Siting public electric vehicle charging stations in Beijing using big-data informed travel patterns of the taxi
fleet. Transp. Res. D 33, 39–46.
Cilio, L., Babacan, O., 2021. Allocation optimisation of rapid charging stations in large urban areas to support fully electric taxi fleets. Appl. Energy 295, 117072.
Dong, J., Liu, C., Lin, Z., 2014. Charging infrastructure planning for promoting battery electric vehicles: An activity-based approach using multiday travel data.
Transp. Res. C 38, 44–55.
Funke, S.Á., Sprei, F., Gnann, T., Plötz, P., 2019. How much charging infrastructure do electric vehicles need? A review of the evidence and international
comparison. Transp. Res. D 77, 224–242.
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. C 68, 389–402.
Habib, S., Khan, M.M., Hashmi, K., Ali, M., Tang, H., 2017. A comparative study of electric vehicles concerning charging infrastructure and power levels. In:
2017 International Conference on Frontiers of Information Technology. FIT, Institute of Electrical and Electronics Engineers, Inc., Islamabad, Pakistan, pp.
327–332.
Hall, D., Lutsey, N., 2017. Emerging Best Practices for Electric Vehicle Charging Infrastructure. Technical Report, The International Council on Clean Transportation
(ICCT), Washington, DC, USA.
Harel, A., 1988. Sharp bounds and simple approximations for the Erlang delay and loss formulas. Manage. Sci. 34 (8), 959–972.
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. C 67, 131–148.
Hokstad, P., 1978. Approximations for the M/G/m queue. Oper. Res. 26 (3), 510–523.
Hosseini, S., Sarder, M., 2019. Development of a Bayesian network model for optimal site selection of electric vehicle charging station. Int. J. Electr. Power
Energy Syst. 105, 110–122.
Islam, M.M., Shareef, H., Mohamed, A., 2018. Optimal location and sizing of fast charging stations for electric vehicles by incorporating traffic and power
networks. IET Intell. Transp. Syst. 12 (8), 947–957.
Jiming, H., Lingyu, K., Yaqi, S., Ying, L., Wenting, X., Hao, W., 2017. A review of demand forecast for charging facilities of electric vehicles. In: IOP Conference
Series: Materials Science and Engineering. IOP Publishing, 012040.
Jung, J., Chow, J.Y.J., Jayakrishnan, R., Park, J.Y., 2014. Stochastic dynamic itinerary interception refueling location problem with queue delay for electric taxi
charging stations. Transp. Res. C 40, 123–142. http://dx.doi.org/10.1016/j.trc.2014.01.008.
Kchaou-Boujelben, M., 2021. Charging station location problem: A comprehensive review on models and solution approaches. Transp. Res. C 132, 103376.
Khazaei, H., Misic, J.V., Misic, V.B., 2011. Modelling of cloud computing centers using M/G/m queues. In: International Conference on Distributed Computing
Systems Workshops.
Khazaei, H., Misic, J., Misic, V.B., 2012. Performance analysis of cloud computing centers. IEEE Trans. Parallel Distrib. Syst. 23 (5), 936–943.
Kimura, T., 1983. Diffusion approximation for an M/G/m queue. Oper. Res. 31 (2), 304–321.
Kınay, Ö.B., Gzara, F., Alumur, S.A., 2021. Full cover charging station location problem with routing. Transp. Res. B 144, 1–22.
Li, M., Jia, Y., Shen, Z., He, F., et al., 2016. Strategic Charging Infrastructure Deployment for Electric Vehicles. Technical Report, California. Department of
Transportation.

17
T. Wu et al. Transportation Research Part D 130 (2024) 104178

Li, N., Jiang, Y., Zhang, Z.-H., 2021. A two-stage ambiguous stochastic program for electric vehicle charging station location problem with valet charging service.
Transp. Res. B 153, 149–171.
Li, J., Xie, C., Bao, Z., 2022. Optimal en-route charging station locations for electric vehicles: A new modeling perspective and a comparative evaluation of
network-based and metanetwork-based approaches. Transp. Res. C 142, 103781.
Maïzi, Y., Zhu, E.C., Wu, T., Zhou, J., 2019. A reliable deployment strategy for public electric vehicle charging stations: A discrete event simulation model for
power grid and traffic networks. In: 2019 Winter Simulation Conference. WSC, IEEE, pp. 1660–1671.
Mak, H.-Y., Rong, Y., Shen, Z.-J.M., 2013. Infrastructure planning for electric vehicles with battery swapping. Manage. Sci. 59 (7), 1557–1575.
Mallig, N., Heilig, M., Weiss, C., Chlond, B., Vortisch, P., 2016. Modelling the weekly electricity demand caused by electric cars. Future Gener. Comput. Syst.
64, 140–150. http://dx.doi.org/10.1016/j.future.2016.01.014.
Marmaras, C., Xydas, E., Cipcigan, L., 2017. Simulation of electric vehicle driver behaviour in road transport and electric power networks. Transp. Res. C 80,
239–256. http://dx.doi.org/10.1016/j.trc.2017.05.004.
Notley, S., Bourne, N., Taylor, N., 2009. Speed, Flow and Density of Motorway Traffic. TRL INSIGHT REPORT.
Pourgholamali, M., Homem de Almeida Correia, G., Tarighati Tabesh, M., Esmaeilzadeh Seilabi, S., Miralinaghi, M., Labi, S., 2023. Robust design of electric
charging infrastructure locations under travel demand uncertainty and driving range heterogeneity. J. Infrastruct. Syst. 29 (2), 04023016.
Quddus, M.A., Shahvari, O., Marufuzzaman, M., Ekşioğlu, S.D., Castillo-Villar, K.K., 2021. Designing a reliable electric vehicle charging station expansion under
uncertainty. Int. J. Prod. Econ. 236, 108132.
Schiffer, M., Walther, G., 2018. Strategic planning of electric logistics fleet networks: A robust location-routing approach. Omega 80, 31–42.
Schmidt, E., 2019. Electric vehicles sales update Q3 2018, Canada. https://www.fleetcarma.com/electric-vehicles-sales-update-q3-2018-canada/.
Setiawan, A.D., Zahari, T.N., Anderson, K., Moeis, A.O., Hidayatno, A., 2023. Examining the effectiveness of policies for developing battery swapping service
industry. Energy Rep. 9, 4682–4700.
Shanmukhappa, T., Ho, I.W.H., Tse, C.K., 2018. Spatial analysis of bus transport networks using network theory. Physica A 502, 295–314. http://dx.doi.org/10.
1016/j.physa.2018.02.111.
Shen, Z.-J.M., Feng, B., Mao, C., Ran, L., 2019. Optimization models for electric vehicle service operations: A literature review. Transp. Res. B 128, 462–477.
Shortle, J.F., Thompson, J.M., Gross, D., Harris, C.M., 2018. Fundamentals of Queueing Theory, fifth ed. John Wiley & Sons.
Sun, H., Yang, J., Yang, C., 2019. A robust optimization approach to multi-interval location-inventory and recharging planning for electric vehicles. Omega 86,
59–75.
Thomin, P., Gibaud, A., Koutcherawy, P., 2013. Deployment of a fully distributed system for improving urban traffic flows: A simulation-based performance
analysis. Simul. Model. Pract. Theory 31, 22–38.
Tijms, H.C., Hoorn, M.H.V., Federgruen, A., 1981. Approximations for the steady-state probabilities in the M/G/c queue. Adv. Appl. Probab. 13 (1), 186–206.
Union of Concerned Scientists, 2018. Electric vehicle charging: Types, time, cost and savings. https://www.ucsusa.org/resources/electric-vehicle-charging-types-
time-cost-and-savings.
Wang, H., Zhao, D., Meng, Q., Ong, G.P., Lee, D.-H., 2019. A four-step method for electric-vehicle charging facility deployment in a dense city: An empirical
study in Singapore. Transp. Res. A 119, 224–237.
Whitt, W., You, W., 2018. Using robust queueing to expose the impact of dependence in single-server queues. Oper. Res. 66 (1), 184–199.
Wolbertus, R., Kroesen, M., van den Hoed, R., Chorus, C., 2018. Fully charged: An empirical study into the factors that influence connection times at EV-charging
stations. Energy Policy 123, 1–7.
Xi, X., Sioshansi, R., Marano, V., 2013. Simulation–optimization model for location of a public electric vehicle charging infrastructure. Transp. Res. D 22, 60–69.
Xie, F., Liu, C., Li, S., Lin, Z., Huang, Y., 2018. Long-term strategic planning of inter-city fast charging infrastructure for battery electric vehicles. Transp. Res.
E 109, 261–276.
Yang, W., 2018. A user-choice model for locating congested fast charging stations. Transp. Res. E 110, 189–213.
Yang, J., Dong, J., Hu, L., 2017. A data-driven optimization-based approach for siting and sizing of electric taxi charging stations. Transp. Res. C 77, 462–477.
Yao, D.D., 1985. Refining the diffusion approximation for the M/G/m queue. Oper. Res. 33 (6), 1266–1277.
Yi, Z., Liu, X.C., Wei, R., 2022. Electric vehicle demand estimation and charging station allocation using urban informatics. Transp. Res. D 106, 103264.
Yılmaz, H., Yagmahan, B., 2022. Range coverage location model: An optimization model for the charging station location problem in a transportation network
to cover intercity travels. Int. J. Energy Res. 46 (2), 1538–1552.

18