This article was downloaded by: [131.175.147.

3] On: 27 June 2023, At: 07:22

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Transportation Science
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

Charging Station Location and Sizing for Electric Vehicles

Under Congestion
Ömer Burak Kınay, Fatma Gzara, Sibel A. Alumur

To cite this article:

Ömer Burak Kınay, Fatma Gzara, Sibel A. Alumur (2023) Charging Station Location and Sizing for Electric Vehicles Under
Congestion. Transportation Science

Published online in Articles in Advance 16 Jun 2023

. https://doi.org/10.1287/trsc.2021.0494

Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

Copyright © 2023, INFORMS

Please scroll down for article—it is on subsequent pages

With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)
and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual
professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to
transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 TRANSPORTATION SCIENCE
Articles in Advance, pp. 1–19
https://pubsonline.informs.org/journal/trsc ISSN 0041-1655 (print), ISSN 1526-5447 (online)

Charging Station Location and Sizing for Electric Vehicles

Under Congestion
Ömer Burak Kınay,a,b Fatma Gzara,a Sibel A. Alumura,*
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.

a
Department of Management Sciences, University of Waterloo, Waterloo, Ontario N2L 3G1, Canada; b Amazon.com, Seattle, Washington 98109
*Corresponding author
Contact: obkinay@uwaterloo.ca (ÖBK); fgzara@uwaterloo.ca, https://orcid.org/0000-0003-2893-8597 (FG); sibel.alumur@uwaterloo.ca,
https://orcid.org/0000-0003-1550-8053 (SAA)

Received: November 26, 2021
Revised: October 31, 2022; March 10, 2023
Accepted: May 8, 2023
Published Online in Articles in Advance:
June 16, 2023
https://doi.org/10.1287/trsc.2021.0494
Copyright: © 2023 INFORMS
Abstract. This paper studies the problem of determining the strategic location of charging
stations and their capacity levels under stochastic electric vehicle flows and charging times
taking into account the route choice response of users. The problem is modeled using bilevel
optimization, where the network planner or leader minimizes the total infrastructure cost of
locating and sizing charging stations while ensuring a probabilistic service requirement on
the waiting time to charge. Electric vehicle users or followers, on the other hand, minimize
route length and may be cooperative or noncooperative. Their choice of route in turn deter­
mines the charging demand and waiting times at the charging stations and hence, the need
to account for their decisions by the leader. The bilevel problem reduces to a single-level
mixed-integer model using the optimality conditions of the follower’s problem when the
charging stations operate as M/M/c queues and the followers are cooperative. To solve the
bilevel model, a decomposition-based solution methodology is developed that uses a new
logic-based Benders algorithm for the location-only problem. Computational experiments
are performed on benchmark and real-life highway networks, including a new eastern U.S.
network. The impact of route choice response, service requirements, and deviation tolerance
on the location and sizing decisions are analyzed. The analysis demonstrates that stringent
service requirements increase the capacity levels at open charging stations rather than their
number and that solutions allowing higher deviations are less costly. Moreover, the differ­
ence between solutions under cooperative and uncooperative route choices is more signifi­
cant when the deviation tolerance is lower.

History: This paper has been accepted for the Transportation Science Special Issue on 2021 TSL Work­
shop: Supply and Demand Interplay in Transport and Logistics.
Funding: This research was supported by the Ontario Graduate Scholarship when Ö. B. Kınay was a
PhD candidate at the University of Waterloo, and this support is acknowledged.
Supplemental Material: The online appendix is available at https://doi.org/10.1287/trsc.2021.0494.

Keywords: facility location • capacity allocation • charging station • electric vehicles • bilevel optimization

1. Introduction
The transportation sector contributes to approximately
one quarter of global direct CO2 emissions, where pas­
senger and freight road vehicles account for nearly 74%
of that share (IEA 2020). The widespread adoption of
zero-emission vehicles (ZEVs), which consist of battery
electric vehicles (EVs) and fuel cell electric vehicles, is
considered to be one of the most effective ways to miti­
gate the negative environmental impacts of transporta­
tion and to promote environmental sustainability (Zeng
et al. 2021).
The Electric Vehicles Initiative, which was established
by 15 countries including Canada, China, Germany,
and the United Kingdom, has set an aspirational goal of
reaching a 30% sales share of EVs by 2030. The United
States declared their goal of achieving 50% of all new
passenger cars and light trucks sold in 2030 to be ZEVs
(The White House 2021). The majority of the European
Union countries and the United Kingdom aim by
2025–2035 and Canada aims by 2040 to reach 100% ZEV
sales (IEA 2021). As for the private sector commitments
toward electric mobility, the Climate Group’s EV100 Ini­
tiative brings together over 100 companies dedicated to
making electric transport the new normal by 2030 (The
Climate Group 2021a). Many corporations are aiming to
speed up the transition to electric mobility by converting
their fleets to EVs and designing private networks by
locating charging stations.
Despite these global initiatives and endeavors, signifi­
cant adoption barriers remain. In particular, the EV100
members reported the lack of charging infrastructure

1

2 Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
as their top concern (The Climate Group 2021b). To ad­
dress this challenge and improve spatial distribution of
charging stations, many countries have set investment
goals for the deployment of publicly accessible charging
infrastructure. For example, EU countries are mandated
to set deployment targets for publicly accessible EV
chargers for the decade ahead, with a suggested ratio of
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
one charging stall per 10 EVs (European Commission
2019).
The majority of shorter-distance trips using EVs typ­
ically rely on the private slow chargers readily avail­
able at homes and workplaces. However, private slow
charging has to be complemented by publicly acces­
sible en route fast-charging stations to support EV
charging on the go for long-distance trips. Further­
more, recent advancements in battery and charging
technologies for passenger EVs are paving the way
toward long-haul freight distribution with heavy-duty
EVs (i.e., electric trucks (ETs)). Lately, manufacturers,
such as Tesla, Daimler AG, and Volvo, have unveiled
their plans for mass production of ETs within this
decade. The International Energy Agency states that
planning needs to start now for megachargers to enable
long-distance electric trucking as there is currently no
infrastructure to support it (IEA 2021). Therefore, the fea­
sibility and convenience of long-distance trips with electric
vehicles depend on the spatial distribution of adequately
capacitated Direct Current (DC) fast-charging stations.
The capacity of a DC fast-charging station is defined
as the number of charging stalls, also known as outlets,
chargers, or electric vehicle supply equipment. Consid­
ering the relatively long recharging times for EVs even
using fast chargers, waiting times may become a major
inconvenience. Moreover, the demand for charging at a
charging station is hardly known in advance. Hence,
the stochasticity of demand and service times should
not be neglected when determining the locations and
capacities of the charging stations.
In this work, we seek to find the optimal locations of
DC fast-charging stations and their respective number of
chargers to serve a given set of origin-destination (OD)
EV trips on a transportation network. To ensure quality
of service considering stochasticity in charging demand
and service times, probabilistic service-level requirements
are defined that make sure that the expected waiting
times at a charging station do not exceed inconvenience
limits. To provide cost-effective solutions, the framework
of this problem allows the possibility of detours from
shortest paths within a threshold, referred to as the devia­
tion tolerance (Arslan et al. 2019).
Given the possibility of detours, when there exist
alternative routing options, the decision makers need to
consider the choice of charging stations to be visited
by rational service seekers who act with self-interest.
This necessitates the route choice response of EV users
to be taken into account when the location and sizing
Kınay, Gzara, and Alumur: Charging Station Location and Sizing
decisions are made. We use a bilevel structure where
the leader is a regulatory body that seeks to establish the
charging infrastructure by determining the strategic
locations and sizes of the charging stations (e.g., govern­
ments or private EV network operators) and the follower
is the set of EV users who seek to complete their long-
distance EV trips.
The leader’s problem minimizes the total cost of locat­
ing and sizing charging stations while ensuring a con­
venient service level taking follower’s response into
account. The follower’s problem, on the other hand, is
defined to incorporate the route and charging station
choice of EV users while minimizing the length of their
trips. The choice of facilities by the EV users (i.e., the
allocation of recharging demand to charging stations) is
crucial for the leader because the congestion level at sta­
tions will directly impact the required capacity.
The generic model that we propose assumes a stochas­
tic stream of EV flow, stochastic service times at charging
stations, and a homogeneous fleet of EVs with a deter­
ministic battery range. The EVs energy consumption is
assumed to be proportional to the travel distance. There
is a shortest path deviation tolerance for EV users, and
they cannot exceed this tolerance while completing their
OD trips. In order to take range anxiety and energy con­
sumption variability into account in our computational
experiments, without loss of generality, we assume that
EV users start their journey with 80% of the available bat­
tery range, and they may charge up to 80% at a charging
station visit. Charging stalls at stations are considered to
be homogeneous, and their count at a single station can­
not exceed a predefined limit. There is a fixed cost of
locating a charging station and a variable cost of capacity
allocation, which is a nondecreasing function of the num­
ber of charging stalls installed at a station. As a special
case of this generic setting, we additionally introduce a
model that assumes an M/M/c queuing system for
charging stations and configures the problem on an open
multiserver Jackson network (i.e., with Poisson arrivals,
exponential service times, and first come, first served and
single-line queues at stations). The cooperative and unco­
operative user responses further the analysis of the bile­
vel modeling. Under a cooperative response, EV users
make the minimum number of stops required to com­
plete their trips, and the model is converted into a single-
level formulation. Under the uncooperative response, we
assume EV users prefer to make the minimum number of
stops as well.
To the extent of our knowledge, this study is the
first to model a chance-constrained bilevel optimization
framework and that makes charging station sizing deci­
sions based on stochastic queuing models focusing on
long-distance transportation with EVs. The main contri­
butions of our work are (i) introducing a new bilevel
mathematical model that allows shortest path devia­
tions without path pregeneration or enumeration of
 Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
alternative routes, (ii) merging the bilevel nature of the
problem with realistic queuing models to incorporate
stochastic travel demand, (iii) developing an exact solu­
tion algorithm for the uncapacitated version of the
problem, and (iv) developing a novel decomposition-
based algorithm (DA) that is capable of rapidly finding
high-quality solutions and also able to handle both
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
cooperative and uncooperative responses of EV users.
Through computational experimentation, we evaluate
the performance of the solution algorithm and show
how the follower’s position, deviation tolerance, and
service-level parameters affect the obtained solutions
on benchmark instances and real-life California and
eastern U.S. highway networks.
The next section of the paper provides the literature
review of the location problems regarding refueling
and charging stations while emphasizing our contribu­
tions. Section 3 gives the problem statement, defines the
underlying expanded network and its reduction proce­
dure, and presents the general bilevel modeling frame­
work. Section 4 provides the analysis of the bilevel
optimization model under cooperative and uncoopera­
tive EV user response and M/M/c queuing systems.
Section 5 presents the logic-based Benders decomposi­
tion algorithm developed for the uncapacitated version
of the problem and a decomposition-based algorithm
that can account for the cooperative or uncooperative
user response for the original problem. Section 6 intro­
duces the data sets and test instances and presents the
results of extensive computational experiments per­
formed to derive managerial insights as well as to
assess the performance of the algorithm. Lastly, Sec­
tion 7 includes a summary of the insights, concluding
remarks, and key takeaways.
2. Literature Review
Many researchers focused on the spatial dimension of
the charging infrastructure problem within an uncapaci­
tated framework excluding nonspatial attributes and
capacity limitations. These problems can be considered
as extensions of a general facility location problem with
additional scopes of EV-specific characteristics, such as
limited vehicle range. Alternative research streams con­
cerning the node-based and flow-based models emerged
to cover different perspectives of the problem (Upchurch
and Kuby 2010). The former approach, which consists of
p-median and node coverage-based location models, is
more suitable for applications in an urban setting (Good­
child and Noronha 1987, Frade et al. 2011, Tu et al. 2016).
On the other hand, the flow-based models, led by the
pioneering study by Kuby and Lim (2005), received
more attention as they are applicable for enabling longer
EV trips without running out of battery.
The flow-refueling location model (FRLM) proposed
by Kuby and Lim (2005) is a maximum covering location
3
problem where the coverage is in terms of capturing the
vehicle flow associated with the OD trips. The main
assumptions of this formulation includes uncapacitated
stations, deterministic demand, identical vehicles, and
no deviation from the shortest path. FRLM received
many modeling improvements and extensions over the
last decade, including but not limited to Kuby and Lim
(2007), Kuby et al. (2009), Upchurch, Kuby, and Lim
(2009), Capar and Kuby (2012), Capar et al. (2013), Mir­
Hassani and Ebrazi (2013), Arslan and Karaşan (2016),
and Hong and Kuby (2016). The reader can refer to Ko,
Gim, and Guensler (2017) and Shen et al. (2019) for a
comprehensive review on these studies and a synthesis
of the most relevant works that adopt variations, such as
the set-covering approach (e.g., Wang and Lin 2009,
Wang and Wang 2010, Li and Huang 2014).
A key scope of extension to FRLM involves the con­
sideration of allowing shortest path deviations, which
incorporates EV users’ routing choices in the mathemati­
cal models (Kim and Kuby 2012, 2013). Given the addi­
tional computational complexity this extension brings,
the initial computational experiments were performed
on small-sized benchmark networks with 25 nodes. The
computational challenges posed by this extension are
addressed by Yıldız, Arslan, and Karaşan (2016), Arslan
et al. (2019), Göpfert and Bock (2019), and Kınay, Gzara,
and Alumur (2021), and the size of the solvable instances
increased significantly to over 500-node networks. All of
these studies, except Kınay, Gzara, and Alumur (2021),
focus on the traditional max-cover objective and put a
limit on the extent of allowed deviation to maintain a
tractable problem size and to obtain applicable solutions
within the max-cover framework. Kınay, Gzara, and
Alumur (2021) proposes a new full-cover approach and
studies optimal routing of all OD trips by minimizing
the total en route recharging as well as minimizing total
cost of locating charging stations. The authors show that
this approach is superior to the max-cover one in terms
of routing solution quality even when a deviation toler­
ance is not involved.
None of these formerly mentioned studies consider
a capacitated framework; thus, they neglect EV users’
route choice response to charging station locations or
congestion. On the other hand, the studies that model
the capacitated extensions of the charging station loca­
tion problem (e.g., Upchurch, Kuby, and Lim 2009;
Jiang et al. 2012; Hosseini, MirHassani, and Hoosh­
mand 2017) adopt approaches that impose hard limits
on the number of vehicles that can be served by each
open station instead of assuming that all flow passing
through a station can be served regardless of their
volume. Additionally, there are studies that consider
capacity-level expansions under a multiperiod frame­
work. Among these, Zhang, Kang, and Kwon (2017)
extends the capacitated FRLM considering the demand
dynamics (increasing EV market share) throughout the

4 Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
planning horizon. Anjos, Gendron, and Joyce-Moniz
(2020) uses a hybrid approach that takes both node-
based (urban) and flow-based (long-distance) travel
demands into account.
There are stochastic modeling approaches that use
queuing theory to determine station capacity levels. Xie
et al. (2018) and Xie and Lin (2021) use M/M/c queuing
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
systems to model facility congestion under the context
of multiperiod capacity expansion and charging station
location problem. The former proposes a stochastic
optimization model that minimizes total system cost
defined as the sum of the capital cost of locating and
expanding charging stations and the penalty cost of not
satisfying (covering) an OD trip. The deviations from
the shortest paths are taken into account by means of
pregenerating a number of alternative paths for each
OD trip. Even though deviations are considered, all
alternative paths are assumed to be equivalent for the
EV users (i.e., EV users’ responses to charging station
locations are neglected). The resulting mathematical
model is solved by a genetic algorithm under the
assumption of last-minute charging, which may cause
suboptimal demand allocation to charging stations. The
computational experiments are based on a case study in
California. Xie and Lin (2021) extends this framework
by introducing an inconvenience cost function of trave­
lers in their system cost definition. This model is also
solved by a genetic algorithm, but it no longer allows
deviations from the shortest paths.
There are several studies that take the route choice
behavior of EV users into account using bilevel location
models under an uncapacitated framework (e.g., Jing
et al. 2017; Zheng et al. 2017; Guo, Yang, and Lu 2018;
He et al. 2018; Tran et al. 2021). These studies are more
suitable for an urban setting as they consider traffic
congestion using stochastic or deterministic user equi­
librium. Hence, the case studies and numerical tests fea­
tured in these works cover smaller geographical regions.
These mathematical models rely on path pregeneration
or enumeration, except for Zheng et al. (2017), in which
the state of charge (SoC) is tracked in the model to
ensure path feasibility. Bilevel optimization problems
are known to be intrinsically hard to solve. Even the
models with both linear leader and follower’s problems,
which are generally the simplest to solve, are shown to
be strongly NP hard (Labbé and Marcotte 2021). Typi­
cally, solution methods used for these studies are meta­
heuristics, such as a genetic algorithm or large-scale
neighborhood search, or a single-level reformulation of
the bilevel problems is proposed to be able to solve the
models using commercial solvers.
To the best of our knowledge, there are only two
studies to date that make charging station location and
sizing decisions while taking EV users’ response into
account. Huang and Kockelman (2020) use a bilevel
profit maximization problem for multiperiod charging
Kınay, Gzara, and Alumur: Charging Station Location and Sizing
station location and sizing. Their model allows for incor­
porating charging price elasticity and station congestion
into a discrete set of route choices for EV users under
elastic demand. However, station capacity deployment
is still modeled in a deterministic way. The numerical
experiments are performed on an urban network with
74 nodes. A genetic algorithm is used to find solutions
to the proposed model. This algorithm uses a queuing
averse logit choice model that determines users’ alloca­
tions to charging stations. Makhlouf, Kchaou-Boujelben,
and Gicquel (2019) also propose a bilevel problem where
the upper-level problem is a max cover-type station loca­
tion and sizing problem and the lower-level problem
represents EV-user behavior in terms of making the min­
imum number of stops to reach their destination. The
solution approach is based on a single-level reduction
assuming cooperative response, and they solve the re­
sulting reformulation using commercial solvers on ran­
domly generated network instances of 100 nodes.
To sum up, none of the studies consider a bilevel
framework that simultaneously takes the infrastructure
cost of locating and sizing charging stations, route choice
response of EV users, and stochastic waiting times at
facilities into account. Moreover, the existing studies
that use a bilevel model assume only a cooperative user
response, whereas we develop a solution methodology
that can be tailored for both the cooperative and unco­
operative responses of the follower. Our algorithm is
capable of solving large-scale transportation networks
with up to 420 nodes, which is the largest network size
that has been solved within a bilevel framework in the
literature.
3. The Bilevel Charging Station Location
and Sizing Problem
Let the underlying transportation network be defined
as Go � (N, Ao ), where N denotes the set of nodes (e.g.,
highway intersections and population centers) and Ao
denotes the set of directed arcs (e.g., highway segments).
Consider K as the set of OD pairs, where Ok and Dk
denote the origin and destination nodes for OD pair
k ∈ K. Each OD pair represents a stochastic stream of EV
flow between two population centers for which the
travel distance exceeds the EV battery range R, and thus,
recharging en route is necessary.
A charging station may host up to C charging stalls,
and each stall offers stochastic service times. As a result
of limited capacity, stochastic arrivals, and stochastic ser­
vice times, charging stations may experience congestion.
Consequently, EV users may have to enter a service
queue when the system is busy (i.e., when all charging
stalls are occupied).
The waiting time at a charging station is a function of
its capacity given by the number of charging stalls and
the total allocated EV charging demand. In order to
 Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
ensure service convenience, a service-level requirement
is introduced to guarantee that the probability of wait­
ing at most α minutes is greater than or equal to β% at
each open station. For example, a service-level require­
ment is for the waiting time to be less than or equal to
α � 10 minutes, β � 90% of the time.
A deviation tolerance is defined to model EV users’
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
willingness to deviate from their shortest path to charge.
This deviation tolerance τ is expressed in terms of the
percentage difference of the length of the path taken
and the length of a shortest path (e.g., Kim and Kuby
2012; Yıldız, Arslan, and Karaşan 2016; Arslan et al.
2019; Göpfert and Bock 2019).
Let f (·) denote the fixed cost function of establishing a
charging station and g(·) denote the variable cost func­
tion of installing charging stalls at the stations. The for­
mer is typically a function of economic characteristics of
the land of each candidate station location, whereas the
latter is a nondecreasing function of the number of
charging stalls installed at a station.
The bilevel optimization model minimizes the total
infrastructure cost of locating charging stations and
installing charging stalls while ensuring feasible OD
routes and probabilistic service-level requirements. Fea­
sible OD routes visit en route charging stations to satisfy
range limitations of EVs and are within the respective
deviation tolerance.
Before presenting the model, we introduce the net­
work transformation that allows us to model OD trip
feasibility without tracking the state of charge of the
vehicles or pregenerating all the feasible stopping com­
binations. The expanded network introduced by Mir­
Hassani and Ebrazi (2013) defines additional arcs on Go
between nodes i, j ∈ N, given that the shortest travel dis­
tance between them is less than or equal to the EV
range. This allows for embedding the EV range infor­
mation in feasible network arcs because if the shortest
path between any two nodes can be completed without
recharging, a stop for a recharge is not required at inter­
mediary nodes. Although this network expansion is
first proposed for given paths (MirHassani and Ebrazi
2013), Kınay, Gzara, and Alumur (2021) showed that it
Figure 1. (Color online) An Illustration of Permissable Arc Expanded Network Generation with R � 200 and τ � 10%
(a)
5
may be used when shortest path deviations are allowed.
Accordingly, we define G � (N, A) to be the expanded
network of Go for a given range R. We now show how
to further reduce the expanded network by defining a
permissible arc for OD pair k ∈ K as follows.
Definition 1. An expanded network arc (i, j) ∈ A is
called a permissible arc for OD k ∈ K if it connects Ok to
Dk within the allowed deviation tolerance. The set of
permissible arcs Ak, k ∈ K is given by
Ak � {(i, j) : (i, j) ∈ A and
δOk i + δij + δjDk ≤ (1 + τ)δOk Dk }, k ∈ K,
where δij is the shortest path length from i ∈ N to j ∈ N.
We then construct Gk � (Nk , Ak ) as the permissible
arc expanded network for k ∈ K, where Nk consists of
the nodes that are incident to the arcs in the set Ak . This
reduced set Ak allows us to define the mathematical
model concisely by removing redundant constraints
because of arcs that cannot be in a feasible path within
the deviation tolerance. However, modeling the prob­
lem on ∪k∈K Gk rather than G does not ensure that the
deviation tolerance requirement is met.
Figure 1 illustrates the permissible arc expanded
network generation on an example. In this figure, the
shortest path between Ok and Dk uses nodes 3 and 4,
and δOk Dk � 250 as shown with bold green arcs in
Figure 1(a). With R � 200 and τ � 10% and assuming
that the initial state of charge is 100%, the dashed red
arcs in Figure 1(b), (Ok , 2) and (3, Dk ), are in Ak because
the shortest distances between respective nodes are
less than R. Specifically, the corresponding subpaths
can be completed without recharging, and the paths
Ok → 2 → Dk and Ok → 3 → Dk are both within the
deviation tolerance. On the other hand, even though
arc (1, 4) has a distance of 195, which is below R, it is
not a permissible arc as the shortest length of the path
that uses this arc is 295 and beyond the 10% deviation
tolerance.
The structure of such an expanded network is versa­
tile enough to embed the fixed cost of charging stops to
(b)
 Kınay, Gzara, and Alumur: Charging Station Location and Sizing
6 Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS

the expanded network arc costs. To achieve this, the
cost of expanded network arcs that do not enter the des­
tination node would be increased by the fixed cost of
charging.
To formulate the bilevel charging station location and
sizing problem under congestion, we introduce the fol­
lowing decision variables:
8
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
context, the leader makes a decision, and the follower
responds. The framework of such a problem is based
on the fact that the leader anticipates the follower’s
optimal reaction to their decisions.
The objective function of the leader (1) minimizes
total infrastructure cost composed of the cost of locat­
ing charging stations and installing charging stalls.

1, if candidate location j ∈ N Constraint (2) ensures that station capacity may be

>
< nonzero only for open charging stations. Chance con­
yj � has a charging station, straints associated with service-level requirements are
>
:
0, otherwise: defined by (3). For each charging station, this con­
straint ensures that the probability of waiting less
zj � number of charging stalls installed at
than or equal to α minutes is greater than or equal to
charging station j ∈ N: β. Constraints (4)–(6) define the domains of the lea­
8 der’s decisions.
>
> 1, if expanded arc (i, j) ∈ Ak is
>
< The objective of the follower’s problem is to mini­
xijk � traversed by OD pair k ∈ K, mize the total travel distance. Constraint (7) defines
>
> the feasible solution space of the follower’s routing
> 0, otherwise:
:
decisions in response to the leader’s charging station
Wj � waiting time at charging station j ∈ N: location decisions, which in itself is an optimization
problem. Constraints (8) are flow balance constraints
The bilevel model BLP is then formulated as follows: for OD k ∈ K. Constraint (9) ensures that there exists a
X charging station, located by the leader, at the head of
[BLP] minimize (f (yj ) + g(zj )) (1) each used arc, except the final arc of each OD path. By
j∈N
the definition of the expanded network, this ensures
s:t: zj ≤ Cyj j∈N (2) vehicle range feasibility besides determining which
charging stations to use on each OD path. Constraint
P(Wj ≤ α) ≥ β j∈N (3) (10) makes sure that each OD path is within the devia­
tion tolerance τ. Lastly, Constraint (11) defines the
yj ∈ {0, 1} j∈N (4)
domain of the flow variables xijk.
There is a lack of clarity in (1) in case there are alter­
zj ≥ 0 and integer j∈N (5)
native optimal solutions for the follower’s problem for
Wj ≥ 0 j∈N (6) a given solution vector of the leader’s problem. This
is resolved by defining the notions of cooperative and
X X
xijk ∈ argmin δij xijk (7) uncooperative responses of the follower. In a coop­
k∈K (i, j)∈Ak erative response, the leader anticipates the follower to
s:t: choose, among the optimal solutions of the lower-level
8 problem, the one that yields the best objective value for
< 1,
> if i � Ok ,
X X the leader’s problem (i.e., the follower cooperates with
xijk � xjik � �1, if i � Dk , the leader). In the latter, the leader assumes a reaction
j:(i, j)∈Ak j:(j, i)∈Ak
>
:
0, otherwise: of the follower that generates the worst possible out­
i ∈ Nk , k ∈ K (8) come for the leader’s objective function.
X In the next section, we mathematically define both
xijk ≤ yj j ∈ Nk , j ≠ Dk , k ∈ K (9) response types of the follower and then, in Section 4.1,
i:(i, j)∈Ak show how BLP may be reformulated as a single-level
X problem under the cooperative response. In Section 4.2,
δij xijk ≤ (1 + τ)δOk Dk k ∈ K (10) we introduce the characterization of a DC fast-charging
(i, j)∈Ak station network as an open multiserver Jackson net­
work, and using the Erlang-C function and queuing sys­
xijk ∈ {0,1} (i,j) ∈ Ak , k ∈ K: (11)
tem stability constraints, we present the deterministic
Bilevel optimization problems are often interpreted as equivalents of the probabilistic service-level constraints
static and open-loop Stackelberg games that are struc­ and modify the single-level reformulation of BLP under
tured to have an upper-level (leader’s) problem and a cooperative response as a mixed-integer linear optimi­
lower-level (follower’s) problem (Bard 1991). In this zation problem (MILP).
 Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS 7

4. Analysis of the Bilevel the integrality property of the follower’s problem under
Modeling Framework cooperative response.
Let Ψ(y) define the reaction set of the follower based on Proposition 1. The follower’s problem of BLP always has
the leader’s decisions y � {yj , j ∈ N} as an integer optimal solution when feasible.
(
Proof. Assume that there exists an OD pair k ∈ K with
Ψ(y) � x � {xijk , (i, j) ∈ Ak , k ∈ K} : fractional xijk in an optimal solution. Without loss of
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.

( )) generality, let p1 and p2 be two, not necessarily disjoint,

X X paths with distances z1 and z2, respectively. Let χ1 and
x ∈ argmin δij xijk : (8)–(11) :
k∈K (i, j)∈Ak
χ2 denote the fractional flow on these paths, such that
χ1 + χ2 � 1 by Constraint (8). By the optimality of this
The optimal solution of BLP under cooperative (optimis­ solution, z1 � z2. Otherwise, one may construct a solu­
tic) response xo is tion that only uses path p∗ � argmin{z1 , z2 } with objec­
8 9 tive function value min{z1 , z2 } ≤ z1 χ1 + z2 χ2 . Because
<X = z1 � z2, an integral solution may be obtained by selecting
xo � argmin (f (yj ) + g(zj )) : xo ∈ Ψ(y) : one of the equal-length paths arbitrarily and setting the
: j∈N ;
corresponding xijk values to one on this path and to
The uncooperative position reflects the case where the zero on the rest. w
leader protects the leader’s self against the worst possi­ Proposition 1 means that the domain Constraint (11)
ble outcome of the follower’s response. Based on the def­ of the follower’s problem may be dropped, and its opti­
inition of the follower’s alternative reaction set Ψ(y), the mal solution is characterized by the optimality condi­
optimal uncooperative (pessimistic) response solutions tions. Let θjk, γjk, and φk be the dual variables associated
of BLP are defined as follows: with (8), (9), and (10), respectively. The single-level refor­
8 9 mulation of BLP, denoted by SLP, is
<X =
xp � argmax (f (yj ) + g(zj )) : xp ∈ Ψ(y) : [SLP] minimize (1)
: j∈N ;
s:t: (2)–(6), (8)–(11)
X
For the case of cooperative response and in the presence δij xijk � (θOk k � θDk k )
of convex follower’s problems, it is possible to reformu­ (i, j)∈Ak
late a bilevel optimization model into a single-level one X
+ yj γjk + δOk Dk (1 + τ)φk k ∈ K
using optimality conditions of the follower’s problem
j∈Nk
(Bard 2013).
Uncooperative response is known to be less tractable (12)
compared with the cooperative response even in the case θik � θDk k + δiDk φk ≤ δiDk k ∈ K, (i, Dk ) ∈ Ak (13)
of a convex follower’s problem. In fact, pessimistic bilevel θik � θjk + γjk + δij φk ≤ δij k ∈ K, (i, j) ∈ Ak , j ≠ Dk
optimization problems, so far, do not have any computa­
(14)
tional methods that are generally applicable (Zeng 2020).
Commonly, solving these problems requires processing θjk free j ∈ Nk , k ∈ K (15)
all alternative optimal solutions of the follower’s problem γjk ≤ 0 j ∈ Nk , k ∈ K (16)
to determine the solution that is the worst for the leader’s φk ≤ 0 k ∈ K: (17)
objective function, which is deemed not tractable (Sinha,
Malo, and Deb 2017). In this reformulation, (12) is the strong duality condi­
We study both the cooperative and uncooperative tion of the follower’s problem, and its dual feasibility
responses in this study by developing the single-level requirements are defined by Constraints (13)–(17). Note
reformulation of BLP under the cooperative response that (12) is written for k ∈ K because the dual problem is
case as well as providing a solution algorithm to solve decomposable for each OD pair. This constraint includes
BLP for the uncooperative response case and M/M/c a quadratic term that can be linearized by introducing
queues. auxiliary decision variables qjk � yj γjk , j ∈ Nk , k ∈ K. Ac­
cordingly, Constraint (12) is linearized by replacing it
with (18)–(21):
X X
4.1. Single-Level Reduction of BLP Under δij xijk � (θOk k � θDk k ) + qjk + δOk Dk (1 + τ)φk
Cooperative Response (i, j)∈Ak j∈Nk
In this subsection, we show that the follower’s problem
k∈K (18)
always has an integer optimal solution when feasible and
then, reformulate BLP into an MILP taking advantage of qjk ≥ γ jk yj j ∈ Nk , k ∈ K (19)

8 Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
qjk ≥ γjk j ∈ Nk , k ∈ K (20)
qjk ≤ γjk � (1 � yj )γ jk j ∈ Nk , k ∈ K (21)
qjk ≤ 0 j ∈ Nk , k ∈ K, (22)
where γ jk denotes the lower bound for the nonpositive
dual variable γjk. Constraint (22) defines the domain of
γ jk as nonpositive, and (19) ensures that γjk � 0, when
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
yj � 0. On the other hand, Constraints (20) and (21) en­
sure that qjk � γjk , when yj � 1.
The reformulation of BLP into a single-level mixed-
integer problem is a result of Proposition 1, which states
that the lower-level problem may be solved as a linear
optimization. This property of the follower’s problem
holds because of its objective function, which optimizes
the length of the path. Understandably, one may con­
sider other plausible objectives for the follower, like
minimizing the number of stops or minimizing the
waiting time for charging. These objectives are interest­
ing extensions of this work in which a linearization of
BLP may not be possible and changes to the solution
method are required.
Both BLP and SLP include variable Wj, which is a
function of xijk and zj as well as demand and service
rates. The characterization of Wj and the associated
probabilistic service-level Constraints (3) depend on the
underlying queuing system at the charging stations.
Consequently, the solvability of SLP still depends on
the ability to characterize chance Constraints (3). Next,
we discuss the case of Poisson arrivals and exponential
service times, which allows the reformulation of (3)
using mixed-integer optimization.
4.2. Case of Poisson Arrivals and Exponential
Service Times: M/M/c Queuing System
DC fast-charging stations for EVs typically include
multiple charging stalls that can deliver 80–400 kW at
50–1,000 V (IEA 2021). Each charging station provides
simultaneous service to multiple EVs, where each stall
acts as a server utilized by service seekers on a first
come, first served basis. Moreover, common courtesy
dictates that there is a single queue for each charging
station in case of concurrent utilization of the servers.
This single-queue, multiple-parallel server structure
constitutes the main framework of our queuing system
models of the DC fast-charging stations.
Existing public charging stations are often located
near discretionary service facilities (e.g., restaurants,
stores, plazas, or shopping centers) in order to mitigate
the relatively longer recharging times for EVs. As a
result, not every customer unplugs their EV right after
getting up to their desired SoC. This is one of the main
reasons for existing studies to consider stochastic and
typically, exponential service times (Xie et al. 2018, Xie
and Lin 2021). The stochastic flow rate of an OD is
assumed to have Poisson distribution. The rate of the
Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Poisson process may be determined as a function of
the population of two cities, their respective EV adop­
tion rates, and the travel distance between them. For
long-distance trucking, for example, it may be a func­
tion of the number of ETs dispatched per day from an
origin to a destination.
Under a single-queue, multiple-parallel server, Pois­
son arrivals, and exponential service time system, each
charging station may be modeled as an M/M/c queue.
However, a long-distance OD route might require mul­
tiple charging stops, and ODs may have varying flow
rates, resulting in a queuing network.
Burke’s theorem states that the departure rate of a sta­
tionary M/M/1 queue is identical to its arrival rate as
long as the queuing system is stable (Burke 1956). This
concept of Poisson flow conservation is also general-
izable to M/M/c queues, which allows the characteri­
zation of a queuing network that comprises several
interconnected M/M/c queues (Bose 2013). Using these
results, an arbitrary transportation network with M/
M/c charging stations is an open multiserver Jackson
network for which the routing probabilities are defined
as the probability that a vehicle leaves one node to go to
another. In the context of our problem, these routing
probabilities are directly proportional to the Poisson
arrival rates of the OD pairs that use them and set by
the route choices of EV users.
We define λk to denote the Poisson flow rate of OD
pair k ∈ K. Let Λj denote the flow rate faced by open sta­
tion j ∈ N, determined as the total Poisson OD demand:
X X
Λj � λk xijk j ∈ N: (23)
i:(i, j)∈A k∈K
Recall that the expanded network ensures that xijk � 1 of
OD k ∈ K charges at j ∈ N. Moreover, the underlying
open multiserver Jackson network ensures that the
arrival rate of an OD to a station is equal to their depar­
ture rate as long as all queues on the network are stable.
The queue stability requirement is defined as
Λj ≤ µzj � ɛ j ∈ N, (24)
where µ is the exponential service rate of each server of
a station and ɛ is an infinitesimal constant. This con­
straint ensures that the Poisson flow rate faced by a sta­
tion is strictly less than its processing rate in order to
have a finite queue length.
We define the average load of a stall at a charging sta­
tion by ρj � Λj =µzj and rewrite the left-hand side of Con­
straint (3) using a modification of the Erlang-C formula
(Chromy, Misuth, and Weber 2012; Xie et al. 2018):
z
(zj ρj ) j
zj !(1�ρj )
P(Wj ≤ α) � 1 � z
Pzj �1 zj ρj t e�µ(zj �ρj zj )α j ∈ N:
(zj ρj ) j
zj !(1�ρj ) + t�1 t!
(25)
 Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS 9

Although this substitution in (3) involves a quadratic DLP may be solved using off-the-shelf solvers, our pre­
function, it may be rewritten using a step function on liminary computational experiments revealed that these
integer zj: solvers are not able to handle the problem complexity
for large-sized networks. In order to solve DLP for real-
8
>
> 0, Λj � 0 life instances, we develop a decomposition-based algo­
>
> rithm that exploits the problem structure as described
> ɛ ≤ Λj < b1
< 1,
>
in the next section.
zj � 2, b1 ≤ Λj < b2 if P(Wj ≤ α) ≥ β
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.

>
>
>
> ::: :::
>
>
:
5. Solution Methodology
C, bC�1 ≤ Λj < bC : There are three decisions to make in our problem set­
(26) ting: the locations of the charging stations, the number
of stalls at open stations, and the OD routes. We pro­
Service-level parameters α and β of Constraint (3) deter­ pose a solution method that decomposes these decisions
mine the values of the break points (b1 , b2 , : : : , bC ) of this and iteratively solves two single-level problems: a loca­
step function by a preprocessing procedure based on tion problem and a sizing problem. As the location
(25). Such a function can be easily incorporated within problem turns out to be computationally demanding,
any MILP model using special ordered set of type 1 vari­ we develop a tailored logic-based Benders algorithm
ables or using the following set of additional constraints (LBBA) that solves all instances within seconds.
to substitute (3): The proposed decomposition structure first solves a
location problem taking into account OD flows to obtain
X
C
the charging station location vector y and the corre­
sjm � 1 j∈N (27)
m�1
sponding optimal reaction set of the follower defined
on y:
X
C
( (
Λj ≤ bm sjm j∈N (28) X X
m�1 Ψ(y) � x : x ∈ argmin δij xijk :
k∈K (i, j)∈Ak
X
C ))
Λj ≥ bm�1 sjm j ∈ N (29) (8), (10), (11), (32) :
m�2

X
C
X
zj ≥ msjm j∈N (30) where xijk ≤ y j j ∈ Nk , j ≠ Dk , k ∈ K: (32)
m�1
i:(i, j)∈Ak

sjm ∈ {0, 1} j ∈ N, m ∈ {1, 2, : : : , C}:
Constraints (28)–(29) determine the two consecutive
break points of the step function that Λj value lies
between. Constraints (30) exploit the fact that both m
and zj are integer and determine the minimum zj value
in conjunction with the minimization objective function
(1) of BLP.
The number of sjm variables increases with C, so
choosing the right value of C is important. In this study,
we determine this value so that it is large enough to
keep Constraint (2) nonbinding for the optimal solution
but small enough so that the number of binary sjm vari­
ables is not excessively many.
Substituting a step function instead of the chance
Constraint (3) by pregenerating its break points based
on the Erlang-C formula and linearizing this step func­
tion with the constraints, the single-level reduction of
the bilevel model is provided in Online Appendix A.
To summarize, when OD flows are Poisson distrib­
uted and service rate at a stall is exponential, SLP is
reformulated as an MILP denoted as DLP. Even though
(31) For given y and Ψ(y), the sizing problem reduces to
determining the number of stalls at open stations (z)
and allocating charging demand of OD flows to open
stations (x). The sizing problem accounts for route
choice response by considering all alternative optimal
paths on the network defined by y and modeling both
cooperative and uncooperative response cases. The siz­
ing problem determines a partial solution (z, x) that is
optimal for given y and a complete solution (y, z, x)
that is bilevel feasible. The flowchart of the solution
methodology is shown in Figure 2. At each iteration,
an elimination cut is generated after solving the sizing
Figure 2. The Flowchart of the Decomposition-Based Solu­
tion Method for BLP

10
problem and is fed back into the location problem. Sim­
ply put, this algorithm performs a smart enumeration
by iteratively constructing bilevel feasible solutions and
tracking the incumbent. When stopped prematurely, it
has no guarantee of optimality.
The rest of the section is organized as follows. Section
5.1 describes the location problem and the suggested
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
logic-based Benders algorithm. Section 5.2 models the
sizing problem under M/M/c and provides the com­
plete decomposition-based algorithm for the coopera­
tive response in Section 5.2.1 as well as an adaptation of
the algorithm for the uncooperative response case in
Section 5.2.2.
5.1. Solution of the Location Problem
When the sizing decisions and service-level Constraint
(3) are dropped, the follower’s choice of route will not
affect the leader’s objective. Consequently, the remain­
ing problem reduces to a single-level location problem
and determines the minimum cost charging station
locations subject to flow balance and deviation tolerance
constraints. The location problem PL is
X region of MP and (ii) do not eliminate any feasible solu­
[PL ] minimize f (yj ) (33)
j∈N
s:t: (4), (8)–(11):
PL may be solved directly by a solver or using Benders
decomposition, where the master problem determines
charging station locations and the subproblems check
path feasibility for each OD. Recall that the subpro­
blems are linear because of Proposition 1.
Despite being a relaxation of BLP, preliminary testing
showed that PL is still challenging to solve in reasonable
times both using the solver and using Benders decom­
position. Because PL is solved repeatedly within the
solution method of BLP, we develop a logic-based
Benders algorithm that is able to solve PL optimally
within a few seconds for the majority of the large-scale
instances. The logic-based Benders master problem
(MP) is given by
X cases, feasibility cuts are generated and added to MP.
[MP] minimize f (yj ) (34)
j∈N
s:t: yj ∈ {0, 1} j∈N (4)
+ logic-based Benders cuts,
and the subproblem is a path feasibility problem with
deviation constraints on the network defined by y, the
solution of MP.
We further modify the subproblems in two ways. First,
we drop the deviation tolerance constraints and deal
with them using feasibility cuts. Second, we keep the
objective function of the follower and solve a minimum
cost path problem instead of a feasibility problem. It is
Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
redundant to say that this does not impact the correctness
of the logic-based Benders algorithm, even though it may
lead to an alternative optimal solution for PL. However,
the objective function of the follower becomes critical
within the solution methodology of BLP, where it is nec­
essary to account for their minimum-length path choice.
The subproblem SPk for OD k ∈ K is given by
X
[SPk ] minimize δij xijk (35)
(i, j)∈Ak
8
< 1,
> if i � Ok ,
X X
s:t: xijk � xjik � �1, if i � Dk , i ∈ Nk
j:(i, j)∈Ak j:(j, i)∈Ak
>
:
0, otherwise:
(36)
X
xijk ≤ y j j ∈ Nk , j ≠ Dk (37)
i:(i, j)∈Ak
xijk ∈ {0, 1} (i, j) ∈ Ak : (38)
When a solution y of MP is infeasible to SPk, the aim is
to derive valid cut(s) that (i) remove y from the feasible
tion of PL. The valid cuts are added to MP, and the latter
is solved repeatedly to find a new set of station loca­
tions. When the valid cuts satisfy the two conditions,
this approach is guaranteed to converge to an optimal
solution (Hooker and Ottosson 2003). We next discuss
the derivation of the valid cuts and prove the correct­
ness of the algorithm.
Once a solution y is obtained, the node restrictions
enforced by Constraint (37) are used to form a reduced
graph Grk � (Nkr , Ark ) for each OD pair k ∈ K such that
Nkr � {j ∈ Nk |y j � 1} ∪ {Ok , Dk } and Ark � {(i, j) ∈ Ak |i, j ∈
Nkr }. Subproblems SPk, ∀k ∈ K, are solved on their re­
spective reduced graphs Grk using a labeling algorithm.
A given subproblem SPk may be feasible, yet the opti­
mal path length may not satisfy the deviation tolerance
constraints of the original problem. Alternatively, a
given subproblem SPk may be infeasible if there is no
path that connects OD in the reduced graph. In both
We first discuss the case where SPk is infeasible for
some OD k̂ ∈ K. In this case, the corresponding graph
Grk̂ has at least two connected components that sepa­
rately include Ok̂ and Dk̂ . Let Gok̂ be the connected com­
ponent of Grk̂ that includes Ok̂ and Gdk̂ be the reverse
graph of the component of Grk̂ that includes Dk̂ . Let Arev k
denote the reverse Ak. We define N f as the set of nodes
that connect Gok̂ to Gdk̂ and N b as the set of nodes that
connect Gdk̂ to Gok̂ :
N f � {j ∈ Nk̂ \Nk̂r : (i, j) ∈ Ak̂ , i ∈ Nk̂o } (39)
N b � {j ∈ Nk̂ \Nk̂r : (i, j) ∈ Arev , i ∈ Nk̂d }: (40)
k̂
 Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
Using N f and N b , we introduce the following feasibility
cuts for an infeasible SPk̂ :
X tion tolerance. It remains to determine the number of
yi ≥ 1 (41)
i∈N f
X requirement, which we refer to as the sizing problem.
yi ≥ 1:
i∈N b
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
Cuts (41) and (42) are added to MP to cut the current
infeasible solution. Note that adding either of them is
sufficient to cut the current infeasible solution.
For the case when SPk is feasible and deviation toler­
P always equip each station with a sufficient number of
ance is exceeded for some OD k̂ ∈ K (i.e., (i, j)∈A δij xijk̂ >
(1 + τ)δOk̂ Dk̂ ), we introduce the following cut:
k̂
X vided stall capacity C is sufficiently large. To determine
yi ≥ 1: (43)
i∈Nk̂ \Nr
k̂
Cut (43) is added to MP to eliminate the current solution
that is infeasible with respect to deviation tolerance.
This cut ensures that at least one charging station from
Nk̂ \Nk̂r is open because opening facilities at i ∈ Nk̂r is not
sufficient to yield a feasible solution.
The logic-based Benders decomposition algorithm
solves the relaxed master problem MP and subproblem
SPk, ∀k ∈ K, iteratively and adds feasibility cuts (41),
(42), and (43) to MP as needed. We present the pseudo­
code of the LBBA in Algorithm 1 in Online Appendix D
and prove its correctness in Proposition 2. Note that PL
is feasible as long as there exists y that is feasible for the
follower’s problem of BLP.
Proposition 2. LBBA determines an optimal solution to
PL given that it is feasible.
Proof.
i. It is trivial that cuts (41), (42), and (43) remove a
solution y infeasible with respect to (8) and (43) from
the feasible region of the master problem MP.
ii. To show that cuts (41) and (42) do not remove any
feasible solution to PL, assume to the contrary that there
exists a feasible solution ŷ that does not satisfy (41) or
(42) for an OD pair k̂ ∈ K. Because this solution is feasi­
ble, there exists at least one open charging station at
node iˆ that is connected to Ok̂ . Therefore, iˆ ∈ N f , and ŷ
satisfies (41). This is a contradiction because ŷ satisfies
cut (41) as this specific node is already in the corre­
sponding N f . The same argument follows for Dk̂ , N b ,
and (42). Hence, valid cuts (41) and (42) do not elimi­
nate any feasible solutions.
To show that cuts (43) do not remove any feasible
solution to PL, assume to the contrary that there exists a
feasible solution that does not satisfy this cut. This means
that none of the facilities in Nk \Nkr are open. This creates
a contradiction because opening any subset of facilities
in Nkr cannot improve path lengths, and the solution
remains infeasible. Hence, cuts (43) do not eliminate any
feasible solutions. w
11
The solution of the location problem opens charging
stations and determines OD flow paths within the devia­
charging stalls at each open location to satisfy the service
(42) These two problems are solved iteratively to generate
solutions for DLP. Next, we present the sizing problem
and the complete solution for BLP.
5.2. The Sizing Problem and DLP Solution
When the charging station locations are fixed, one can
stalls to satisfy the service-level Constraints (3), pro­
a solution that minimizes the total cost of capacity
installation, one has to account for the amount of flow
through each station, which in turn, depends on the
shortest OD paths and the stopping response of EV
users when there are alternative stopping options on
their shortest path(s). Consider the example in Figure 3.
This path has three open charging stations, represented
with square nodes. Artificial expanded network arcs,
represented with dashed red lines, indicate that rechar­
ging at any of the three open stations results in a feasible
trip. There are seven alternative shortest paths for this
OD, each corresponding to different feasible stopping
combinations. Each shortest path may lead to different
total flows through the stations and, consequently, dif­
ferent requirements in the number of stalls. Hence,
to model the sizing problem correctly, we have to con­
sider all possible stopping combinations on the network
defined by y.
The set of alternative solutions for OD pair k ∈ K
determines its optimal reaction set Ψk (y), where Ψk (y)
is populated using a shortest path algorithm on Grk . We
next discuss how to formulate the sizing problem and
present the decomposition-based algorithm under coop­
erative and uncooperative responses of the follower.
5.2.1. The Cooperative Response. Under the coopera­
tive response of the follower, the sizing problem deter­
mines the number of stalls in open charging stations
and allocates charging demand to the stations so that
the service-level requirement is satisfied while minimiz­
ing total stall installation cost. Let ajrk be a binary param­
eter that takes the value of one if an open station j ∈ Nk
Figure 3. (Color online) An Example OD Path with Alterna­
tive Stopping Options

12
is used in path r ∈ Ψk (y) of OD pair k ∈ K, and let urk be
a binary decision variable that takes the value of one if
OD pair k ∈ K is assigned to path r ∈ Ψk (y). The sizing
problem PS under the cooperative response and M/M/c
case is
X the uncooperative response or the worst possible out­
[PS ] minimize g(zj ) (44)
j∈Nk | y j �1
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
X problem PS is modified as follows. First, the objective
s:t: urk � 1 k∈K
r∈Ψk (y)
X X X outcome of the uncooperative EV user response defined
Λj � λk ajrk urk
k∈K i:(i, j)∈Ak r∈Ψk (y)
j ∈ Nk | y j � 1
(5),(27)–(31)
urk ∈ {0,1} r ∈ Ψk (y), k ∈ K:
The objective function (44) minimizes the total cost of
installing charging stalls at open stations. Constraint
(45) ensures that each OD pair is assigned to exactly one
of the alternative paths in its optimal reaction set. Con­
straint (46) calculates the total allocated demand to a
charging station. Constraints (27)–(31) ensure that the
required capacity is calculated based on the lineariza­
tion of the Erlang-C function as discussed in Section 4.2.
Finally, Constraints (5) and (47) define the domains of
the capacity variables zj and the allocation variables urk,
respectively. Although not decomposable by k ∈ K, pre­
liminary computational experiments showed that PS
may be solved within seconds using the commercial
solver Gurobi for sufficiently large instances.
Upon solving PS, we update the solution x to reflect
the selected paths determined by optimal urk. The par­
tial solution (x, z) is optimal for a given y, and solution
(y, x, z) is feasible to DLP. The algorithm to solve DLP
proceeds by generating elimination cuts (48) to cut the
current solution and generate new ones:
X more than the minimal number of stations on the same
yj ≥ 1:
j:{j∈N | y j ≠1}
When appended to MP, this elimination cut ensures
that the current set of open facilities is never generated
again. The complete algorithm for DLP is depicted in
Algorithm 2 in Online Appendix D. Step 3 utilizes
LBBA to determine a minimal cost, feasible set of charg­
ing station locations. Steps 4–6 are the solution proce­
dure for PS, and steps 7–10 are for incumbent tracking.
Finally, step 11 adds cuts (48) to MP to eliminate the
current solution from the feasible region of PL. In the
computational experiments, an iteration of the algo­
rithm refers to steps 3–11, and the algorithm terminates
using time or iteration limits. We refer to this algorithm
as the DA.
Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
5.2.2. The Uncooperative Response. The cooperative
response assumes that EV drivers will use a shortest
path that minimizes the leader’s objective. In practice,
however, it is difficult to enforce such a choice. Conse­
quently, the leader may be interested in investigating
come of the follower’s choices. To account for the un­
cooperative response, the formulation of the sizing
(45)
function is converted to a maximization to reflect the
by xp in Section 4. Next, in order the prevent unbound­
edness, the sign of Constraint (30) is changed to equal­
(46) ity. The rest of PS remains the same, and algorithm DA
can be implemented accordingly.
Upon preliminary computational experiments, it is
observed that this definition of xp generally leads to solu­
(47)
tions where the worst case path is one that uses all charg­
ing stations regardless of their proximity and the need to
charge. This is unlikely to happen in reality because EV
users would generally prefer fewer stops. Although the
follower is not mindful of the leader’s objective in the
uncooperative case, they are self-interested and would
favor stopping fewer times on the same path. As a result,
an uncooperative response solution should exclude un­
justifiable charging stops. This response may be reflected
in the BLP formulation under uncooperative response by
changing Ψk (y). For each OD pair k ∈ K, we modify the
selection of OD paths in the optimal reaction set Ψk (y) to
minimal paths where such a path is one that becomes
infeasible if one charging station is removed. We use this
definition of the uncooperative response in the computa­
tional testing. We consider the minimal paths because
any feasible solution with more than the minimal number
of necessary stops is an alternative optimal solution.
Hence, removing it would not change the optimal objec­
tive function value. The reason for this is the contradic­
tion that if there exists a feasible stopping option with
(48) shortest available route, an additional stop could only
require more or the same level of capacity at charging sta­
tions given that the service-level requirements are kept
the same.
6. Computational Experiments
In this section, we present extensive computational anal­
ysis to assess the performance of the proposed solution
algorithm as well as to derive insights from effects of
the follower’s position, deviation tolerance, and service-
level parameters on optimal solutions. We compare the
quality of the solutions obtained by DA with optimal
solutions found by solving DLP using a commercial
solver on smaller-sized instances. We also derive several
managerial insights from optimal solutions. Additionally,
we carry out experiments on two real-world highway
 Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
networks to assess the proposed solution methodology
by evaluating its computational performance on larger
instances. Experiments are performed using Gurobi
v.9.1.1 with Python API and NetworkX package v.2.5.1
(Hagberg, Swart, and Chult 2008) using up to eight
threads on a Windows OS computer with an Intel
i7-9700K 3.60-GHz processor and 32 GB of RAM.
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
6.1. Data Sets and Instance Generation Specifics
We use three test networks of different sizes, one of
which is a network that we introduce to the literature
based on the highway network of the United States. The
other two are from the literature: namely, a 25-node net­
work and a California road network. The smaller-sized
25-node network, introduced by Simchi-Levi and Ber­
man (1988), is one of the benchmark networks used for
testing flow capturing and refueling station location
problems, which will be referred to as N25. The Califor­
nia road network introduced by Arslan, Yıldız, and Kar­
aşan (2014), on the other hand, is a larger-sized network
that contains 339 nodes, which will be referred to as
CA339. The N25 and CA339 networks are depicted in
Figure 1 in Online Appendix C. The blue nodes on these
networks highlight the potential OD nodes, which cor­
respond to population centers. Even though we test
our methodology on United States-based networks, we
would like to remark that it is implementable for any
highway network that serves origin-destination pairs
with a travel distance greater than the EV range.
We introduce a new transportation network to the lit­
erature, referred to as US-E, based on the highway
network of the United States. This new data set is formed
based on the geographic information system (GIS) shape
files provided by the U.S. Census Bureau, Department of
Commerce, which include all primary roads that cover
the continental United States. US-E contains the eastern
half of this output and has 2,575 nodes and 6,736 arcs. We
simplify the raw GIS data by removing some node clus­
ters that represent highway intersections and many of
the midarc nodes that are used to represent the geo­
graphical direction changes on highway sections. After
these reductions, the resulting network, referred to as
US-E420, contains 420 nodes and 1,276 arcs. The visual
representations of US-E and US-E420 are shown in
Figure 2 in Online Appendix C. The specifications of the
three networks are summarized in Table 1.
Table 1. Specifications of the Three Networks Used in
Computational Experiments
N25 CA339
Number of nodes 25 339
Number of potential OD nodes 25 51
Number of arcs 86 1,234
Mean arc length (km) 4.60 18.96
Standard deviation of arc lengths 1.81 12.96
13
For this computational analysis, we assume that every
EV user starts their journey with 80% of the available
battery range, and they may charge up to 80% SoC level
at a charging station visit. The reason for this assump­
tion is twofold. First, the battery charging process at
public charging stations gets exponentially slower after
80% SoC, which may cause additional congestion. In
practice, such a limit is currently enforced at busy Tesla
Superchargers in the United States (Electrek 2019). Sec­
ond, this assumption regards EV user battery degrada­
tion concerns by continuously maintaining an SoC
above 80%, which is shown to significantly accelerate
degradation (Pelletier et al. 2017).
We test three R values for N25 (15, 20, and 25), two
for CA339 (200 and 250 km), and two for the larger-scale
US-E420 (400 and 600 km) to represent the state-of-the-
art EV and electric truck ranges. We further filter all
possible long-distance trips between any potential OD
nodes for which the EVs need to stop for recharging at
least once based on the selection of R. Each R setting
forms a new problem instance as the number of arcs on
the expanded network changes with this parameter.
Typically, when R increases, the size of the expanded
network also increases.
For the deviation tolerance τ, we test 0%, 10%, 25%,
and 50%. Even though 50% deviation may be considered
unrealistic, high-deviation tolerance instances are known
to require significantly more computational effort and
have been tested as a benchmark. Moreover, being able
to solve deviations up to 50% is important as it makes
our study compatible with autonomous EV travel, where
users may be less concerned about how much they devi­
ate from the shortest paths.
We test 1, 5, 10, and 30 minutes for service-level wait­
ing threshold α and 95%, 90%, and 80% for service-level
probability β. As we do not have access to real-world
OD flow data for the transportation networks used in
this study, we assume that the Poisson OD flow rate λk
is randomly distributed between (0, 2) vehicles per hour
for every trip. Considering that the potential OD nodes
are typically higher-degree nodes of the underlying
graphs and that we take all possible long-distance OD
pairs into account, the aggregate EV flow rate out of an
origin node is observed to be realistic using this scale of
individual random rates. We take the exponential ser­
vice time µ as 30 minutes (Xie et al. 2018, Xie and Lin
2021). Moreover, throughout our computational analy­
sis, we set the fixed cost of building a charging station
as $150,000 and the cost of deploying one DC charging
stall as $75,000 (Nicholas 2019).
US-E420
420 6.2. Computational Performance of the
39 Decomposition-Based Algorithm
1,276 In this section, we first compare the optimal solutions
91.35 under cooperative response found by the solver with the
67.25
solutions of our DA on the benchmark N25 network.

14
Then, we test the computational performance of this algo­
rithm on the large-scale CA339 and US-E420 networks.
For N25, we generate 48 instances for each R value
we test, which add up to a total of 144 test instances.
For each instance, the numbers of all possible long-
distance OD pairs (i.e., the cardinality of set K) are 167,
108, and 57 for R � 15, 20, and 25, respectively. Each
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
instance is solved both with the optimization model
DLP and with DA, and their solutions are compared to
assess the performance of the latter. The stopping con­
dition of the algorithm is set to 60 iterations. Our algo­
rithm is able to find the optimal solutions for all 144
instances. The results of these experiments are detailed
in Table 1 of the online appendix. We further visualize
the average solution time comparisons in Figure 3 in
Online Appendix C.
The solver finds the optimal solutions quicker only
when the deviation tolerance is zero as the size of the
expanded network is smaller because of our introduc­
tion of the set of permissible arcs. Instances with higher
R values are also relatively quicker to solve by the
solver as they include fewer long-distance OD pairs.
On the other hand, time savings by our algorithm are
drastically higher for R � 15 instances with deviation
tolerance greater than 0%. Particularly, the average
solution time obtained with Gurobi for 50% deviation
tolerance instances of R � 15 is over 1,000 seconds,
whereas our algorithm is capable of returning the opti­
mal solution identified by Gurobi even when it runs
less than 4 seconds.
Now, we focus on the large-sized CA339 and US-
420E instances to further evaluate the computational
performance of the algorithm. We generate 32 instances
on the CA339 network at two different R values, 200
and 250 km, with a fixed β of 90%. The R � 200 and
R � 250 instances include 502 and 335 long-distance OD
pairs, respectively. The same 32 configurations as with
CA339 are used to create the US-E420 instances, with
higher values of R (400 and 600 km). For these instances,
we include a trip distance threshold of 1,600 km as those
longer than this threshold would require overnight
stays. This results in totals of 469 and 429 long-distance
OD pairs defined on the US-E420 network for R � 400
and R � 600 instances, respectively.
We first attempted to solve each instance using Gurobi
and were unable to obtain even an incumbent solution
for deviation tolerance levels greater than 0% within a
two-hour time limit. When this time limit is increased,
Gurobi terminates the solution procedure with an out of
memory error because of the size of the branch and
bound search tree. Note that these instances require
over 1 billion decision variables. This indicates that
the state-of-the-art solvers are not able to handle prob­
lem instances on real-world networks. Consequently,
we utilize DA to solve these instances with a stopping
condition of one hour. The results of these extensive
Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
experiments are summarized in Tables 2 and 3 in
the online appendix and are presented in the online
appendix.
The performance of DA is evaluated in terms of the
number of logic-based Benders cuts added (i.e., feasibil­
ity cuts (41) and (42) and deviation tolerance cuts (43))
and the number of iterations completed within the time
limit. The averages of these performance metrics are
depicted in Figures 4 and 5 in the online appendix for
CA339 and US-E420, respectively.
It is observed from these results that the number of
iterations DA can complete in the one-hour time limit
tend to decrease slightly for CA339 instances as devia­
tion tolerance increases. This is because of the decreas­
ing number of required feasibility cuts. Figure 4 in
Online Appendix C also shows that the deviation toler­
ance cuts (43) are utilized just 16 times throughout the
experiments. This is a clear indication that introducing
permissible arcs is effective for generating solutions that
are deviation tolerance feasible.
The expanded networks of US-E420 include over
700,000 arcs for R � 400 and 1 million arcs for R � 600
at high levels of deviation tolerances. This results in
tens of millions of decision variables and constraints
for the bilevel optimization model. When R is lower
and the number of OD pairs is higher, the likelihood
of obtaining an infeasible set of station locations in the
algorithm increases. Consequently, it is inevitable to
add thousands of feasibility cuts even with the effi­
cient approach that we are proposing. In particular,
observe from Figure 5(a) in Online Appendix C that
when R � 400 and the deviation tolerance is 50%, the
algorithm adds over 14,000 feasibility cuts, whereas it
can only complete one iteration. Moreover, the num­
ber of times deviation tolerance cuts are utilized is sig­
nificantly higher for US-E420 compared with CA339
instances.
6.3. Insights from Bilevel Solutions
In this section, we derive insights from the solutions
by evaluating the effects of deviation tolerance, vehicle
range, and service-level parameters under the cooper­
ative response. We first analyze the N25 solutions and
then move on to the large-scale instances.
For all R configurations of the small-sized N25 net­
work, 0% and 10% deviation tolerance solutions turn
out to be identical. Moreover, the number of located
stations does not change with varying α or β values for
a fixed deviation tolerance. This indicates that the ser­
vice level only affects the capacities of the stations for
the tested configurations. In all instances, stricter α
and
β result in installing more capacity and hence, a higher
objective function value. For example, when R � 15 and
the deviation tolerance is 25%, the α � 1 minute and β �
95% combination opens 31% more charging stalls com­
pared with the case where α � 30 minutes and β � 80%.
 Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
Higher-deviation tolerance solutions require fewer sta­
tions and/or fewer stalls and result in a lower objective
function value (e.g., when R � 15, 0% deviation toler­
ance solutions utilize eight stations and an average of
136 charging stalls in total, whereas 25% deviation toler­
ance solutions require six stations and 132 charging
stalls on average). The value of using a bilevel frame­
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
work is higher when solving higher-deviation tolerance
instances because the solution space of the follower’s
route choice response grows larger. Within the bilevel
framework, the follower picks the shortest available
route, and the average deviation from the shortest paths
never exceeds 5.65% among the optimal solutions of all
the 50% deviation tolerance instances.
Next, we analyze CA339 and US-E420 solutions ob­
tained by DA to derive insights from sensitivity analysis
on the problem parameters. For CA339 solutions, similar
to the N25 results, varying α only changes the capacity
levels deployed but not the number of located stations.
As the geographical shape of the CA339 network is nar­
row and potential OD nodes are not scattered to the
boundaries, between two and seven stations are observed
to be able to sufficiently serve various parameter settings
that we tested on this network.
In Figure 6 in Online Appendix C, we further analyze
the difference in the objective values based on the wait­
ing time threshold and deviation tolerance of CA339
solutions. These charts demonstrate how the objective
value decreases with higher deviation tolerance (τ) and
increases with stricter waiting time threshold (α). The
most significant cost savings are achieved when 0%
deviation tolerance is increased to 10%.
Figure 7 in Online Appendix C shows the average and
maximum deviation metrics at 10%, 25%, and 50% devia­
tion tolerance levels for α � 10 minutes instances of CA339.
For R � 250, average deviation does not exceed 0.88%
even when the deviation tolerance is set to 50%. However,
this performance metric can be as high as 7.07% at 50%
deviation tolerance for R � 200. The maximum deviation
metric follows the deviation tolerance limit closely for this
range level, whereas it has nearly 30% slack for R � 250
instances at 50% deviation tolerance.
The US-E420 instances require significantly more
facilities than the other networks that we tested. The
highest number of stations located is 97 for R � 400,
whereas the lowest is 25 for R � 600 instances. We also
observe that R � 400 instances require almost twice as
many stations as their R � 600 counterparts, although
the number of OD trips is only 9% more for the latter.
As it is the case with the other two networks, lower
values of α might require up to 20% additional capacity
at the established stations, whereas it does not change
the station count in any of the experiments. The station
locations are observed to be insensitive to the changes
in α in the majority of the instances.
15
Figure 8 in Online Appendix C shows the sensitivity
of the objective values based on the tested waiting time
threshold and deviation tolerance values for US-E420
solutions. Similar to the CA339 results, the difference
between the objective values of the solutions with 0%
and 10% deviation tolerances is the highest. On the
other hand, the objective function value differences
between 25% and 50% deviation tolerance instances are
relatively smaller, particularly for higher R. For both
R � 400 and R � 600 instances, the α � 1-minute solu­
tions are approximately 5.5%, 10.2%, and 20% more
costly than α � 5, 10, 30-minute solutions, respectively.
The average and maximum deviation comparison of
US-E420 solutions for α � 10 instances are illustrated in
Figure 9 in Online Appendix C. For both R � 400 and
R � 600 solutions, the value of maximum deviation
closely follows the deviation threshold. On the other
hand, the average deviation is only as high as 13% and
as low as 1.42% over all solutions.
When deviation tolerance is 0% (i.e., every EV user can
complete their shortest path without running out of
energy) and R � 600 for the instances of the US-E420 net­
work, 55 charging stations and over 600 charging stalls
are required, which necessitates an approximately $60
million investment. When the deviation tolerance is set to
25%, only 31 charging stations and a similar number of
stalls are required, and the required investment decreases
to approximately $50 million. In order to achieve about
$10 million in savings in the infrastructure cost, the deci­
sion maker can allow for a maximum of 25% deviation.
In this case, note that the average deviation per EV user
will only be 5.92% when α � 10 minutes and β � 90%
(Table 3 in the online appendix).
Figure 10 in Online Appendix C presents the spatial
comparison of charging station locations and sizes on
two representative solutions with two different devia­
tion tolerances for R � 600 km, α � 10 minutes, and β �
90% on US-E420. The sizes of the charging stations are
depicted using different sizes for the green triangles.
When the deviation tolerance is 10% (Figure 10(a) in the
online appendix), there are 39 charging stations. The
largest stations are located near the center of the net­
work, whereas numerous smaller stations are scattered
throughout the region. On the other hand, when the
deviation tolerance is 50% (Figure 10(b) in the online
appendix), there are only 25 charging stations, and the
majority are of larger sizes that are located closer to the
center; the number of smaller charging stations is signif­
icantly fewer.
For a more extensive analysis on how the deviation
tolerance and vehicle range affect the size of the facilities,
we provide the histograms in Figure 11 in Online
Appendix C that compare the solutions with fixed α and
β values at 10 minutes and 90%, respectively. For both
R � 600 and R � 400, instances with smaller-deviation
tolerances require significantly more stations of smaller

16
sizes. Higher deviation tolerances allow us to aggregate
the charging demand at fewer stations that are of larger
size. This aggregation not only locates fewer facilities,
but it also requires less capacity in total (see Table 3 in
the online appendix). This indicates the importance of
determining the right deviation tolerance for the right
applications.
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
6.4. The Effect of the Variable Cost on Algorithm
Performance
Our decomposition-based algorithm is geared toward
a problem setting where the fixed cost of locating a
charging station is greater than the variable cost of
capacity installation. This is motivated by real-life cost
configurations that follow this relationship (Nicholas
2019). In order to evaluate computational performance
of the algorithm on instances with varying levels of
ratios between fixed and variable cost components, we
perform additional computational experiments on the
N25 network. We use the same cooperative response
instance configurations introduced in Section 6.1 (i.e.,
R � {15, 20, 25}, α � {1, 5, 10, 30}, β � {80, 90, 95}, and τ �
{1:0, 1:1, 1:25, 1:5}), keep the fixed cost of building a
charging station as $150,000, and test five variable
charging stall cost values: $75,000, $100,000, $150,000,
$200,000, and $300,000. This generates 720 test in­
stances. We first solve each of these instances by Gurobi
with a one-hour time limit and then use DA with a
20-second time limit. Table 4 in Online Appendix B
compares the solution quality in terms of mean gap
(percentage) and reports the average solution time of
Gurobi. The mean gap (percentage) columns compare
the objective function value of the algorithm with that
of Gurobi. For each R and deviation tolerance level,
reported means are taken over 12 different configura­
tions of α and β.
It is observed that the required computational effort
to find an optimal solution by the solver increases as
the ratio between fixed and variable cost decreases.
Similarly, R � 15 solutions tend to have a significantly
longer computation time compared with the higher R
levels. Among the 240 instances with R � 15, Gurobi
failed to find the optimal solution for 16 of them within
the one-hour time limit. Among them, DA found better
solutions than Gurobi within the 20-second time limit.
In 674 of the 720 instances, DA finds the optimal or bet­
ter solution than Gurobi within 20 seconds. It is able to
find the optimal solutions reported by Gurobi for all
R � 20 and R � 25 instances. However, as expected, the
mean optimality gap is observed to increase as the
ratio between fixed and variable costs decreases. The
largest mean optimality gap is only 0.53% for R � 15,
τ � 1:25, and stall cost equals $300,000 instances. In
turn, the highest individual optimality gap is observed
to be 1.28%.
Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
6.5. Analyzing the Types of OD Pairs Affected
from Deviation Allowance
In order to provide more insights into the trade-off
between the satisfaction of EV users and infrastructure
deployment, we define the shortest path for all OD pairs
as a benchmark and then investigate the types of OD
pairs that are mostly affected by deviation allowance.
We use the same cooperative response instance config­
urations used in the previous subsection for N25 and set
the fixed cost of building a charging station as $150,000
and charging stall cost as $75,000. The aim is to compare
the impact of 25% and 50% deviation tolerance policies
and evaluate any correlation to shortest path distance
and Poisson OD flow rate (λk). The analysis of the for­
mer is presented on Figure 12 in Online Appendix C,
and the latter is presented on Figure 13 in Online
Appendix C. These analyses are performed on the set of
aggregate instances of N25.
The results on Figure 12 in the online appendix dem­
onstrate that the OD pairs that have shorter shortest
paths are impacted more significantly from higher devi­
ation tolerances. In other words, the longer the shortest
paths are, the less likely an OD pair is to deviate in an
optimal solution based on the observed data. For exam­
ple, when τ � 25%, the group of longest shortest path
OD pairs (shortest path lengths greater than 33 units)
does not deviate more than 10%, and they do not devi­
ate at all when τ � 50%. On the other hand, when
τ � 50%, the group of shortest OD pairs (shortest path
length less than 17 units) is subject to much greater
levels of deviation within 50%.
Based on the analysis depicted in Figure 13 in the
online appendix, a straightforward correlation between
OD flow rate and deviation tolerance policy cannot be
observed (i.e., the OD pairs that deviate more do not
necessarily have high or low OD flow rate). This is not
an unexpected outcome as the OD flow rate does
impact the lower-level decision in terms of the chosen
path, but it drives upper-level decisions to determine
the level of capacity required to ensure that the service-
level requirements are met.
6.6. Analyzing EV User Response Behavior
In this section, we compare optimal cooperative re­
sponse solutions with uncooperative response solutions
on select N25 instances. Recall that the uncooperative
response solutions are obtained by the adaptation of
the algorithm with the modification of PS as described in
Section 5.2.2. The optimality of the uncooperative re­
sponse solutions can only be claimed by exhausting all
possible alternative solutions of the problem, which is
not tractable in practice. Hence, we use the same stop­
ping condition for the uncooperative response version
of the algorithm as with the cooperative response.
Because the follower’s problem ensures that each OD
pair follows their minimum-length path, the average
 Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
and maximum deviation metrics stay the same when
exactly the same sets of stations are open in the coopera­
tive and uncooperative response solutions. In particu­
lar, for the solutions of R � 20 and R � 15 instances, the
set of charging station locations stays the same, whereas
the uncooperative response solutions result in different
required capacity levels. On the other hand, the sets of
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
open stations are different for all R � 25 instances when
the deviation tolerance is greater than 0%. This indicates
that the locations of the charging stations may change
based on the position determined for the follower. The
complete results of this comparison are provided in
Table 5 in the online appendix.
Interestingly, smaller differences in the objective func­
tion values are observed when the deviation tolerance is
higher. For that matter, the cooperative and uncoopera­
tive response solutions turn out to be equivalent when
the deviation tolerance is 50% for both R � 25 and R � 20.
This is an insightful result for the decision makers
because it shows that having a higher deviation toler­
ance mitigates the difference that might result from an
uncooperative response of the EV users. In other words,
higher deviation tolerances for the cooperative response
case play an important role to provide more robust solu­
tions for the cases in which the follower may not behave
as desired. On the other hand, lower-deviation tolerance
levels tend to yield higher differences in the objective
function values between the cooperative and uncoopera­
tive response solutions. Intuitively, one might think that
our bilevel model does not make a difference when
the deviation tolerance is set to 0% as the route choices
are no longer relevant. However, our computational
results reveal that the objective function difference
between the cooperative and uncooperative response
solutions is usually higher for the 0% deviation toler­
ance instances. This indicates that using lower devia­
tion tolerances, especially 0%, might generate solutions
that may easily be overridden by uncooperative re­
sponses of the EV users in terms of choosing at which
station(s) to recharge.
7. Conclusion
In this paper, we develop a bilevel optimization model
to determine the strategic locations and capacities of
DC fast-charging stations under stochastic demand to
enable long-distance travel with EVs. The model allows
the possibility of detours from shortest paths within a
deviation tolerance without endogenously determining
the deviation paths. Through the bilevel framework, the
model takes into account the route choice response of
EV users, as well as their cooperative or uncooperative
charging station selection responses, when alternative
solutions exist. Besides providing a general formulation
of the problem, we model the M/M/c queuing system
characterization of the DC fast-charging stations for
17
the case of Poisson arrivals and exponential service
times. We derive the MILP equivalent of the chance-
constrained stochastic bilevel optimization formulation
by modeling the underlying charging station network as
an open multiserver Jackson network.
We develop a decomposition-based algorithm to solve
the proposed bilevel optimization model under both
cooperative and uncooperative follower responses. The
algorithm utilizes a new exact logic-based Benders meth­
odology for the uncapacitated version of the problem as
a subroutine. The solution quality and the computational
performance of the algorithm are verified on smaller-
sized network instances, where it finds optimal solutions
for all of the 144 test instances. The computational per­
formance comparison with a commercial solver reveals
that our algorithm is much more efficient on instances
with higher deviation tolerances and lower range values.
Allowing for higher deviation tolerances (up to 50%) can
especially be relevant for autonomous long-distance
travel with EVs.
The results demonstrate that the number of located
charging stations is insensitive to the changes in the
service-level requirements. Typically, more stringent
service-level thresholds require more capacity to be
installed at the charging stations. The solutions with
higher deviation tolerance are observed to require less
infrastructure cost and are not prone to generate alter­
native stopping options for the EV users. This leads to
robust solutions that are equivalent for the cooperative
and uncooperative responses of the EV users and pro­
tect the decision maker against the worst possible out­
come. On the other hand, the solutions allowing for
lower deviation tolerance tend to increase the number
of required stops and result in higher differences in the
optimal objective function values obtained under the
cooperative and uncooperative responses.
In this work, the route choice behavior of EV users is
driven by shortest path given the locations of charging sta­
tions. This allows us to solve our problem on large-scale
instances by developing solution methods that exploit
properties of the shortest path problem. Although the
trip distance is the only term that is explicitly considered
in the objective function of the follower’s problem, to
ensure that the charging network offers convenient ser­
vice, service and waiting times are considered by the
probabilistic service-level constraint. Assuming that traf­
fic congestion is negligible during long-distance trips,
what leads to faster travel are the trip distance, service
time, and waiting time. Incorporating features, such as
traffic congestion or dynamic charge pricing and toll
pricing, rather than relying on the shortest path route
choice assumption would require new solution algo­
rithms. Under our modeling framework, the service-
level constraint ensures that the levels of congestion
experienced at all located facilities are comparable. Since
the actual congestion, hence trip completion time, will be

18
driven by real-time dependencies (e.g., traffic condition),
this is a feature that may be observed by performing a
discrete-event simulation. Another interesting extension
of our work is to evaluate various path choice settings,
such as real-time congestion, toll pricing, and charge
pricing, through simulation.
Future work may focus on different types of queuing
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
system characterizations for charging stations. For exam­
ple, an extension with M/D/c systems may be a viable
option for centralized systems with autonomous vehicles
or for commercial applications that can enforce deter­
ministic service times to the users of a charging station.
Another research dimension may consider a multiperiod
framework to determine infrastructure expansion plans
considering the growth in charging demand because of
the increasing adoption rate of EVs.
Acknowledgments
The authors express their appreciation to Research Assistant
Jonathan Wong during the data set generation phase of this
work. The authors also thank the anonymous referees who
have contributed to the further development of this manu­
script by their suggestions and comments.
References
Anjos MF, Gendron B, Joyce-Moniz M (2020) Increasing electric vehicle
adoption through the optimal deployment of fast-charging stations
for local and long-distance travel. Eur. J. Oper. Res. 285(1):263–278.
Arslan O, Karaşan OE (2016) A Benders decomposition approach for
the charging station location problem with plug-in hybrid electric
vehicles. Transportation Res. Part B Methodological 93:670–695.
Arslan O, Yıldız B, Karaşan OE (2014) Impacts of battery character­
istics, driver preferences and road network features on travel
costs of a plug-in hybrid electric vehicle for long-distance trips.
Energy Policy 74:168–178.
Arslan O, Karaşan OE, Mahjoub AR, Yaman H (2019) A branch-
and-cut algorithm for the alternative fuel refueling station loca­
tion problem with routing. Transportation Sci. 53(4):1107–1125.
Bard JF (1991) Some properties of the bilevel programming problem. J.
Optim. Theory Appl. 68(2):371–378.
Bard JF (2013) Practical Bilevel Optimization: Algorithms and Applica­
tions, vol. 30 (Springer Science & Business Media, New York).
Bose SK (2013) An Introduction to Queueing Systems (Springer Science
& Business Media, New York).
Burke PJ (1956) The output of a queuing system. Oper. Res. 4(6):699–704.
Capar I, Kuby M (2012) An efficient formulation of the flow refueling
location model for alternative-fuel stations. IIE Trans. 44(8):622–636.
Capar I, Kuby M, Leon VJ, Tsai YJ (2013) An arc cover–path-cover for­
mulation and strategic analysis of alternative-fuel station locations.
Eur. J. Oper. Res. 227(1):142–151.
Chromy E, Misuth T, Weber A (2012) Application of erlang formulae in
next generation networks. Internat. J. Comput. Network Inform. Secu­
rity 4(1):59–66.
Electrek (2019) Tesla starts limiting charge to 80% at busy superchargers
to reduce wait times. Accessed June 10, 2021, https://electrek.co/
2019/05/24/tesla-limiting-supercharger-busy/.
European Commission (2019) The European Green Deal: Sustain­
able mobility. Accessed September 24, 2021, https://ec.europa.
eu/commission/presscorner/api/files/attachment/860070/
Sustainable_mobility_en.pdf.pdf.
Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
Frade I, Ribeiro A, Gonçalves G, Antunes AP (2011) Optimal loca­
tion of charging stations for electric vehicles in a neighborhood
in Lisbon, Portugal. Transportation Res. Rec. 2252(1):91–98.
Goodchild MF, Noronha VT (1987) Location-allocation and impul­
sive shopping: The case of gasoline retailing. Ghosh A, Ruston
G, eds. Spatial Analysis and Location-Allocation Models (van Nos­
trand Reinhold, New York), 121–136.
Göpfert P, Bock S (2019) A branch & cut approach to recharging and
refueling infrastructure planning. Eur. J. Oper. Res. 279(3):808–823.
Guo F, Yang J, Lu J (2018) The battery charging station location
problem: Impact of users’ range anxiety and distance conve­
nience. Transportation Res. Part E Logist. Trans. Rev. 114:1–18.
Hagberg A, Swart PS, Chult D (2008) Exploring network structure,
dynamics, and function using NetworkX. Technical report, Los
Alamos National Laboratory, Los Alamos, NM.
He J, Yang H, Tang TQ, Huang HJ (2018) An optimal charging station
location model with the consideration of electric vehicle’s driving
range. Transportation Res. Part C Emerging Tech. 86:641–654.
Hong S, Kuby M (2016) A threshold covering flow-based location
model to build a critical mass of alternative-fuel stations. J.
Transportation Geography 56:128–137.
Hooker JN, Ottosson G (2003) Logic-based benders decomposition.
Math. Programming 96(1):33–60.
Hosseini M, MirHassani SA, Hooshmand F (2017) Deviation-flow refuel­
ing location problem with capacitated facilities: Model and algo­
rithm. Transportation Res. Part D Transportation Environ. 54:269–281.
Huang Y, Kockelman KM (2020) Electric vehicle charging station loca­
tions: Elastic demand, station congestion, and network equilibrium.
Transportation Res. Part D Transportation Environ. 78:102179.
IEA (2020) Improving the sustainability of passenger and freight
transport. Tracking Transport 2020. Accessed September 1, 2021,
https://www.iea.org/topics/transport.
IEA (2021) Global EV outlook 2021: Accelerating ambitions despite the
pandemic. Report, IEA. Accessed September 21, 2021, https://
www.iea.org/reports/global-ev-outlook-2021.
Jiang Y, Zhang Y, Zhang C, Fan J (2012) Capacitated deviation-flow fuel­
ing location model for sitting battery charging stations. CICTP 2012
Multimodal Transportation Systems—Convenient Safe Cost-Effective Effi­
cient, 2771–2778. https://doi.org/10.1061/9780784412442.282.
Jing W, An K, Ramezani M, Kim I (2017) Location design of electric vehi­
cle charging facilities: A path-distance constrained stochastic user
equilibrium approach. J. Adv. Transportation 2017:4252946.
Kim JG, Kuby M (2012) The deviation-flow refueling location model
for optimizing a network of refueling stations. Internat. J. Hydro­
gen Energy 37(6):5406–5420.
Kim JG, Kuby M (2013) A network transformation heuristic approach
for the deviation flow refueling location model. Comput. Oper. Res.
40(4):1122–1131.
Kınay ÖB, Gzara F, Alumur SA (2021) Full cover charging station
location problem with routing. Transportation Res. Part B Meth­
odological 144:1–22.
Ko J, Gim THT, Guensler R (2017) Locating refuelling stations for
alternative fuel vehicles: A review on models and applications.
Transportation Rev. 37(5):551–570.
Kuby M, Lim S (2005) The flow-refueling location problem for
alternative-fuel vehicles. Socio-Economic Planning Sci. 39(2):125–145.
Kuby M, Lim S (2007) Location of alternative-fuel stations using the
flow-refueling location model and dispersion of candidate sites
on arcs. Network Spatial Econom. 7(2):129–152.
Kuby M, Lines L, Schultz R, Xie Z, Kim JG, Lim S (2009) Optimiza­
tion of hydrogen stations in Florida using the flow-refueling
location model. Internat. J. Hydrogen Energy 34(15):6045–6064.
Labbé M, Marcotte P (2021) Bilevel network design. Crainic TG, Gen­
dreau M, Gendron B, eds. Network Design with Applications to Trans­
portation and Logistics (Springer, Cham, Switzerland), 255–281.
 Kınay, Gzara, and Alumur: Charging Station Location and Sizing
Transportation Science, Articles in Advance, pp. 1–19, © 2023 INFORMS
Li S, Huang Y (2014) Heuristic approaches for the flow-based set
covering problem with deviation paths. Transportation Res. Part
E Logist. Trans. Rev. 72:144–158.
Makhlouf W, Kchaou-Boujelben M, Gicquel C (2019) A bi-level program­
ming approach to locate capacitated electric vehicle charging sta­
tions. 2019 6th Internat. Conf. Control Decision Inform. Tech. (CoDIT)
(IEEE, Piscataway, NJ), 133–138.
MirHassani S, Ebrazi R (2013) A flexible reformulation of the refuel­
ing station location problem. Transportation Sci. 47(4):617–628.
Downloaded from informs.org by [131.175.147.3] on 27 June 2023, at 07:22 . For personal use only, all rights reserved.
Nicholas M (2019) Estimating electric vehicle charging infrastructure
costs across major us metropolitan areas. Accessed September
20, 2021, https://theicct.org/sites/default/files/publications/
ICCT_EV_Charging_Cost_20190813.pdf.
Pelletier S, Jabali O, Laporte G, Veneroni M (2017) Battery degrada­
tion and behaviour for electric vehicles: Review and numerical
analyses of several models. Transportation Res. Part B Methodo­
logical 103:158–187.
Shen ZJM, Feng B, Mao C, Ran L (2019) Optimization models for
electric vehicle service operations: A literature review. Transpor­
tation Res. Part B Methodological 128:462–477.
Simchi-Levi D, Berman O (1988) A heuristic algorithm for the traveling
salesman location problem on networks. Oper. Res. 36(3):478–484.
Sinha A, Malo P, Deb K (2017) A review on bilevel optimization:
From classical to evolutionary approaches and applications.
IEEE Trans. Evolution Comput. 22(2):276–295.
The Climate Group (2021a) EV100 Initiative. Accessed September
23, 2021, https://www.theclimategroup.org/ev100.
The Climate Group (2021b) EV100 Progress and Insights Report 2021.
Accessed September 23, 2021, https://www.theclimategroup.org/
media/7941/download.
The White House (2021) Executive order on strengthening American
leadership in clean cars and trucks. Accessed September 24, 2021,
https://www.whitehouse.gov/briefing-room/presidential-
actions/2021/08/05/executive-order-on-strengthening-american-
leadership-in-clean-cars-and-trucks/.
Tran CQ, Ngoduy D, Keyvan-Ekbatani M, Watling D (2021) A user
equilibrium-based fast-charging location model considering
19
heterogeneous vehicles in urban networks. Transportmetrica A
Transportation Sci. 17(4):439–461.
Tu W, Li Q, Fang Z, Shaw S-l, Zhou B, Chang X (2016) Optimizing the
locations of electric taxi charging stations: A spatial–temporal
demand coverage approach. Transportation Res. Part C Emerging
Tech. 65:172–189.
Upchurch C, Kuby M (2010) Comparing the p-median and flow-
refueling models for locating alternative-fuel stations. J. Trans­
portation Geography 18(6):750–758.
Upchurch C, Kuby M, Lim S (2009) A model for location of capaci­
tated alternative-fuel stations. Geographical Anal. 41(1):85–106.
Wang YW, Lin CC (2009) Locating road-vehicle refueling stations.
Transportation Res. Part E Logist. Trans. Rev. 45(5):821–829.
Wang YW, Wang CR (2010) Locating passenger vehicle refueling sta­
tions. Transportation Res. Part E Logist. Trans. Rev. 46(5):791–801.
Xie F, Lin Z (2021) Integrated us nationwide corridor charging infra­
structure planning for mass electrification of inter-city trips.
Appl. Energy 298:117142.
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.
Transportation Res. Part E Logist. Trans. Rev. 109:261–276.
Yıldız B, Arslan O, Karaşan OE (2016) A branch and price approach
for routing and refueling station location model. Eur. J. Oper.
Res. 248(3):815–826.
Zeng B (2020) A practical scheme to compute the pessimistic bilevel
optimization problem. INFORMS J. Comput. 32(4):1128–1142.
Zeng D, Dong Y, Cao H, Li Y, Wang J, Li Z, Hauschild MZ (2021)
Are the electric vehicles more sustainable than the conventional
ones? Influences of the assumptions and modeling approaches
in the case of typical cars in China. Resources Conservation Recy­
cling 167:105210.
Zhang A, Kang JE, Kwon C (2017) Incorporating demand dynamics in
multi-period capacitated fast-charging location planning for electric
vehicles. Transportation Res. Part B Methodological 103:5–29.
Zheng H, He X, Li Y, Peeta S (2017) Traffic equilibrium and charging
facility locations for electric vehicles. Networks Spatial Econom.
17(2):435–457.