23858 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO.

12, DECEMBER 2022

A Novel Data-Driven Approach for Solving the

Electric Vehicle Charging Station
Location-Routing Problem
Ying-Chao Hung and George Michailidis , Member, IEEE

Abstract— Due to increasing rates of adoption of electric adoption is driven by many factors including, on the technical
vehicles (EVs), there is a strong need to deploy the necessary side, advances in battery and power electronics technologies
charging station infrastructure, together with routing strategies that have increased substantially the EVs driving range, on the
to manage traffic flow and congestion. This study addresses the
location-routing problem (LRP) for a general EV charging sys- consumer side, environmental concerns, on the supply side,
tem with stochastic charging requests regarding their locations, new model offerings by manufacturers, and on the policy
arrival times and charging times. The objective is to develop an side, tighter emission regulations and mandates for energy
efficient routing strategy of EVs to charging stations, as well independence aided by financial incentive programs [4], [5],
as to determine the optimal charging station locations so as [6], [7]. However, a key enabler for faster EV adoption is
to minimize the demand’s mean response time. Under some
regularity assumptions on the mean waiting time at each charging the large scale deployment of an efficient and well managed
station (e.g. system operates in a light or heavy traffic regime), charging station infrastructure. To that end, many cities deloy
we show that the optimization problem can be formulated as chargers in public spaces to allow easy access by EVs.
a partition-based clustering problem with size constraints. This The charging time of an EV depends on the characteris-
relaxation of the problem formulation enables us to develop tics of the vehicle, as well as the technology and types of
a novel data-driven approach for solving the charging station
LRP, without requiring detailed stochastic models for the EV’s the charging system. Chargers can be installed in houses,
charging requests, as well as the queueing behavior of the workplaces, private facilities and public areas. Installation
charging stations. An algorithm along with two size adjustment of chargers in public spaces is considered to be crucial for
strategies are developed to solve the obtained clustering problem various reasons. One reason is that charging at home is time
and illustrated on urban areas of Seattle with various types consuming (a typical full charge can exceed 6 hours) and may
of distance, vehicle speeds, distributions for charging request
locations, and inter-arrival time densities. not be a viable option for many EV owners in dense urban
areas. Another reason is that it alleviates drivers’ concern about
Index Terms— Location-routing problem, mean response time, exhausting their battery charge while being away from home
queueing system, heavy traffic approximation, clustering with
size constraints, shortest route distance. or the workplace. Further, the location of charging stations
impacts the EV flow and consequently traffic conditions on
I. I NTRODUCTION the road network.

W ITH the aid of well-developed information and com-

munication technologies (ICT) that enable transmission
and analysis of real-time transportation information together
There has been a growing literature addressing various
issues related to problems of determining the location of EV
charging stations, including integration with existing distrib-
with historical data, Intelligent Transportation Systems (ITS) ution system infrastructure [8], [9], [10], examining several
are becoming capable of providing solutions for improving practical aspects [11], [12], [13], [14] and proposing different
safety, efficiency, and service quality through an intelligent algorithms to solving the problem [15], [16], [17], [18],
network and a cloud computing support platform [1], [2], [3]. [19]. In a related line of work, a number of papers consider
An important component for ITS that has received increasing the impact of charging station locations on the distribution
attention in the literature is accommodating electric vehicles network and the power grid in general [20], [21], [22], [23].
(EV), whose global sales exceeded 3 million units in 2020 and Another thrust in the literature considers the problems of
are forecasted by the International Energy Agency to surpass how to route EVs to charging stations [24], [25], [26] and
125 million units by 2030. The accelerated pace of EV where to locate the latter jointly based on different problem
formulations, among which a large portion of the works
Manuscript received 29 September 2021; revised 30 March 2022;
accepted 27 July 2022. Date of publication 26 August 2022; date of current formulate the problem as a mixed integer programming one
version 5 December 2022. This work was supported by the Research Grant and find the solution with heuristic algorithms [27], [28], [29],
MOST 110-2118-M-004-004-MY2. The Associate Editor for this article was [30], [31], [32], [33].
B. Singh. (Corresponding author: Ying-Chao Hung.)
Ying-Chao Hung is with the Institute of Industrial Engineering, National Note that in order to provide real time solutions to the
Taiwan University, Taipei 10617, Taiwan (e-mail: [email protected]). charging station location-routing problem, the design of an ITS
George Michailidis is with the Department of Statistics and the Department should be able to respond to changes of EV drivers’ behavior
of Computer Science, Informatics Institute, University of Florida, Gainesville,
FL 32611 USA (e-mail: [email protected]). or traffic patterns. This accentuates the need for employing
Digital Object Identifier 10.1109/TITS.2022.3196835 data-driven approaches that require fewer rigid stochastic
1558-0016 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.
HUNG AND MICHAILIDIS: NOVEL DATA-DRIVEN APPROACH 23859

assumptions on charge requests and rely mainly on historical different types of distances, including shortest and fastest
and geographic charging event data collected from the EVs. route ones.
Recent work taking such an approach focuses on optimizing The remainder of the paper is organized as follows.
various performance metrics, including demand coverage [34], Sec. II introduces a general EV charging system and assump-
profits of new charging stations [35], walking distance [36] tions on the stochastic mechanism for the charging request
and energy requirements [37]. However, there is fairly limited process, locations, EVs speed, and operational characteristics
work addressing the EV routing problem. An optimal routing of the charging stations. Since the charging station is modeled
solution for travel time and energy consumption is obtained by as a queueing system, the fundamental stability issue is
using an adaptive epsilon constraint method [38], while [39] also discussed. Sec. III introduces a location-based routing
leverages a heavy traffic approximation regime to minimize strateg1y, together with the location problem for placing the
the charging system’s mean response time. EV charging stations, while the goal is to minimize the
The goal of this study is to address the joint EV charg- system’s overall mean response time. Under the regularity
ing station location-routing problem, leveraging information assumption on the charging requests’ mean waiting time
obtained from historical data on the location and arrival rate at each station, we show that the optimal routing strategy
of EV charging requests, the mean vehicle speed, together and charging station locations can be selected by solving
with information about the capacity of charging stations (pos- a dual partitioning-based clustering problem with size con-
sibly constrained by the local distribution network), so as straints that satisfy the system’s stability condition. A typical
to minimize the EV’s mean response time (including mean partition-based algorithm along with two novel size adjust-
travel time, waiting time and charging time). With a regularity ment strategies are developed to solve the resulting clustering
assumption on the waiting time at each charging station, problem. Sec. IV contains simulation studies that illustrate
we show that the problem can be reduced to a standard our proposed approach by considering Euclidean and shortest
clustering problem with size constraints. Further, the optimal route distances, which are motivated by the transportation
routing policy reduces to allocating all charging requests to system in the urban areas of Seattle, Washington. Concluding
the station that is at the “center” of these requests. Hence, remarks and discussion are provided in Sec. V.
a complex LRP can be efficiently solved by leveraging approx-
II. T HE EV C HARGING S YSTEM
imations of key quantities by existing results in the queueing
literature and algorithms developed for solving a variant of Suppose instant random charging requests are generated
the clustering problem. It should be highlighted that the by EV drivers at locations x 1 , . . . , x n in a geographic area
regularity assumption on the mean waiting time holds for G with the latitude and longitude coordinates (such as a
charging systems under both light traffic and heavy traffic metropolitan area, a transportation network, or a collection
regimes, with some minor technical conditions on the EV of transportation routes and buildings, and so forth) according
speed, service and/or inter-arrival times of charging requests to a general arrival process with rate λ. We denote the random
(see Sec. III-C for details). However, the heavy traffic regime is variable of these locations by X, which has a probability
of particular interest, since it corresponds to a scenario wherein density function f (x) over G. Note that in practice x 1 , . . . , x n
the charging infrastructure may exhibit constrained power/grid can be technically positioned and collected by a two-way
capacity at peak times. The main contributions of this work communication GPS system, while f (x) can be estimated
are highlighted next: by the so-called kernel method [42] based on x 1 , . . . , x n .
 −1/2estimator in R is given
2
• The proposed approach is data-driven and does not To illustrate, a 2-D  kernel density
impose any distributional assumptions on the charging times. by fˆH (x) = n|H |1/2 i=1 K H
1 n
(x − x i ) , where K (·) is

It simply requires a two-way Global Positioning Sys- a non-negative kernel function satisfying G K (x)d x = 1 (e.g.,
tem (GPS) communication technology that accurately iden- a bivariate Gaussian density) and H is a 2 × 2 symmetric and
tifies the location of instant charging requests sent by EV positive definite bandwidth matrix that controls the smoothness
drivers. of the estimator.
• The proposed approach solves a joint charging station Suppose that at present, either there are no charging stations
location-routing problem under some regularity conditions. Its or, K extant charging stations in G. The facility planner
novelty lies in reducing a complex optimization problem as wants to place or replace the extant K charging stations in
a machine learning one (i.e., clustering problem with size G in some optimal manner, whose locations are denoted by
constraints), for which existing algorithms can be leveraged S1 , S2 , . . . , S K . Once the i -th vehicle places a request for
to provide a high quality solution. charging at location x i at time ti , a service policy π would
• The proposed approach provides adaptive and centralized route it to one of the charging stations Si ∈ {S1 , . . . , SK }
control strategies for both EV routing and charging station and the EV would proceed to the station at speed v(ti , x i , Si ),
location selection. Specifically, EV drivers can sign up an inte- a function of the location and the various service stations (i.e.,
grated data exchange platform in an ITS so that traffic flows travel paths) and the time that the request was made. There
can be well directed/managed [40], [41]. It is often associated are two main questions to be addressed: (i) where to locate the
with a high level design of EV charging infrastructure for service stations, and (ii) what is an efficient vehicle routing
smart cities. policy π, in the sense that it maximizes the service system
• With the aid of platforms such as Google Maps, the throughput -i.e., the long term average number of service
proposed approach is easy to implement and can leverage requests given charging stations’ capacity constraints.

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.
23860 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 12, DECEMBER 2022

Next, we define a number of useful quantities needed to

formally define the routing policy π. The mean vehicle speed
over G is defined as the empirical long-term average

1
n
v = lim v(ti , x i , Si ), (1)
n→∞ n
i=1

which has an upper bound v ≤ v̄ over the region G due to

regulatory speed limits. Hence, the mean travel time from
a random location x to a charging station Sk is given by
Fig. 1. The graphical illustration of the input and density of charging demand
d(x, Sk )/v.1 Since EV drivers tend to have a common concern locations over G (left) and the system with K charging stations in queueing
of range anxiety, d(x, Sk ) mainly refers to the “shortest route context (right).
distance” between x and Sk , whose calculation obviously
depends on the transportation network topology. In practice, (Theorem 1, Sec. III-B).2 Note that we do not impose any
d(x, Sk ) can be directly calculated by using a satellite map additional distributional assumptions on the charging times
(e.g., Google Maps). Finally, it is assumed that once a charging (e.g. exponential distribution or independent and identically
station is assigned to an EV, the latter would honor the distributed charging times), thus contributing to the generality
assignment. of the obtained results. Finally, an EV leaves the system upon
Each charging station Sk comprises of an infinite buffer charging completion.
first-come-first-serve (FCFS) queue with μk charging units Let x 1 , x 2 , . . . , x n be a sequence of locations where charg-
that can serve μk vehicles simultaneously. Note that the ing requests originate. For a given routing policy π, define the
infinite-buffer assumption here allows us to develop rigorous functions
methodology, obtain tractable analysis results, test and design ⎧
⎪
⎨ 1 if charging request originating at x i is
practical systems. If at any point in time there are more than
μk vehicles at Sk , they would queue up and wait for service. h k (x i ) = directed to Sk , (5)
⎪
⎩
Due to uncertainty caused by the EV battery capacity and/or 0 otherwise,
driver’s charging behavior (e.g., a decision to fully or partially
where k = 1, . . . , K . Based on the definition, the long-term
charging the battery), the charging time for any EV at charging
proportion that a charging request is directed to station Sk is
unit j in station Sk is assumed to be random and denoted by
j given by
σk , j = 1, . . . , μk , k = 1, . . . , K . In addition, for technical n
reasons we make the following standard assumptions on the i=1 h k (x i )
charging times: rk = lim , (6)
n→∞ n
K
j
σk are mutually independent for all j and k; (2) where 0 ≤ rk ≤ 1 for all k = 1, . . . , K and k=1 rk = 1.
j Conceptually, a typical routing policy π splits the primary
E[σk ] = 1 for all j and k; (3) traffic flow into K sub-flows and directs an rk portion of it to
station k. Fig. 1 illustrates an EV charging system described
and above.
μ A key issue in the analysis of the service system and
V ar (σk1 ) = · · · = V ar (σk k ) for all k, (4)
the development of efficient routing policies is the notion of
stability. Any routing policy π needs to satisfy
where j = 1, . . . , μk , k = 1, . . . , K .
Assumption (3) indicates that all charging units in the rk λ < μk for all k = 1, . . . , K ; (7)
system charge at the same rate (that without loss of generality
can be scaled to one). Therefore, the capacity of each charging that is, the input load for each sub-flow must be less than
station Sk is simply characterized by the number of associated the service capacity available at the charging station. This
charging units μk , while the total capacity of the entire system definition corresponds to the so-called stability condition in
is μ = μ1 + · · · + μ K . Assumption (4) indicates that the the queueing literature (e.g. [25]) and automatically imposes a
μk units at each station Sk also have the same variance. natural upper bound on the total input rate to the system, which
Thus, the posited service system comprises of homogeneous is given by λ < μ1 + · · · + μ K . The quantity (μ1 + · · · + μ K )
stations/units with respect to the first two moments of charging is referred to as the maximum throughput of the system.
times. It should be mentioned that this assumption is necessary 2 Another way to technically accommodate Assumption (4) is to consider
for obtaining the theoretical results and undertaking numerical multiple types of charging requests (or chargers) at each station (e.g., general
validation in later analysis (see Sec. III). In practice, it may size EV, large goods/passenger EV, etc.). It is then reasonable to assume a
be properly relaxed if some regularity condition is retained mixture charging time distribution over different types of requests. However,
such assumption will divide the charging flow into multiple classes, and
the correspoding stability constraints need to be considered for each class
1 Note that this formulation simplifies the exposition, even though in practice and integrated in the primary optimization problem. Though this requires a
the mean speed v can also be time varying and hence a function of λ; e.g., v more complicated problem formulation and rigorous analysis, it constitutes
is small when the system is heavily loaded. an interesting direction of future work.

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.
HUNG AND MICHAILIDIS: NOVEL DATA-DRIVEN APPROACH 23861

Based on (8), the long-term proportion that a charging

request is directed to station Sk can be equivalently written as

rk = f (x)d x, k = 1, . . . , K . (9)
x∈G k

The latter relationship places a natural constraint on the “size”

of each partitioned subregion G k . We next introduce the
optimization problem of interest.

Fig. 2. A graphical presentation of the DLB routing policy, where the system B. The Charging Station Location Problem
has three charging stations located at S1 , S2 and S3 , associated with a partition
of three subregions G 1 , G 2 and G 3 , respectively. The previous discussion shows that the routing strategy and
the locations of charging stations impact the mean response
Given the above detailed system description, the two key time of all service requests over the region G. The response
issues addressed next are where to locate the charging stations time for a charging request at location x can be expressed as
and how to design the vehicle routing policy.

K
T (x) = h k (x) · (tk + wk + σk ) , (10)
III. L OCATING THE C HARGING S TATIONS AND D ESIGNING k=1
ROUTING P OLICIES
where the first component tk represents the travel time from
Note that the first problem of interest corresponds to the location x to charging station Sk , the second component wk
famous facility location problem [43], [44], [45], where the represents the waiting time for charging at Sk , and the last
objective is to minimize the system’s mean response time. component σk represents the charging time at Sk , respectively.
The latter however depends on the routing policy π selected, Since T (x) is a random variable associated with S1 , . . . , S K
which needs to satisfy the stability condition (7). Nevertheless, and G 1 , . . . , G K , taking its expectation yields
there is an asymmetry in these objectives; whereas, a routing 
policy can be easily redesigned, relocating service stations is 
K
d(x, Sk )
E S,P [T (x)] = h k (x) · + E S,P [wk ] + 1 ,
a time consuming and costly process. v
k=1
To that end, we decouple the two problems by first determin-
(11)
ing the scheme of the routing policy and subsequently solving
the charging station location problem. where S = {S1 , . . . , SK } and P = {G 1 , . . . , G K } represent the
collection of K charging station locations and a partition of
A. A Deterministic Location-Based Routing Policy G, respectively. Taking the expected value again with respect
to the charging demand location, the mean response time is
Consider a class of random multi-allocation routing policies given by
that routes EV charging requests originating at location x ∈ G
to a charging station Sk based on a pre-specified probability E S,P [T (X)]
0 ≤ rk (x) ≤ 1, k = 1, . . . , K . Prior work has established that   K 
d(x, Sk )
such policies can be tuned to minimize the system’s overall = h k (x) + E S,P [wk ] + 1 f (x)d x
mean service time, while maintaining its stability when the G k=1 v
charging stations locations are fixed a priori [25]. Since the K   
d(x, Sk )
locations are not yet determined, the routing policy needs to = + E S,P [wk ] + 1 f (x)d x (12)
be designed in a more nuanced manner. Gk v
k=1
Note that the stability condition (7) directly implies that a
Recall that our goal is to find a solution (S, P) that minimizes
simple version of the random multi-allocation routing policy
E S,P [T (X)], which can be equivalently formulated as
is to choose rk (x) = μk /(μ1 + · · · + μ K ) for all x ∈ G
and k = 1, . . . , K . However, to address the charging station K 
 
d(x, Sk )
location problem, it is reasonable to partition the demand minimize f (x)d x + rk E S,P [wk ]
location region G into K disjoint subregions G 1 , . . . , G K and
S,P Gk v
k=1
consider the following routing strategy: subject to rk λ < μk and 0 ≤ rk ≤ 1 for k = 1, . . . , K ,
 K
h k (x) = 1 with probability one if x ∈ G k . (8) rk = 1. (13)
k=1
In words, such a routing strategy directs all charging requests
in a subregion G k in a deterministic manner to the associated The next result establishes that under certain regularity con-
charging station Sk . Therefore, we call it the deterministic ditions, the solution of (13) can be found by considering a
location-based (DLB) routing policy. Fig. 2 provides a graph- reduced form of the primary objective function.
ical presentation of the DLB routing policy for a system with Theorem 1: Let S ∗ = {S1∗ , . . . , S K∗ } and P ∗ =
three charging stations. {G ∗1 , . . . , G ∗K } be the solution of the following

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.
23862 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 12, DECEMBER 2022

optimization problem: influenced by the primary charging demand arrival process and
K  density, the vehicle speed v, the location of charging stations
 d(x, Sk )
minimize f (x)d x (i.e. S) and the routing policy (i.e. partition P). However,
S,P v in practice this term can be estimated based on data for the
k=1 G k
subject to constraints in (13). (14) inter-arrival times of charging requests at each station and then
by simulating the system. In Sec. IV, we provide numerical
For any given v ≤ v̄ and (S, P) satisfying the constraints evidence based on empirical simulation studies that support
in (13), assume the claim that Cτ2k (or E[wk ]) in (17) does not vary much for
  any given (S, P), as ρk → 1.
rk E S ∗ ,P ∗ [wk ] − E S,P [wk ]
  Another condition for which (15) holds is when the charging
d(x, Sk ) d(x, Sk∗ )
≤ f (x)d x − f (x)d x (15) system operates under a “light traffic regime” - say, rk λ  μk
Gk v G ∗k v for all k. With moment constraints placed on τk and σk , it is
for all k = 1, . . . , K . Then, (S ∗ , P ∗ ) is the solution to the shown that the mean waiting time at each charging station is
optimization problem (13). bounded above [47]. This also implies that the mean waiting
time will be dominated by the mean travel time when the EV
Proof. By (15), it is clear that speed is not high (say, v ≤ v̄ for some v̄). Thus, Theorem 1 is
 also applicable when the system is lightly loaded.
d(x, Sk∗ )
f (x)d x + rk E S ∗,P ∗ [wk ]
G ∗k v
 D. A Dual Problem Based on Clustering
d(x, Sk )
≤ f (x)d x + rk E S,P [wk ] (16) Next, we further elaborate on how to solve for (S ∗ , P ∗ )
Gk v
in (14). Suppose there are n locations x 1 , x 2 , . . . , x n , where
for each k = 1, . . . , K . The result follows in a straightforward charging requests originate, the latter being independent and
manner by taking the summation with respect to k. identically distributed with density function f (x). Given any
Theorem 1 states that when the optimal solution (S ∗ , P ∗ ) is (S, P), the (weak) law of large number yields
replaced by any other one denoted by (S, P), if the increase K  K 

in the mean travel time to each station k is greater than k=1 x i ∈G k d(x i , Sk ) P
−
→ d(x, Sk ) f (x)d x (18)
rk × 100% decrease in the mean waiting time at that station, n Gk
k=1
then (S ∗ , P ∗ ) is also the solution that minimizes the mean
travel time. In words, if such regularity condition is satisfied, P
as n → ∞, where the notation “− →” means “convergence
the objective function in (13) can be further reduced to in probability”. Thus, if the number of observed locations
the mean travel time. Even though the result is intuitively for charging requests is large and the mean EV speed v
straightforward, it nevertheless enables us to find the optimal is given, under the heavy traffic assumption, solving the
routing strategy, as well as the charging station locations in a following optimization problem yields a good approximation
feasible and simple manner. of (S ∗ , P ∗ ):

K 
C. Validation of Assumption (15) minimize d(x i , Sk )
S,P
Let us examine the conditions for which (15) holds and thus k=1 x i ∈G k
the conclusion of Theorem 1 is applicable. We first assume 
K
that the EV system operates under a so-called “heavy traffic subject to |G k | = n k = [rk n] for all k, n k = n.
regime”, namely, λ → (μ1 + · · · + μ K ) with rk λ → μk k=1
for all k = 1, . . . , K . Such a regime provides a ready-at- (19)
hand mathematical formulation for validating assumption (15),
give that v is not particularly large. Since each charging A careful look at (19) yields that the primary optimiza-
station k now represents a G/G/μk queue (or a G I /G/μk tion problem can be reformulated as a partition based clus-
queue if the inter-arrival times for the charging requests are tering problem (also known as K -medoids clustering in
independent), the heavy traffic regime yields the following the machine learning literature [48], [49]) - where S̃ ∗ =
approximation [46]: { S̃1∗ , . . . , S̃K∗ } represents the set of optimal K -medoids3 and
P̃ ∗ = {G̃ ∗1 , . . . , G̃ ∗K } represents the associated partition of
1 ρk Cτ2 + Cσ2k K clusters with predetermined size constraints |G̃ ∗k | = n k
E[wk ] ≈ Ẽ[wk ] = · · k (17)
μk 1 − ρk 2 for k = 1, . . . , K . Note that d(x i , Sk ) represents the shortest
as the traffic intensity ρk = rk λ/μk → 1, where Cτ2k and driving distance between x i and Sk . Therefore, finding the
optimal K -medoids is a rather computationally involved task.
Cσ2k are the squared coefficients of variation for the inter-arrival
Next, we discuss how to accommodate the size constraints
time and charging time distribution at station Sk , respectively.
in (19) over K clusters. Denote by S = {S1 , . . . , S K } the set
From (3) and (4) we obtain that Cσ2k = V ar (σki ), which
of K medoids obtained from a search procedure, which is
remains a constant and characterizes the charging time dis-
tribution for each unit at station Sk . On the other hand, the 3 A medoid is the most central observation in the corresponding clus-
analytical expression for Cτ2k is far from trivial, since it is ter/partition.

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.
HUNG AND MICHAILIDIS: NOVEL DATA-DRIVEN APPROACH 23863

associated with the partition P = {G 1 , . . . , G K }. We provide boundaries between clusters so as to meet the size constraints.
two strategies for adjusting the size of each obtained cluster. We next summarize the whole procedure for finding (S ∗ , P ∗ )
Let us rearrange the order of subsets in P and denote by in the following algorithm, which assumes a fairly large
P = {G (1), . . . , G (K ) }, where subset G (k) has the k-th smallest number of charging request locations x 1 , x 2 , . . . , x n .
value in the sequence ||G 1 |−n 1 |, ||G 2 |−n 2 |, . . . , ||G K |−n K |,
k = 1, . . . , K . The first strategy, named as “smallest size-
Algorithm 1: Finding the Approximate Solution ( S̃ ∗ , P̃ ∗ )
difference first” (SSDF), is to adjust the size of each subset in
P in the order of G (1) , . . . , G (K ) . The second strategy, named 1: Randomly place an initial set of K charging station
as “largest size-difference first” (LSDF), is to adjust the size of locations S = {S1 , . . . , S K } in G (i.e., the K medoids).
each subset in P in a reverse order, i.e., G (K ), . . . , G (1) . The 2: Make a partition of K clusters based on the obtained K
detailed steps of strategies SSDF and LSDF are given next. medoids, say, P = {G 1 , . . . , G K }. This partition is simply
formed by assigning an observation x i to the cluster
whose
The SSDF Strategy
medoid Sk is closest to x i .
1: Select a small ε > 0 and set k = 1. 3: Adjust the size of each cluster by using strategy SSDF or
2: while |G (k) | = n k do LSDF, obtain the updated partition P̃.
3: If |G (k) | > n k , find x ∈ G (k) so that 4: Re-compute the K medoids based on P̃ by a search
0 < d(x, S( j )) − d(x, S(k) ) < ε and method 4 so as to minimize the objective function in (19).
j = arg min d(x, S(i) ). Denote the updated K medoids by S̃ = { S̃1 , . . . , S̃ K }.
i>k
Assign x to G ( j ) and increase the value of ε. 5: Repeat Steps 2-4 until convergence. Return the optimal
4: If |G (k) | < n k , find x ∈ / G (k) so that solution ( S̃ ∗ , P̃ ∗ ).
0 < d(x, S(k) ) − d(x, S( j )) < ε and
j = arg max d(x, S(i) ). Note that the computation complexity of executing Algo-
i>k
Assign x to G (k) and increase the value of ε. rithm 1 is of the order O(i j K n), where i is the mean number
5: end while of steps for finding a medoid S̃k based on any search method
6: Let k := k + 1. and j is the number of iterations for updating the partition P̃.
7. If k < K , reorder G (k) , . . . , G (K ) and go back to line 2; Basically, i depends on the choice of ε in strategies SSDF and
otherwise return all adjusted clusters and denote the LSDF.
updated partition by P̃ = {G̃ 1 , . . . , G̃ K }.
E. Incorporating Charging Station Operation or Installation
Costs
The LSDF Strategy Facility operation or installation costs are often considered
1: Select a small ε > 0 and set k = K . along with the EV charging station location problem. Let θ (x)
2: while |G (k) | = n k do be the cost/price function of operating/installing one charging
3: If |G (k) | > n k , find x ∈ G (k) so that unit at x ∈ G. The total cost for the K charging stations at
0 < d(x, S( j )) − d(x, S(k) ) < ε and S = (S1 , . . . , SK ) is then denoted by
j = arg min d(x, S(i) ).
i<k 
K
Assign x to G ( j ) and increase the value of ε. Cost(S) = μk · θ (Sk ). (20)
4: If |G (k) | < n k , find x ∈ / G (k) so that k=1
0 < d(x, S(k) ) − d(x, S( j )) < ε and
In practice, a preset maximum allowable operation/installation
j = arg max d(x, S(i) ).
i<k cost is often considered for this type of problem, i.e.,
Assign x to G (k) and increase the value of ε.
5: end while 
K

6: Let k := k − 1. μk · θ (Sk ) ≤ C0 (21)

k=1
7. If k > 1, reorder G (k) , . . . , G (K ) and go back to line 2;
otherwise return all adjusted clusters and denote the for some C0 > 0. In this case, constraint (21) can be included
updated partition by P̃ = {G̃ 1 , . . . , G̃ K }. in the search procedure of Algorithm 1 so as to find the optimal
solution ( S̃ ∗ , P̃ ∗ ). Note that there exists a substantial body of
Although the SSDF and LSDF strategies adjust the size literature discussing how to optimize the charging station’s
of clusters in a different order, they share the same intuition installation and operation costs, which are often related to the
in determining the cluster boundaries. Specifically, if |G (k) | EV performance and management of power grids. For more
is larger than its constrained size, then observations in G (k) information, the readers can refer to [53], [54], [55], [56].
and furtherest away from the medoid S(k) are moved to the
neighboring clusters. On the other hand, if |G (k) | is smaller IV. I LLUSTRATION BASED ON A S IMULATION S TUDY
than its constrained size, then observations in the neighboring In practice, there are three main sources from which the
clusters and nearest to the medoid S(k) are moved to G (k) . charging demand data can be collected: (i) questionnaires
It should be noted that both strategies would produce nonlinear and/or surveys; (ii) chargers at public charging stations;

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.
23864 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 12, DECEMBER 2022

μ3 = 1 charging units, respectively, and the charging request

demand process has an input rate λ = 5.94. Since the total
charging capacity is μ1 +μ2 +μ3 = 6, the system then operates
under a heavy traffic regime with intensity ρ = 5.94/6 = 0.99.
Therefore, it is straightforward to choose rk proportional to
μk so that the three clusters have constrained sizes n 1 : n 2 :
n 3 = μ1 : μ2 : μ3 = 3 : 2 : 1. To mimic the charging
behavior of EV drivers in the two urban areas of Fig. 3 and
incorporate stochastic variations of charging demand requests,
we simulate the system by considering two demand location
densitiesfResults: (i) uniform distribution over G and (ii)
truncated bivariate Gaussian distribution so that it is skewed
Fig. 3. A map with uniformly meshed transportation system (a) and a map
with charging requests skewed to the upper-right corner (b), both extracted to the upper-right of G; associated respectively with two
from downtown Seattle, Washington. demand inter-arrival time distributions: (i) Gamma(1/5.94, 1)
and (ii) E x ponenti al(5.94) under various vehicle speeds v.
and (iii) EV driving patterns measured by GPS. However, Then, the proposed methodology for finding the optimal DLB
current data collected from these sources are not suitable routing and charging station locations is illustrated based
for our analysis due to lack of population and geographical on synthetic data. Note that we choose ε = 0.01 in both
location representation [57]. To overcome this issue, we will SSDF and LSDF strategies for adjusting the cluster size.
use synthetic demand location data generated by various Further, we choose the commonly used stochastic gradient
simulation models to illustrate the proposed approach. Fur- descent (SGD) method [58] in Algorithm 1 for finding the
ther, we consider the problem for three selected urban areas medoids. All the numerical results were obtained by using the
extracted from the Seattle metropolitan one in Washington software package RStudio (Version 1.1.456) and executed on
state (available at https://openchargemap.org/site). Note that 1.6 GHz Intel Core i5 processors with 4 GB of cache under
the first two areas, as shown in Fig. 3, have well-developed the 64-bit version of Microsoft Windows 10.
road networks with grid-like layout of streets, where d(x, Sk ) 1) Uniform Charging Request Locations: Suppose
is assumed to be proportional to the “Euclidean distance”, i.e., 1000 random charging request locations are uniformly
d(x, Sk ) ∝ ||x − Sk ||. Such setup facilitates the computation placed over G. Fig. 4(a) and Fig. 4(b) show the estimated
of cluster centers and the objective function in (19), thus clustering results and three charging station locations by
allowing us to conduct extensive numerical studies within a running Algorithm 1 with 40 iterations under the SSDF and
reasonable time, such as the validation of assumption (15), LSDF strategy, respectively. As can be seen, both strategies
execution of Algorithm 1, and comparisons of elapsed time yield very similar routing subregions and charging station
for various algorithms. The third area, as shown in Fig. 10, locations. For comparison purposes, the clustering result
has a bigger and irregular shape surrounding Lake Union obtained by using another size adjustment strategy proposed
and Portage Bay in the northbound of Seattle. In this case, by Zhu et al. [59] is given in Fig. 5(a). Note that this strategy
we consider a more realistic setting with d(x, Sk ) correspoding formulates the problem as a mixed-integer programming
to the shortest route distance between a demand location x and one and the solution is obtained by using the R package
a charging station Sk , which is estimated by Google Maps “Rsymphony” [60]. However, as can be seen, the strategy
API (https://developers.google.com/maps). Note that such a by Zhu et al. usually produces overlapping clusters. This
setting can improve the accuracy of finding the optimal routing makes the routing policy difficult to implement in practice.
boundaries and charging station locations, but nevertheless Further numerical comparisons of these three size adjustment
increases the computational complexity and requires a certain strategies are given in Fig. 5(b), where each strategy is run
amount of execution time. The numerical results for the first with 500 random initial medoids and the resulting values
two areas and the third area are given in Sec. IV-A and Sec. of the objective function in (19) are summarized in a box
IV-B, respectively. plot. Fig. 5(b) shows that the SSDF and LSDF strategy
perform fairly similar in terms of the resulting distribution
for the objective function in (19). On the other hand, the
A. Approximated Results Based on Euclidean Distance distributions of both strategies are significantly smaller than
As can be seen, Fig. 3(a) is associated with a uniformly that of the strategy by Zhu et al. Numerical results show that
meshed transportation system. Thus, it is reasonable to expect the strategy proposed by Zhu et al. yields a minimum value
uniform charging requests over the area. On the other hand, 2.743 × 103 of (19) among 500 replicated searches, which is
Fig. 3(b) is associated with a transportation system with charg- even larger than the maximum values of (19) under strategies
ing requests skewed to the upper-right corner - an area with SSDF and LSDF (which are 2.641 × 103 and 2.682 × 103 ,
exit ramps from Interstate Freeway I-5. Next, we consider an respectively).
EV charging system analogous to the transportation network of Fig. 6 depicts the mean travel and mean waiting time
these two urban areas, wherein three charging stations are to be over the station-search stage of Algorithm 1, where system
placed in. Assume each area is scaled to a square region G = is simulated with various EV speeds v and the charging
[0, 10] × [0, 10], the three stations have μ1 = 3, μ2 = 2 and requests inter-arrival time distribution E x ponenti al(5.94).

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.
HUNG AND MICHAILIDIS: NOVEL DATA-DRIVEN APPROACH 23865

Fig. 4. Illustrative clustering results based on 103 simulated charging request

locations from the uniform distribution under the DLB routing with size Fig. 6. The mean travel time (a) and mean waiting time (b) over the
adjustment strategies SSDF (a) and LSDF (b). station-search stage of Algorithm 1 for the system with various vehicle speeds
and demand inter-arrival time distribution E x ponential(5.94).

Fig. 5. An illustrative clustering result based on 103 simulated charging Fig. 7. Illustrative clustering results based on 103 simulated charging request
request locations from the uniform distribution under the size adjustment locations from the truncated bivariate Gaussian distribution under the DLB
strategy by Zhu et al. (a). Note that due to overlapping clusters, the charging routing with size adjustment strategies SSDF (a) and LSDF (b).
request locations associated with different routing stations are highlighted as
different colors. The box plots based on 500 estimated values of (19) under
three size adjustment strategies are shown in (b).

As can be seen, the mean travel time for all considered EV

speeds decreases monotonically as the station-search stage
increases and converges at S̃ ∗ . In addition, the mean travel time
decreases as the EV speed becomes larger, as expected. On the
other hand, the mean waiting time curve is rather flat over
the station-search stage for all EV speeds. These numerical
results render strong support for assumption (15) in Theorem
1. Consequently, the minimum mean response time for all
EV speeds and both demand inter-arrival time distributions
is also obtained at S̃ ∗ . Finally, note that for high EV speeds Fig. 8. An illustrative clustering result based on 103 simulated charging
(i.e. when the assumption v ≤ v̄ in Theorem 1 is violated), request locations from the truncated bivariate Gaussian distribution under the
size adjustment strategy by Zhu et al. (a). Note that due to overlapping clus-
the selection of charging stations locations will not exert a ters, the charging request locations associated with different routing regions
significant influence on the mean response time. are highlighted as different colors. The box plots based on 500 estimated
2) Skewed Charging Request Locations: Analogous clus- values of (19) under three size adjustment strategies are shown in (b).
tering results to the case with truncated bivariate Gaussian
charging request locations are shown in Fig. 7(a) and Fig. 7(b), strategies are shown in Fig. 8(b). As can be seen, the SSDF and
where Algorithm 1 was run for 40 iterations under the SSDF LSDF strategies yield similar results regarding the distribution
and LSDF strategy, respectively. As can be seen, both strate- of the objective function in (19). In addition, the distributions
gies yield very similar clustering results. It is worth noting of both strategies are significantly smaller than that of strategy
that the truncated bivariate Gaussian distribution concentrates by Zhu et al. Numerical results show that that the strategy by
the demand in the upper right corner, which determines the Zhu et al. yields a minimum value 1.618 × 103 of (19) among
location of the charging stations. Analogously, for compari- 500 replicated searches, while the minimum values under
son purpose the clustering result obtained by using the size strategies SSDF and LSDF are 1.468 × 103 and 1.447 × 103 ,
adjustment strategy by Zhu et al. is given in Fig. 8(a), respectively.
while the box plots for the objective function in (19) based Fig. 9 yields the detailed mean travel and mean waiting time
on 500 replicated searches under the three size adjustment over the station-search stage of Algorithm 1, where the system

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.
23866 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 12, DECEMBER 2022

Fig. 9. The mean travel time (a) and mean waiting time (b) over the Fig. 10. An illustration of shortest route distance calculated by Google
station-search stage of Algorithm 1 for the system with various vehicle speeds Maps (a) and the clustering result by running Algorithm 1 with grid search
and demand inter-arrival time distribution Gamma(1/5.94, 1). for finding the medoids, size constraint n1 : n2 : n3 = 3 : 2 : 1 and size
adjustment strategy SSDF based on 120 simulated uniform charging demand
TABLE I
locations (b). Note that dots of the same color constitute a cluster (or routing
T HE E LAPSED T IME ( IN S ECONDS ) OF RUNNING THE C LUSTERING A LGO - region), while the three large colored icons refer to the estimated cluster
RITHM 500 T IMES W ITH T HREE S IZE A DJUSTMENT S TRATEGIES centers (i.e., charging station locations).
LSDF, SSDF AND Z HU et al

with the grid search for finding the medoids, size constraint
n1 : n2 : n3 = 3 : 2 : 1 and size adjustment strategy
SSDF, is given in Fig. 14(b). As can be seen, the area with
the same colored dots refers to the routing subregion, while
the associated charging station is placed on the center found
is simulated with various EV speeds v and charging demand by grid search. Note that this clustering result yields an
inter-arrival time distribution Gamma(1/5.94, 1). As can be approximate solution for (19). To improve the approximation,
seen, all results are fairly similar to those shown in Fig. 6 and one may perform the search a number of times by varying
hence similar remarks apply. the initial condition in Algorithm 1 and find the best solution
3) Algorithm Efficiency: Tab. I shows the elapsed time of among all search results.
running the clustering algorithm 500 times for the following
size adjustment strategies: LSDF, SSDF and Zhu et al. The V. C ONCLUSION
elapsed time of the SSDF strategy is the smallest for both uni-
form and truncated Gaussian service location densities. On the In this paper, we address the location-routing problem
other hand, the elapsed time of the strategy by Zhu et al. is for a general EV charging system with a fixed number of
20% and 92% larger than that of LSDF and SSDF for the charging stations (with predetermined capacity) and stochastic
uniform density and 30% and 63% larger than that of LSDF charging demands. The objective is to determine an optimal
and SSDF for the truncated Gaussian density. In summary, routing strategy, as well as the locations of the charging
numerical evidence from these experiments shows that the stations, so as to minimize the system’s mean response
proposed strategies LSDF and SSDF significantly outperform time. We establish that under both light and heavy traffic
the strategy by Zhu et al. in terms of both accuracy and speed. regimes, the primary optimization problem can be approxi-
mately formulated as the K-medoids clustering problem with
size constraints. Thus, a complex LRP can be efficiently solved
B. Results Based on Shortest Route Distance by utilizing information collected from a large number of
Next, we consider a larger and non-regular urban area observed/historical charging request data. In this clustering
of Seattle shown in Fig. 10(a), wherein each point on the formulation, the “medoids” correspond to the optimal locations
map is assigned its latitude and longitude. Note that d(x, Sk ) of charging stations, while each resulting cluster corresponds
now represents the shortest route distance calculated by the to a subregion in which all the charging requests are routed
Distance Matrix API provided by the Google Maps Plat- to the station located at the medoid. Note that solving the
form (https://developers.google.com/maps), and three charging K-medoids clustering problem is not a trivial task, especially
stations with capacities μ1 : μ2 : μ3 = 3 : 2 : 1 are when the distance is non-Euclidean and a size constraint is
to be placed in. Since d(x, Sk ) is obviously non-Euclidean, placed on each cluster. Two strategies SSDF and LSDF were
traditional gradient search or particle swarm algorithms are presented for the latter task, while Algorithm 1 is designed
no longer applicable. Therefore, we will use “grid search” for based on gradient or grid search for finding the medoids.
finding/updating the clusters’ medoids in Algorithm 1. Due to Selected simulations show very good results for systems with
limited computation resources, we consider a simplified setup various charging request inter-arrival distributions and location
with (i) 120 simulated charging demand locations uniformly densities for vehicle speeds below a predetermined threshold.
distributed over the service region; and (ii) 124 candidate grid In other situations when the mean waiting time becomes non-
points uniformly imposed on the service region for identifying negligible, one may search the neighborhood of S̃ ∗ so as to
cluster centers. The result based on running Algorithm 1, find the optimal solution.

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.
HUNG AND MICHAILIDIS: NOVEL DATA-DRIVEN APPROACH 23867

Discussion: There are several problems that constitute topics [2] X. Liu, R. Anand, G. Xiong, X. Shang, and X. Liu, Big Data and Smart
of future research, including: (i) how to best allocate the Service Systems. London, U.K.: Academic Press, 2017.
[3] L. Silva et al., “Computing paradigms in emerging vehicular envi-
charging station capacity along with determining location; ronments: A review,” IEEE/CAA J. Autom. Sinica, vol. 8, no. 3,
(ii) how to formulate the optimization problem, if bottleneck pp. 491–511, Mar. 2021.
effects and/or congestion of the transportation network are [4] J. Shao, H. Yang, and A. Zhang, “Adoption of electric vehicles: Man-
ufacturers’ incentive and government policy,” J. Transp. Econ. Policy,
taken into account; (iii) how to modify the proposed routing vol. 53, no. 2, pp. 175–198, 2019.
strategy so as to deal with the problem raised by EV battery [5] A. Jenn, K. Springel, and A. R. Gopal, “Effectiveness of electric vehicle
range; (iv) how to solve the optimization problem if the system incentives in the United States,” Energy Policy, vol. 119, pp. 349–356,
Aug. 2018.
has a time-varying charging request arrival rate and/or location [6] C. Luo, M. Leng, J. Huang, and L. Liang, “Supply chain analysis under
density; and (v) how to deal with the problem if finite buffers a price-discount incentive scheme for electric vehicles,” Eur. J. Oper.
are assumed. For (i), suppose the total charging capacity Res., vol. 235, no. 1, pp. 329–333, May 2014.
μ = μ1 + . . . + μ K is controlled and the planner wishes [7] E. S. Rigas, S. D. Ramchurn, and N. Bassiliades, “Managing electric
vehicles in the smart grid using artificial intelligence: A survey,” IEEE
to determine the best capacity allocation (μ1 , . . . , μ K ) in the Trans. Intell. Transp. Syst., vol. 16, no. 4, pp. 1619–1635, Aug. 2015.
primary optimization problem. Then, even for a simple setting [8] C. Luo, Y.-F. Huang, and V. Gupta, “Placement of EV charging
with K = 3 and μ = 10, there are 2×(1+2+3+4) = 20 ways stations—Balancing benefits among multiple entities,” IEEE Trans.
Smart Grid, vol. 8, no. 2, pp. 759–768, Mar. 2017.
of capacity allocation for finding the optimal solution, thus [9] R. Mehta, D. Srinivasan, A. Khambadkone, J. Yang, and A. Trivedi,
requiring 20 times computational cost. Note that capacity “Smart charging strategies for optimal integration of plug-in electric
allocation and charging station locations are closely related vehicles within existing distribution system infrastructure,” IEEE Trans.
Smart Grid, vol. 9, no. 1, pp. 299–312, Jan. 2016.
to the operation of power systems. Therefore, they are often [10] H. Zhang, S. Moura, Z. Hu, and Y. Song, “PEVfast-charging station
considered together with the energy management and security siting and sizing on coupled transportation and power networks,” IEEE
control of power grids (see [20], [22], [61] and references Trans. Smart Grid, vol. 9, no. 4, pp. 2595–2605, Jul. 2018.
[11] T. D. Chen, K. M. Kockelman, and M. Khan, “Locating electric
therein). For (ii), we can consider a “time-dependent” DLB vehicle charging stations: Parking-based assignment method for Seattle,
routing scheme so that the problem can be solved based on Washington,” Transp. Res. Rec., J. Transp. Res. Board, vol. 2385, no. 1,
data collected over the time period with bottleneck effect pp. 28–36, 2013.
[12] A. Y. S. Lam, Y.-W. Leung, and X. Chu, “Electric vehicle charging
or traffic congestion. Further, the “shortest route” can be station placement: Formulation, complexity, and solutions,” IEEE Trans.
replaced by the “fastest route” so as to better estimate the Smart Grid, vol. 5, no. 6, pp. 2846–2856, Nov. 2014.
EV travel time (note that Google Maps can also provide [13] X. Wang, C. Yuen, N. U. Hassan, N. An, and W. Wu, “Electric vehicle
charging station placement for urban public bus systems,” IEEE Trans.
instant fastest route estimation). Then, the primary objective Intell. Transp. Syst., vol. 18, no. 1, pp. 128–139, Jan. 2017.
function can be directly formulated as the total travel time [14] Q. Cui, Y. Weng, and C.-W. Tan, “Electric vehicle charging station
and an analogous clustering problem with size constraints can placement method for urban areas,” IEEE Trans. Smart Grid, vol. 10,
no. 6, pp. 6552–6565, Nov. 2019, doi: 10.1109/TSG.2019.2907262.
be solved. For (iv), we would like to highlight an important
[15] F. Baouche, R. Billot, R. Trigui, and N.-E. El Faouzi, “Efficient
fact that the primarily observed charging demand locations allocation of electric vehicles charging stations: Optimization model and
x 1 , . . . , x n are often dependent on the routes to extant charging application to a dense urban network,” IEEE Intell. Transp. Syst. Mag.,
stations. Thus, once the charging stations are replaced based vol. 6, no. 3, pp. 33–43, Jul. 2014.
[16] F. He, D. Wu, Y. Yin, and Y. Guan, “Optimal deployment of public
on the solution provided by Algorithm 1, the demand location charging stations for plug-in hybrid electric vehicles,” Transp. Res. B,
density f (x) may change as well. Such time-varying property Methodol., vol. 47, pp. 87–101, Jan. 2013.
suggests that monitoring the change of f (x), obtaining more [17] N. Sathaye and S. Kelley, “An approach for the optimal planning of
electric vehicle infrastructure for highway corridors,” Transp. Res. E,
representative data, and resolving the location-routing problem Logistics Transp. Rev., vol. 59, pp. 15–33, Nov. 2013.
are necessary for finding a stable solution. For (v), we can [18] Y. Xiong, J. Gan, B. An, C. Miao, and A. L. C. Bazzan, “Optimal electric
examine if condition (15) in Theorem 1 is still applicable vehicle charging station placement,” in Proc. 24th Int. Conf. Artif. Intell.
Buenos Aires, 2015, pp. 2662–2668.
(but it may require other constraints) or simply formulate the [19] H. Xu, S. Miao, C. Zhang, and D. Shi, “Optimal placement of charging
problem that minimizes the mean travel time (the finite-buffer infrastructures for large-scale integration of pure electric vehicles into
assumption mainly influences the waiting time for charging). grid,” Int. J. Elect. Power Energy Syst., vol. 53, pp. 159–165, Dec. 2013.
[20] S. Deb, K. Tammi, K. Kalita, and P. Mahanta, “Impact of electric vehicle
For the later case, Algorithm 1 can still be used to find the charging station load on distribution network,” Energies, vol. 11, no. 1,
solution, while the current size constraints may be modified p. 178, Jan. 2018.
or dropped. In practice, solving these problems will require [21] Z. Jiang, H. Tian, M. J. Beshir, S. Vohra, and A. Mazloomzadeh,
“Analysis of electric vehicle charging impact on the electric power grid:
a considerable amount of computational load, reformulation Based on smart grid regional demonstration project—Los Angeles,” in
of the problem, and/or the development of more sophisticated Proc. IEEE PES Transmiss. Distrib. Conf. Expo.-Latin Amer. (PESTD-
routing strategies. LA), Morelia, Mexico, Sep. 2016, pp. 1–5.
[22] W. Khan, F. Ahmad, and M. S. Alam, “Fast EV charging station
ACKNOWLEDGMENT integration with grid ensuring optimal and quality power exchange,”
Eng. Sci. Technol., Int. J., vol. 22, no. 1, pp. 143–152, Feb. 2019.
The authors would like to thank the Editor, Associate Editor, [23] Q. Kong, M. Fowler, E. Entchev, H. Ribberink, and R. McCallum,
and five anonymous referees for their valuable comments and “The role of charging infrastructure in electric vehicle implementation
within smart grids,” Energies, vol. 11, no. 12, p. 3362, Dec. 2018.
suggestions.
[24] T. Chen, B. Zhang, H. Pourbabak, A. Kavousi-Fard, and W. Su, “Optimal
R EFERENCES routing and charging of an electric vehicle fleet for high-efficiency
dynamic transit systems,” IEEE Trans. Smart Grid, vol. 9, no. 4,
[1] S. Campos-Cordobés et al., “Chapter 5—Big data in road transport and pp. 3563–3572, Jul. 2018.
mobility research,” in Intelligent Vehicles, F. Jiménez, Ed. Oxford, U.K.: [25] Y.-C. Hung and G. Michailidis, “Optimal routing for electric vehicle
Butterworth-Heinemann, 2018, pp. 175–205. service systems,” Eur. J. Oper. Res., vol. 247, no. 2, pp. 515–524, 2015.

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.
23868 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 23, NO. 12, DECEMBER 2022

[26] W. Yuan, J. Huang, and Y. J. A. Zhang, “Competitive charging station [50] J. Snyman, Practical Mathematical Optimization: An Introduction to
pricing for plug-in electric vehicles,” IEEE Trans. Smart Grid, vol. 8, Basic Optimization Theory and Classical and New Gradient-Based
no. 2, pp. 627–639, Mar. 2017. Algorithms. Boston, MA, USA: Springer, 2005.
[27] A. Aghalari, D. Salamah, C. Marino, and M. Marufuzzaman, “The elec- [51] Q. Kang, S. Liu, M. C. Zhou, and S. Li, “A weight-incorporated
tric location routing problem under ambient temperature,” in Proc. IISE similarity-based clustering ensemble method based on swarm intelli-
Annu. Conf., 2021, pp. 710–715. gence,” Knowl.-Based Syst., vol. 104, pp. 156–164, Jul. 2016.
[28] A. Arias, J. D. Sanchez, and M. Granada, “Integrated planning of electric [52] X. Xu, J. Li, M. Zhou, J. Xu, and J. Cao, “Accelerated two-stage particle
vehicles routing and charging stations location considering transportation swarm optimization for clustering not-well-separated data,” IEEE Trans.
networks and power distribution systems,” Int. J. Ind. Eng. Comput., Syst., Man, Cybern. Syst., vol. 50, no. 11, pp. 4212–4223, Nov. 2020.
vol. 9, no. 4, pp. 535–550, 2018. [53] J. Hu, S. You, M. Lind, and J. Ostergaard, “Coordinated charging of
[29] H. Çalik, A. Oulamara, C. Prodhon, and S. Salhi, “The electric location- electric vehicles for congestion prevention in the distribution grid,” IEEE
routing problem with heterogeneous fleet: Formulation and benders Trans. Smart Grid, vol. 5, no. 2, pp. 703–711, Mar. 2014.
decomposition approach,” Comput. Oper. Res., vol. 131, Jul. 2021, [54] Y. Qin, Z. Bei, and M. Kezunovic, “Optimized operational cost reduction
Art. no. 105251. for an EV charging station integrated with battery energy storage and
PV generation,” IEEE Trans. Smart Grid, vol. 10, no. 2, pp. 2096–2106,
[30] J. C. Paz, M. Granada-Echeverri, and J. W. Escobar, “The multi-depot
Mar. 2018.
electric vehicle location routing problem with time Windows,” Int. J. Ind.
[55] M. Z. Zeb et al., “Optimal placement of electric vehicle charging
Eng. Comput., vol. 9, no. 1, pp. 123–136, 2018.
stations in the active distribution network,” IEEE Access, vol. 8,
[31] L. Wang and Y. Song, “Multiple charging station location-routing pp. 68124–68134, 2020.
problem with time window of electric vehicle,” J. Eng. Sci. Technol. [56] G. Zhou, Z. Zhu, and S. Luo, “Location optimization of electric vehicle
Rev., vol. 8, no. 5, pp. 190–201, 2015. charging stations: Based on cost model and genetic algorithm,” Energy,
[32] J. Yang and H. Sun, “Battery swap station location-routing problem with vol. 247, May 2022, Art. no. 123437.
capacitated electric vehicles,” Comput. Oper. Res., vol. 55, pp. 217–232, [57] E. Hartvigsson, N. Jakobsson, M. Taljegard, and M. Odenberger, “Com-
Mar. 2015. parison and analysis of GPS measured electric vehicle charging demand:
[33] H. Zheng and S. Peeta, “Routing and charging locations for electric The case of western Sweden and Seattle,” Frontiers Energy Res., vol. 9,
vehicles for intercity trips,” Transp. Planning Technol., vol. 40, no. 4, Oct. 2021, Art. no. 730242.
pp. 393–419, May 2017. [58] S. Shalev-Shwartz and S. Ben-David, Understanding Machine Learning:
[34] D. Efthymiou, K. Chrysostomou, M. Morfoulaki, and G. Aifantopoulou, From Theory to Algorithms. Cambridge, U.K.: Cambridge Univ. Press,
“Electric vehicles charging infrastructure location: A genetic algorithm 2014.
approach,” Eur. Transp. Res. Rev., vol. 9, no. 27, pp. 1–9, Jun. 2017. [59] S. Zhu, D. Wang, and T. Li, “Data clustering with size constraints,”
[35] C. Bian, H. Li, F. Wallin, A. Avelin, L. Lin, and Z. Yu, “Finding Knowl.-Based Syst., vol. 23, no. 8, pp. 883–889, Dec. 2010.
the optimal location for public charging stations—A GIS-based MILP [60] R. Harter, K. Hornik, S. Theussl, C. Szymanski, and
approach,” Eenrgey Proc., vol. 158, pp. 6582–6588, Feb. 2019. F. Schwendinger. (2017). Rsymphony: SYMPHONY in R. R. Package
[36] R. Pagany, A. Marquardt, and R. Zink, “Electric charging demand loca- Version 0.1-28. [Online]. Available: https://cran.r-project.org/web/
tion model—A user and destination-based locating approach for electric packages/Rsymphony/Rsymphony.pdf
vehicle charging stations,” Sustainability, vol. 11, no. 8, pp. 1–15, 2019. [61] I. S. Bayram, G. Michailidis, M. Devetsikiotis, and F. Granelli, “Electric
[37] N. Andrenacci, R. Ragona, and G. Valenti, “A demand-side approach power allocation in a network of fast charging stations,” IEEE J. Sel.
to the optimal deployment of electric vehicle charging stations in Areas Commun., vol. 31, no. 7, pp. 1235–1246, Jul. 2013.
metropolitan areas,” Appl. Energy, vol. 182, pp. 39–46, Nov. 2016.
[38] A. Sarker, H. Shen, and J. A. Stankovic, “MORP: Data-driven multi-
objective route planning and optimization for electric vehicles,” Proc.
ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 1, no. 4,
pp. 1–35, Jan. 2018.
[39] Y.-C. Hung, H. P. Lok, and G. Michailidis, “Optimal routing for
electric vehicle charging systems with stochastic demand: A heavy Ying-Chao Hung received the B.A. degree in
traffic approximation approach,” Eur. J. Oper. Res., vol. 299, no. 2, mathematics from the National Taiwan University,
pp. 526–541, Jun. 2021, doi: 10.1016/j.ejor.2021.06.058. Taiwan, in 1993, the M.B.A. degree from the
[40] A. G. Budrin, E. V. Budrina, A. S. Lebedeva, L. I. Rogavichene, and National Chengchi University, Taiwan, in 1995, and
K. B. Kvitko, “Intelligent transport systems as an integration platform the Ph.D. degree in statistics from the University of
for creating a network of regional transport and logistics complexes,” Michigan in 2002. In 2009, he joined the Depart-
in Proc. Int. Conf. Quality Manage., Transp. Inf. Secur., Inf. Technol. ment of Statistics, National Chengchi University,
(ITQMIS), Sep. 2019, pp. 176–181. where he became a Full Professor in 2015. In 2022,
[41] H.-H. Pan, S.-C. Wang, and K.-Q. Yan, “An integrated data exchange he joined the Institute of Industrial Engineering,
platform for intelligent transportation systems,” Comput. Standards National Taiwan University. His research interests
Interfaces, vol. 36, no. 3, pp. 657–671, Mar. 2014. include control and optimization of stochastic sys-
[42] B. W. Silverman, Density Estimation for Statistics and Data Analysis. tems, statistical machine learning, applied probability, and computational
New York, NY, USA: Chapman & Hall, 1986. statistics.
[43] Z. Drezner and H. W. Hamacher, Facility Location: Applications and
Theory. New York, NY, USA: Springer, 2004.
[44] H. Min, V. Jayaraman, and R. Srivastava, “Combined location-routing
problems: A synthesis and future research directions,” Eur. J. Oper. Res.,
vol. 108, no. 1, pp. 1–15, Jul. 1998.
[45] C. Ortiz-Astorquiza, I. Contreras, and G. Laporte, “Multi-level facility
George Michailidis (Member, IEEE) received the
location problems,” Eur. J. Oper. Res., vol. 267, no. 3, pp. 791–805,
B.S. degree in economics from the University of
Jun. 2018.
Athens, Greece, in 1987, and the M.A. degrees in
[46] W. Whitt, “Approximations for the GI/G/M queue,” Prod. Oper. Manag., economics and mathematics and the Ph.D. degree in
vol. 2, no. 2, pp. 114–161, Jan. 2009. mathematics from UCLA. After a post-doctoral in
[47] Z. Rosberg, “Bounds on the expected waiting time in a GI/G/1 queue: operations research at Stanford University, in 1998,
Upgrading for low traffic intensity,” J. Appl. Probab., vol. 24, no. 3, he joined the Department of Statistics, University
pp. 749–757, Sep. 1987. of Michigan, where he became a Full Professor in
[48] L. Kaufman and P. J. Rousseeuw, Clustering by Means of Medoids. 2008. In 2015, he joined the University of Florida
Amsterdam, The Netherlands: North-Holland, 1987. as the Founding Director of the Informatics Insti-
[49] E. Schubert and P. J. Rousseeuw, “Faster k-medoids clustering: tute. His research interests include network analysis,
Improving the PAM, CLARA, and CLARANS algorithms,” 2018, queueing theory, stochastic control and optimization, applied probability, and
arXiv:1810.05691. machine learning.

Authorized licensed use limited to: VIT University. Downloaded on January 05,2023 at 06:43:36 UTC from IEEE Xplore. Restrictions apply.