Contents lists available at ScienceDirect

,: -
Expert
Systems

-
with

ed
Applications
Expert Systems With Applications Anlnle<nollOnol

ELSEVIER journal homepage: www.elsevier.com/locate/eswa

iew
Optimizing a System of Electric Vehicle Charging Stations using Mixed Integer Linear
Programming Computer Experiments
Ukesh Chawal a,b, Jay Rosenberger a, Victoria C.P. Chen a, Wei J. Leec Fellow, IEEE, Mewan
Wijemannea, Raghavendra K. Punugua, Asama Kulvanitchaiyanunta

ev
a Department of Industrial and Manufacturing Systems Engineering, The University of Texas at Arlington, Arlington, TX 76019, United States
b Boeing Research & Technology Integrated Vehicle Systems, The Boeing Company, 2750 Regent Blvd, Dallas, TX 75261, United States
c Department of Electrical Engineering. The University of Texas at Arlington, Arlington, TX 76019, United States

ARTICLE INFO ABSTRACT

r
Keywords: Owing to concerns related to both the economy and the environment regarding fossil fuels, along with an
Design and Analysis of Computer- increasing emphasis on sustainability, the utilization of electric vehicles (EVs) presents itself as a viable remedy.
Experiments
Electric Vehicle Charging Stations
Latin Hypercube Sampling
Metamodel
Mixed Integer Linear Programming
er
Nevertheless, for a widespread adoption of EVs, having convenient access to charging infrastructure is a
fundamental requirement. This paper presents a model based on mixed integer linear programming (MILP) to
efficiently plan a network of EV charging stations, pinpoint suitable locations for these stations, determine the
optimal count of charging slots for deployment, and estimate potential profits stemming from meeting the demand
while also engaging in electricity trading with the grid. Our methodology introduces a two-stage framework that
blends the first-stage system design problem with a second-stage control problem of the EV charging stations and
pe
Multivariate Adaptive Regression-
develops a design and analysis of computer experiments (DACE) based system design optimization solution
Splines
method. In the first phase, the system layout is configured to maximize potential profits. Our DACE strategy
creates a metamodel to predict revenue from the control problem using multivariate adaptive regression splines
(MARS), built upon a categorized Latin hypercube (LH) experimental setup. With the revenue insights derived
from the control problem, the system design conundrum is subsequently tackled. The outcomes derived from the
DACE based system design approach, upon comparison with the MILP model, exhibit a solution that is nearly
optimal. Additionally, the DACE approach yields highly interpretable profit functions that show the marginal
profits as a function of the number of slots opened at each station. Moreover, the computational time needed by
ot

the DACE approach is substantially reduced, rendering it a more practical choice for real-world applications.

1. Introduction operational cost for EV charging stations and maximizes the profit in
The energy shortages of the 1970s spurred the exploration of running the stations. Through the generation of electricity using
tn

alternative energy sources for vehicles, leading to the initiation of renewable sources, charging stations could actively engage in the
electric vehicle (EV) research. In the present times, the increasing electricity market. This additional dynamic provides even more
demand for sustainability has greatly elevated the significance of motivation for the U.S. transportation sector to seek out cleaner
electric vehicles (EVs). A recent environmental evaluation conducted alternatives.
by the Electric Power Research Institute and the Natural Resource
Defense Council has indicated that opting for electricity in place of 1.1 Literature Review
gasoline/petroleum has the potential to substantially decrease emissions Several papers in the literature have studied optimal planning for EV
rin

of greenhouse gases and other airborne pollutants (He et al. 2015). To charging station, Plug-in Hybrid vehicles (PHEV), charging pads, and
bolster the usage of EVs, some governments have taken a variety of battery swapping.
initiatives. Norway is one of the countries working towards the goal of
having all new car sales be electric vehicles by 2025, to meet their 1.1.1 EV Charging Station Literature Review
emission reduction targets (Lambert 2016). By endorsing both current Some papers proposed using heuristics for optimal planning of
and forthcoming technologies in electric power-based vehicular charging station locations. Tang et al. (2013) proposed a Particle Swarm
ep

products, the transportation industry anticipates transitioning from its Optimization for the planning of EV charging stations. This model
oil-dependent design to a cleaner and more environmentally sustainable considered an initial fixed investment cost, charging cost, operating costs,
electric-based design. One of the critical factors that requires attention service radius and capacity of charging station, etc., to determine the
when it comes to EV charging infrastructure is the driving range and layout of EV charging stations. Lin & Hua (2015) proposed a flow
accessibility of charging stations. This limitation has led to a capturing location model that uses Particle Swarm Optimization for
constrained adoption of electric vehicles. One of the solutions to this selecting locations for EV charging stations. This study considered 25
problem is installation of EV charging stations, which minimizes nodes with 42 arcs for setting up EV charging stations while accounting
Pr

Abbreviations: DACE, Design and Analysis of Computer Experiments; EV, Electric Vehicle; LH, Latin Hypercube; MARS, Multivariate Adaptive Regression
Splines; MILP, Mixed Integer Linear Programming.
* Corresponding author.
E-mail address: [email protected] (U. Chawal).

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4572175
for initial construction costs of facilities, the maximum charging al. (2022) proposed a model for determining optimal locations of EV
distance, and the cost of power network loss. Awasthi et al. (2017) charging stations to maximize the number of EVs while considering user-
proposed a hybrid algorithm based on a Genetic Algorithm and an specific characteristics, which represented the decision of the user to
improved version of a conventional Particle Swarm Optimization to find purchase EVs. This study focused on supporting a plan for maximizing EV

ed
an optimal placement of charging stations while accounting for initial adoption. Chang et al. (2014) proposed an extension to the Flow Refueling
investment cost and distribution grid power quality. Vazifeh et al. (2019) Location Model, which considered the allocation of both charging stations
developed a data-driven approach for EV charging stations employing a and charging pads to optimize the flow of recharged EVs. This study
Genetic Algorithm applied to a geographical grid to minimize total excess proposed deploying charging pads on the road network to capture more
driving distance to charging stations, energy overhead, and the number of traffic than captured by charging stations alone.
charging stations. Yang et al. (2021) presented a hybrid approach of You et al. (2015) proposed a novel cooperative charging strategy for an
differential evolutionary algorithm and Particle Swarm Optimization intelligent charging station. This strategy was designed to operate effectively

iew
which was applied to solve a charging station location model while in a dynamically shifting electricity pricing context. In this system, electric
considering charging station capacity and total charging costs. In this vehicles (EVs) shared energy stored in their batteries amongst each other
research, a Voronoi diagram was used to partition the service coverage area under the supervision of an aggregator. This collaborative approach granted
of the charging stations. the aggregator enhanced flexibility in managing schedules. This problem
Biesinger et al. (2017) presented an urban station-based car-sharing was formulated as a scheduling MILP model to capture discrete states of the
approach where the users can rent and return publicly available EV cars battery (charging, idle, and discharging). Zhang et al. (2023) presented a
from charging stations. In this research, a heuristic algorithm for finding day-ahead optimized dispatching technique for a distribution network (DN)
and designing charging stations was formulated as a bi-level model in that incorporates a fast-charging station (FCS). This FCS is integrated with
which the first level is to decide the station location, the number of charging photovoltaic (PV) systems and energy storage (ES) mechanisms, aiming to

ev
slots per station, and the total number of cars using a variable neighborhood alleviate the adverse effects of the FCS on the distribution network. Initially,
search algorithm. A path-based heuristic was used in the second level to historical vehicle travel data served as the foundation for employing a Monte
determine which trips are accepted by the system. To solve a dynamic Carlo simulation method (MCSM) to replicate the fast-charging load.
control of a system of plug-in hybrid vehicle charging stations problem, Finally, the uncertainties associated with photovoltaic (PV) power were
Kulvanitchaiyanunt et al. (2015) proposed and formulated a finite horizon managed through the implementation of a robust optimization model that
stochastic program. The objective of the study was to maximize profit, pertained to the economic operation of the distribution network (DN). Some

r
which is the revenue from selling power back to the power grid and the researchers proposed battery swapping, either as an alternative to charging
charging of vehicles minus the cost of buying electricity from the power pads or in addition to charging stations, for greater reach. Mak et al. (2013)
grid. This research considered wind energy, solar power generation, total proposed a robust optimization model for deploying battery-swapping
demand at each station and market price at each node. Sadeghi et al. (2022) infrastructure where a depleted battery can be exchanged for a recharged one
proposed a bi-objective mixed integer linear programming (MILP) model
to determine optimal locations for charging stations while taking into
consideration the number of chargers to be set up at each station and their
types. This model aimed to minimize the total cost as well as users’
dissatisfaction. This study used Lagrangian Relaxation to handle the
er in the middle of long trips. Kang et al. (2016) proposed a centralized
charging approach for electric vehicles (EVs) incorporating battery
swapping. This strategy factored in optimal charging precedence and
charging site selection based on real-time electricity prices at specific
locations. In this study, a population based heuristic approach was suggested
pe
complexity of the model. Arslan & Karasan (2016) presented a charging to minimize the total charging cost, as well as to reduce the power loss and
station location problem with plug-in hybrid vehicles as a generalization to voltage deviation of the power network. To minimize the cost and land use,
the flow recapture location problem to maximize the vehicle miles traveled Chen & Hua (2014) proposed a new location model based on set covering.
using electricity and thereby minimize the total cost of transportation under This model hinged on the use of existing gas stations as potential locations
the existing structure between electricity and gasoline. The authors to determine the arrangement of charging and battery swapping stations. Wu
proposed an arc-cover formulation and Benders decomposition algorithms et al. (2015) proposed using a Genetic Algorithm to solve an optimization
as exact solution methods. model to maximize the number of batteries in stock and minimize the cost
Brandstatter et al. (2017) proposed a robust integer linear optimization due to different charging schemes.
method to determine optimal locations for charging stations of electric car-
ot

sharing systems. Battistelli et al. (2012) developed a stochastic 1.1.2 Design and Analysis of Computer Experiments Literature Review
programming framework that considered uncertainty associated with In this paper, we develop a two-stage framework, which addresses the
vehicle-to-grid and wind power scenarios. Khosrojerdi et al. (2012) design of a system of EV charging stations using a design and analysis of
presented a linear mathematical model to optimize the cost of power computer experiments (DACE) based system design optimization approach.
trading, which used auto regressive methods to forecast wind power output DACE based optimization was first conducted for continuous-state
tn

and market clearing price for energy. Ma & Zhang (2018) proposed a stochastic dynamic programming (SDP) by Chen et al. 1999 and Chen 1999
queuing theory model to optimize the sizing and the location of charging and has since been applied for numerous high-dimensional dynamic
stations and solve using an exhaustion search method. This research used optimization applications, including airline optimization (Chen et al. 2003,
a Bass model developed by Frank Bass to predict the total number of EVs Siddappa et al. 2007), water resources (Tsai et al. 2004, Cervellera et al.
and to calculate the size of the charging stations. Gorbunova & Anisimov 2006), environmental quality control strategies (Yang et al. 2009, Sule et al.
(2020) developed a model for optimal selection of the limited number of 2011, Fan et al. 2018, Ariyajunya et al. 2021), and pain management (Lin et
charging stations to meet maximum demand while minimizing the total al. 2014). The concept of DACE based optimization for two-stage stochastic
rin

cost of operating the charging infrastructure. This model considered programming was first proposed by Chen 2001 and first demonstrated by
characteristics of the road network, traffic flow at nodes and places of Pilla et al. 2008 and Shih et al. 2014 using an airline fleet assignment case
attractions for the population. Tan & Lin (2014) proposed a stochastic study.
model for back-up flow capturing demand to ensure stability in service
coverage. Soltani et al. (2014) put forward a similar stochastic model to 1.2 Contribution
maximize profit based on price responsiveness of customers. While the previously mentioned papers put forward various methods
ep

Gerding et al. (2013) presented a two-sided market for advanced and implemented several algorithms for the optimization of locations for EV
reservation to reduce the queue faced by customers and uncertainty over charging stations, a global optimal set of stations to be opened, with the
the availability of a charging facility. In this method, the EV owners corresponding number of slots, has never be found while accounting for
reported their preference of time and charging location, while charging factors such as the customer demand obtained from the city population
stations reported their availability and cost. Park et al. (2014) proposed a (hotspots) of EVs, the distances from hotspots to the stations, and available
reservation recommendation algorithm to select the charging stations based solar energy and wind energy generation. To tackle these concerns, we
on distance and route. It provided recommendations of three charging present a deterministic MILP model designed to achieve a global optimal
Pr

stations based on: (i) desired amount of charge without waiting time, (ii) selection of stations for opening, with the primary goal of maximizing
desired amount of charge with waiting time, and (iii) the limited amount of profits. In this study, we examine 11 potential charging station sites across
charge with waiting time. It also helped in time slot management if there various time intervals to cater to 140 demand hotspots.
was a need to reserve a charging station slot. Yudovina & Michaildis The methodological contribution of this paper is the DACE based
(2015) proposed a decentralized policy of assigning electric vehicles to a optimization for system design formulated using a two-stage framework.
network of charging stations with the goal to achieve little to no queue for The first stage specifies the design of the system that maximizes expected
optimal location deployment of the charging infrastructure. Lamontagne et profit, and the second stage solves the system control problem. The first

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4572175
stage design optimization incorporates both the operational costs of charging slots at an opened station for the entire time horizon. Moreover, we
stations and the revenue from a system control problem in the second stage. discretize the time horizon into a fixed set of discrete time periods.
The “design” part of the DACE approach uses design of experiments to The demand model assumes that the demand at charging stations is
organize a set of feasible system designs of EV charging stations. In this based on their proximity to the hotspots. The closest station to a hotspot
paper, we develop a new design of experiments approach, referred to as a captures a fraction of the total hotspot demand, which is a linear function of
binned Latin Hypercube (LH), to sample points in the system design space. the distance between them. All other stations do not capture any demand

ed
Although the binned LH experimental design is developed specifically for from the hotspot. We also assume that stations that are sufficiently far away
the EV charging station problem, it has the potential to be used in other from a hotspot are unable to fulfill any demand of that hotspot. The capacity
capacity planning applications where both the decisions to open and of a charging station is determined by the number of open slots. We assume
determine capacity are considered simultaneously. For each of these each station can open at most a fixed maximum number of slots. In instances
system designs, we execute a second-stage control problem to obtain the where the demand at a charging station surpasses its capacity, a portion of
corresponding expected revenue. The “analysis” part of the DACE customers display a willingness to wait for service in a subsequent time

iew
approach uses the expected revenue data to build a metamodel that interval. This willingness is expressed as a recapture rate. while the
approximates expected revenue as a function of the first-stage system remaining demand will be lost. We use piecewise linear functions for the
design. Finally, we employ this expected revenue approximation in the nonlinearity in the constraints associated with the recaptured demand.
profit objective of the first stage to enable a more computationally efficient Moreover, we assume that the amount of electricity purchased from the grid
and interpretable method to optimize the system design. The results that we at a station in a time period is no larger than the nominal demand.
obtain from the DACE based system design optimization approach, when The subsequent MILP formulation is designed to optimize the selection
compared to solutions of the MILP model from a commercial solver, of charging stations and the allocation of slots while simultaneously
provide near optimal solution with a loss of less than 1% of profit. maximizing total profits. While providing an optimal solution, it's important
Furthermore, the DACE approach reduces the computational time from to note that the proposed MILP formulation's drawback lies in its demand
several hours to roughly 18 minutes, making it a much more suitable option for extensive computational time when utilizing a commercial solver. Hence,

ev
for practical use. In addition, the DACE approach yields highly we develop the DACE based optimization approach to determine the system
interpretable profit functions allowing us to analyze the marginal profits as of EV charging stations.
a function of the number of slots opened at each station. In this paper, the following notations are introduced for the MILP
We organize the rest of this paper as follows: In Section 2, we present formulation and the DACE based system design optimization problem
the system design problem formulation, including system design modeling formulation.
assumptions, formulation of the problem using MILP and the DACE

r
approach in detail. In Section 3, we describe system design experiments Sets
where we discuss the results. Finally, we present conclusions in Section 4. 𝐽 set of potential station locations indexed by j
𝐼 set of demand hot spots indexed by i
2. System Design Problem Formulation
2.1 System Design Layout
er 𝐼(𝑗)

𝑇
𝐾
set of demand hot spots within the max-mile radius
of station j
set of time periods in a day indexed by t
set of basis functions in a multivariate adaptive
regression splines (MARS) function indexed by k
pe
𝑁 set of LH experimental design points

Parameters
𝑚𝑖𝑗 Distance from hotspot i to station j (in miles)
𝜑 Maximum mile radius between the hotspots and the
station
Fig. 1. EV Charging station Layout 𝑝𝑖 Population of EVs at hotspot i
The EV charging station depicted in Figure 1 is devised to harness 𝑑𝑡 Demand percentage in time period t
ot

energy from both wind and solar sources, as well as the conventional power 𝑒 Charging efficiency of the battery
grid. The stored energy within the charging station is subsequently utilized 𝑑𝑐 Discharge rate of the battery
for recharging electric vehicles (EVs). The system has the capability to 𝜙 Recapture rate
store excess energy for future requirements through a battery storage unit. 𝜃 First time period in the time horizon
The surplus stored energy can serve as a reserve for meeting demand in 𝜌 Last time period in the time horizon
tn

case the generated energy falls short. Should the energy produced by 𝑟𝑡
wind/solar sources and the battery storage prove inadequate to meet the Electricity retail price in time period t
𝑐𝑟 Battery charge capacity
demand, the system will procure the necessary energy from the power grid.
𝑐𝑗 Operational cost of station j
Additionally, any surplus energy stored in the unit can be sold back to the
𝑁𝑐𝑗 Operational cost of a slot at station j
power grid, contributing to increased profits.
𝑣 Minimum battery level
𝑢 Maximum battery level
rin

ℶ Maximum number of slots opened per charging

station
𝑠𝑐 Slot capacity
𝑊𝑡 Wind generation (Mwh) in time period t
𝑆𝑡 Solar production (Mwh) in time period t
𝑀𝑡 Market price in time period t
𝛽0 Y-intercept of the MARS function
ep

𝛽𝑘 Least squares estimators for basis function k

𝐿𝐻𝑗 Traditional LH design value between 0 and 1 of
station j

Fig. 2 . The distribution of the station locations Variables

Figure 2 shows the distribution of 11 potential station locations within 𝑥𝑗 ∈ {0,1} Binary variable, if station j is operational
Pr

the Dallas/Fort Worth area in Texas. 140 hotspots (cities) are spread around 𝑦𝑖𝑗 ∈ {0,1} Binary variable, if hotspot i is assigned to station j
these stations. The distance between each hotspot and station is pre- 𝛼𝑡𝑗 ∈ {0,1} Binary variable, if the solar production in time
determined. period t is allocated to station j
𝜔𝑡𝑗 Fraction of the total wind generation in time period t
2.2 System Design Modeling Assumptions allocated to station j
The model assumes that decisions are optimized over a fixed 𝑔+𝑡𝑗 Electricity bought from the power grid by station j in
reoccurring time horizon, such as a day. In addition, there is a fixed cost, time period t
which we refer to as an operational cost, to open a charging station and 𝑔―𝑡𝑗 Electricity sold to the power grid from direct charge

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4572175
 of station j in time period t station j is equal to the total capacity of the charging station j; otherwise, it
𝐵―
𝑡𝑗 Electricity sold to the power grid from the battery is equal to the sum of the total demand and recaptured demand in time period
at t at station j.
station j in time period t When the demand for a time period exceeds the total capacity, it is
𝐷𝑡𝑗 Total demand in time period t at charging station j assumed that a proportion 𝜙 of customers are amenable to waiting for service
𝐷1𝑡𝑗

ed
Demand satisfied by direct charge of station j in in a subsequent time period. Specifically, ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇\{𝜃}, let 𝑏𝑡𝑗 be such that
time period t
𝐷2𝑡𝑗 if 𝜙[𝐷(𝑡―1)𝑗 + 𝑅(𝑡―1)𝑗 ― 𝑇𝑐𝑗] ≥ 0
Demand satisfied by the battery at station j in time
period t
1
𝑏𝑡𝑗 = 0 { o.w.
𝐿𝑡𝑗 Battery level of station j in time period t
𝐵𝑐𝑡𝑗 ∀𝑗 𝜖 𝐽 let 𝑏𝜃𝑗 be such that
Battery charge of station j in time period t
𝑁𝑠𝑗 Number of operational slots at station j if 𝜙[𝐷𝜌𝑗 + 𝑅𝜌𝑗 ― 𝑇𝑐𝑗] ≥ 0

iew
𝑇𝑐𝑗
𝑁𝑑𝑡𝑗
Total capacity of slots at station j
1
𝑏𝜃𝑗 = 0 { o.w.
Nominal demand in time period t at station j
𝑅𝑡𝑗 Recaptured demand from time period t – 1 at ԑ (1-𝑏𝑡𝑗) + 𝜙[𝐷(𝑡―1)𝑗 + 𝑅(𝑡―1)𝑗 ― 𝑇𝑐𝑗] ≥ 0; ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇\{𝜃} (10)
charging station j
𝑎𝑡𝑗, 𝑏𝑡𝑗 Binary decision variables used for the piecewise ԑ (1-𝑏𝜃𝑗) + 𝜙[𝐷𝜌𝑗 + 𝑅𝜌𝑗 ― 𝑇𝑐𝑗] ≥ 0; ∀𝑗 𝜖 𝐽 (11)
linear formulation in time period t at station j
𝑍𝑀𝐼𝐿𝑃 Objective function of the MILP formulation 𝜙[𝐷(𝑡―1)𝑗+𝑅(𝑡―1)𝑗 ― 𝑇𝑐𝑗] ≤ ԑ 𝑏𝑡𝑗; ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇\{𝜃} (12)
(𝑥, 𝑁𝑠 ) System station design
𝑍𝑀𝐼𝐿𝑃(𝑥, 𝑁𝑠 ) Objective function of a system station design 𝜙[𝐷𝜌𝑗+𝑅𝜌𝑗 ― 𝑇𝑐𝑗] ≤ ԑ 𝑏𝜃𝑗;∀𝑗 𝜖 𝐽 (13)

ev
𝑅𝑒𝑣 (𝑁𝑠 ) Revenue of a system station design
𝐵𝐹𝑘 The value of basis function k in the MARS model - ԑ (1-𝑏𝑡𝑗) +𝜙[𝐷(𝑡―1)𝑗 + 𝑅(𝑡―1)𝑗 ― 𝑇𝑐𝑗] ≤ 𝑅𝑡𝑗 ≤ 𝜙[𝐷(𝑡―1)𝑗 + 𝑅(𝑡―1)𝑗 ― 𝑇𝑐𝑗] +
ԑ (1-𝑏𝑡𝑗); ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇 (14)
2.3 MILP Formulation
The objective function in the MILP is given as equation (1).
-ԑ (1-𝑏ϴ𝑗) +𝜙[𝐷𝜌𝑗 + 𝑅𝜌𝑗 ― 𝑇𝑐𝑗 ] ≤ 𝑅ϴ𝑗 ≤ 𝜙[𝐷𝜌𝑗 + 𝑅𝜌𝑗 ― 𝑇𝑐𝑗] + ԑ (1-𝑏ϴ𝑗); ∀𝑗 𝜖 𝐽 (15)

r
max ∑𝑡∈𝑇 ∑𝑗∈𝐽 [(𝑀𝑡(𝑔―
𝑡𝑗 + 𝐵𝑡𝑗 ― 𝑔𝑡𝑗) + 𝑟𝑡𝑁𝑑𝑡𝑗)] ― ∑𝑗∈𝐽(𝑐𝑗𝑥𝑗 +𝑁𝑐𝑗𝑁𝑠𝑗)
― +
(1) -ԑ 𝑏𝑡𝑗 ≤ 𝑅𝑡𝑗 ≤ ԑ 𝑏𝑡𝑗; ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇 (16)

As depicted in equation (1), the primary aim is to optimize the revenue The piecewise constraints in equations (10) - (16) ensure that if the total
generated by selling energy to the power grid. This encompasses both
energies derived from the battery and direct charging operations across all
stations, coupled with the revenue stemming from satisfying the demand.
These gains are then offset by expenses related to energy procurement from
the power grid, the operational cost of an EV station at a prospective site,
and the expenses incurred by opening slots.
er
demand and recaptured demand from time period t-1 at station j is greater
than or equal to the total capacity of the station j, then the recaptured in time
period t at station j is 𝜙 of the demand and recaptured demand in time period
t-1 minus the total capacity of the charging station j. Otherwise, there will
be no recaptured demand. We assume that the recaptured demand at the last
time period is taken into consideration to calculate the recaptured demand of
pe
the demand at the first time period. Observe that ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇\{𝜃}, 𝑎𝑡𝑗 = 𝑏𝑡―1𝑗, and
The constraints of the MILP are as given below in equations (2)-(34). 𝑎ϴ𝑗 = 𝑏𝜌𝑗 by definition, so we can reduce the number of binary variables in the model.

[(
∑𝑖𝜖𝐼(𝑗) 𝑝𝑖 ∗ 𝑑𝑡 𝜑 ― 𝑚𝑖𝑗)
𝜑
]
𝑦𝑖𝑗 = 𝐷𝑡𝑗; ∀𝑗 𝜖 𝐽, ∀𝑡𝜖𝑇 (2) 𝑁𝑠𝑗 = 𝑇𝑐𝑗
; ∀𝑗 ∈ 𝐽 (17)
𝑠𝑐

The constraints in equation (2) serve to guarantee that the total demand in The constraints in equation (17) guarantee that the count of opened slots at
time period t at charging station j is the product of the distance function, charging station j is determined by dividing the total capacity of charging
which is the percentage of the demand of hotspot i assigned to station j with station j by the capacity of each individual slot.
ot

the population of hotspot i and the general demand percentage in time

period t. 𝑁𝑑𝑡𝑗 = (𝐷1𝑡𝑗 + 𝐷2𝑡𝑗); ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇 (18)

𝑦𝑖𝑗 ≤ 𝑥𝑗; ∀𝑖 ∈ 𝐼(𝑗), ∀𝑗 ∈ 𝐽 (3)

The constraints in equation (18) ensure that the total nominal demand in time
tn

period t at charging station j is equal to the demand satisfied by the direct

∑𝑗∈𝐽 𝑦𝑖𝑗 ≤ 1; ∀𝑖 ∈ 𝐼 (4) charge of station j in time period t and the demand satisfied by the battery at
station j in time period t. ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇, let 𝛼𝑡𝑗 be such that
The constraints in equation (3) ensure that demand hotspot i can be
assigned to stations j only if stations j is opened. The constraints in equation
(4) ensure that each hotspot i is served by at most one station. 𝛼𝑡𝑗 = 0 o.w {
1 if solar production 𝑡 is allocated to station 𝑗
rin

𝑥𝑗 + 𝑦𝑖𝑗 ≤ 1; ∀𝑖 ∈ 𝐼, ∀𝑗, 𝑗 ∈ 𝐽,𝑚𝑖𝑗 < 𝑚𝑖𝑗 (5) 𝐿𝑡𝑗 = 𝐿(𝑡―1)𝑗 + 𝐵𝑐𝑡𝑗 ―𝑒𝐷2𝑡𝑗 +𝑒𝐵―
𝑡𝑗; ∀𝑗 ∈ 𝐽, ∀𝑡 ∈ 𝑇 (19)

The constraints in equation (5) ensure that if j is a closer station than ĵ, then 𝐿𝜃𝑗 = 𝐿𝜌𝑗 + 𝐵𝑐𝜃𝑗 ―𝑒𝐷2ϴ𝑗 +𝑒𝐵―
𝜃𝑗; ∀𝑗 ∈ 𝐽 (20)
the hotspot will be assigned to the closer one.
For the piecewise linear function for the recaptured demand, ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇, 𝐵𝑐𝑡𝑗 = 𝑊𝑡𝜔𝑡𝑗 + 𝑆𝑡𝛼𝑡𝑗 + 𝑔+ ― 1
𝑡𝑗 ― 𝑔𝑡𝑗 ― 𝐷𝑡𝑗; ∀𝑗 ∈ 𝐽, ∀𝑡 ∈ 𝑇 (21)
let ε be an upper bound (very large number, + ∞) and 𝑎𝑡𝑗 be such that
The set of energy balance constraints include the battery level transition as
ep

1
𝑎𝑡𝑗 = 0{ o.w.
if 𝐷𝑡𝑗 + 𝑅𝑡𝑗 ≥ 𝑇𝑐𝑗 equation (19), the energy balance for the battery charge as equation (21).
Moreover, the constraints in equation (20) ensure the battery level at the first
stage is calculated using battery level transition equation and the battery
ԑ (1-𝑎𝑡𝑗) + 𝐷𝑡𝑗 + 𝑅𝑡𝑗 ≥ 𝑇𝑐𝑗 (6) level at the last stage.
𝐷𝑡𝑗+𝑅𝑡𝑗 ≤ 𝑇𝑐𝑗 + ԑ 𝑎𝑡𝑗 (7) 𝑔― +
𝑡𝑗 ≤ 𝑊𝑡𝜔𝑡𝑗 + 𝑆𝑡𝛼𝑡𝑗 + 𝑔𝑡𝑗; ∀𝑗 ∈ 𝐽, ∀𝑡 ∈ 𝑇 (22)
Pr

-ԑ (1-𝑎𝑡𝑗) +𝑇𝑐𝑗 ≤ 𝑁𝑑𝑡𝑗 ≤ 𝑇𝑐𝑗 + ԑ (1-𝑎𝑡𝑗); ∀𝑗 𝜖 𝐽, ∀𝑡𝜖𝑇 (8) 𝑔+ ≤ 𝑁𝑑𝑡𝑗; ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇

𝑡𝑗 (23)
-ԑ 𝑎𝑡𝑗 + 𝐷𝑡𝑗 + 𝑅𝑡𝑗 ≤ 𝑁𝑑𝑡𝑗 ≤ 𝐷𝑡𝑗+𝑅𝑡𝑗 + ԑ 𝑎𝑡𝑗; ∀𝑗 𝜖 𝐽, ∀𝑡𝜖𝑇 (9)
The constraints in equation (22) ensure that the electricity sold to the power
The constraints in equations (6) - (9) represent a piecewise linear grid from the direct charge of station j in time period t should be less than or
formulation ensuring that if the sum of the total demand and the recaptured equal to the sum of the total wind purchased by station j in time period t, the
demand in time period t at station j is greater than or equal to the total solar production of station j in time period t, and the electricity bought from
capacity of the station j, then total nominal demand in time period t at the power grid by station j in time period t. However, because there is no
reason to purchase and sell back to the power grid from the same station in

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4572175
 +
the same time period, the purchase term 𝑔𝑡𝑗 in (22) can be omitted. in step 1 and the corresponding revenues generated by step 2.
Similarly, the constraints in equation (23) ensure that the electricity bought 4. The obtained MARS model in step 3 is then optimized by subtracting
from the power grid by station j in time period t should be no more than the the cost component to predict the profit to obtain a best system design.
total nominal demand in time period t at charging station j. 5. True profit is calculated using the system design point from step 4, with
𝐵― 2
𝑡𝑗+𝐷𝑡𝑗 ≤ 𝑑𝑐 ∗ 𝑒 ∗ 𝑥𝑗; ∀𝑗 ∈ 𝐽, ∀𝑡 ∈ 𝑇 (24) the help of the MILP. The obtained result is the solution to our first-
stage system design problem.

ed
𝐵𝑐𝑡𝑗 ≤ 𝑐𝑟 ∗ 𝑥𝑗; ∀𝑗 ∈ 𝐽, ∀𝑡 ∈ 𝑇 (25)
2.4.1 Binned LH Design
𝑣 ∗ 𝑥𝑗 ≤ 𝐿𝑡𝑗 ≤ 𝑢 ∗ 𝑥𝑗; ∀𝑗 ∈ 𝐽, ∀𝑡 ∈ 𝑇 (26) As described in the MILP formulation, the system design variables are
given by the vectors (𝑥, 𝑁𝑠 ), where vector 𝑥 includes binary variables,
The constraints in equation (24) ensure that the sum of the electricity sold indicating which stations are operational, and vector 𝑁𝑠 gives the number of
back to the power grid from the battery at station j in time period t and the open slots. A traditional LH experimental design yields independent values

iew
demand satisfied by the battery at station j in time period t cannot be higher between given limits. However, due to constraint set (30), the vectors 𝑥 and
than the product of discharge rate and storage efficiency of station j. 𝑁𝑠 are dependent in that slots can only be opened at operational stations.
Similarly, the constraints in equation (25) ensure that the battery charge of Consequently, we develop a binned LH, which consists of a traditional LH
station j in time period t should be within the battery charging capacity. experimental design along with a mapping that determines a feasible system
The constraints in equation (26) ensure that the battery inventory is
design (𝑥, 𝑁𝑠 ) from a traditional LH experimental design point. Moreover,
between minimum and maximum battery level for each station. Moreover,
traditional LH experimental designs fill the space, but optimal solutions tend
(24) - (26) are only considered when station j is operational.
to reside in very small portions of the feasible decision space, so the mapping
needs to be intelligent to include “good” system design points so that the
∑𝑗𝜖𝐽 𝜔𝑡𝑗 ≤ 1; ∀𝑡 ∈ 𝑇 (27)
MARS model can accurately model this part of the space. For example,

ev
opening 8 or more stations usually yields more capacity than needed to meet
The constraint in equation (27) ensures that the fraction of the allocated demand, whereas opening fewer than 3 stations will be insufficient to fulfill
wind generation to all the stations is no more than 1. demand. Consequently, we need an experimental design where many of the
system design points have between 4 to 7 operational stations. Specifically,
𝜔𝑡𝑗 ≤ 𝑥𝑗; ∀𝑗 ∈ 𝐽, ∀𝑡 ∈ 𝑇 (28) we used a traditional LH experimental design to generate 325 11-
dimentional experimental design points between 0 and 1 using the

r
𝛼𝑡𝑗 ≤ 𝑥𝑗; ∀𝑗 ∈ 𝐽, ∀𝑡 ∈ 𝑇 (29) MATLAB R2016a function “lhsdesign”. The fractional values obtained
from the first 20 points of the traditional LH experimental design are as
Constraints in equations (28) and (29) ensure that the total wind purchased indicated in Table 1 below. The mapping is shown as the step function given
by station j in time period t, and the solar production of station j in time in Figure 3, which can be defined as follows for a station 𝑗:
period t are only considered if the station is operational.

0 ≤ Nsj ≤ ℶ 𝑥𝑗; ∀𝑗 ∈ 𝐽

The constraints in equation (30) ensure that the number of slots opened is
(30)
er { (0,0)
(𝑥𝑗,𝑁𝑠𝑗) = (1,⌈19(𝐿𝐻 ― 9 )⌉)
𝑗
19
𝑖𝑓 𝐿𝐻𝑗 <
𝑜.𝑤.
9
19
pe
between zero and the maximum possible number of slots opened per Here 𝐿𝐻𝑗 refers to the traditional LH design value between 0 and 1. The
charging station j. Moreover, no slots are open if the station is not mapping translates the traditional LH experimental design in Table 1 into
operational. the binned LH experimental design in Table 2, which indicates the locations
of the charging stations and the number of slots at each charging station (𝑥,
𝑁𝑑𝑡𝑗, 𝐷𝑡𝑗, 𝐷1𝑡𝑗, 𝐷2𝑡𝑗, 𝜔𝑡𝑗, 𝑇𝑐𝑗, 𝐿𝑡𝑗, 𝑔+ ― ―
𝑡𝑗, 𝑔𝑡𝑗, 𝐵𝑡𝑗, 𝐵𝑐𝑡𝑗, 𝑅𝑡𝑗 ≥ 0; ∀𝑗 𝜖 𝐽, ∀𝑡 𝜖 𝑇 (31) 𝑁𝑠 ). Figure 4 also shows the distribution of the number of operational
|𝐽|
𝑥 ∈ 𝔹 (32) stations in the binned LH experimental design. Observe that 80% of the
system design points have between 4 and 7 operational stations, which will
𝑦 ∈ 𝔹|𝐼|∗|𝐽| (33) appropriately capture demand without too much excess capacity. The 325
binned LH experimental design points were split into 250 training data
ot

𝛼,𝑎, 𝑏 ∈ 𝔹|𝑇|∗|𝐽| (34) points and 75 testing data points.

Table 1
The constraints in equations (31) - (34) ensure that the given variables are 20 points LH Design using MATLAB (Partial)
LH1 LH2 LH3 LH4 LH5 LH6 LH7 LH8 LH9 LH10 LH11
nonnegative, and 𝑥,𝑦,𝛼,𝑎 and 𝑏 are binaries of appropriate dimension.
0.74 0.28 0.45 0.17 0.45 0.31 0.78 0.64 0.23 0.51 0.50
tn

Furthermore, the constraints in equations (4), (5), and (27) create a 0.84 0.02 0.66 0.70 0.32 0.62 0.53 0.79 0.76 0.52 0.61
dependent relationship between the stations and prevent the problem from 0.24 0.09 0.19 0.46 0.10 0.29 0.42 0.05 0.89 0.68 0.00
being separable by station. 0.51 0.76 0.53 0.63 0.11 0.82 0.07 0.85 0.85 0.91 0.58
0.46 0.67 0.63 0.96 0.86 0.82 0.24 0.42 0.63 0.94 0.76
2.4 DACE Based System Design Optimization Approach 0.33 0.74 0.03 0.32 0.54 0.99 0.59 0.43 0.72 0.88 0.67
0.83 0.58 0.47 0.07 0.96 0.47 0.44 0.34 0.93 0.58 0.52
DACE stands as a statistical technique designed to efficiently carry 0.99 0.78 0.89 0.79 0.99 0.49 0.90 0.57 0.46 0.90 0.83
out computer experiments, particularly suited for exploring applications 0.59 0.04 0.14 0.38 0.41 0.61 0.06 0.31 0.30 0.12 0.97
rin

governed by intricate computer models (Sacks et al. 1989). These types of 0.18 0.37 0.44 0.56 0.36 0.86 0.48 0.40 0.69 0.34 0.30
computer models find widespread use in engineering domains, often seen 0.12 0.22 0.67 0.19 0.22 0.04 0.57 0.36 0.30 0.03 0.79
in applications like finite element simulations (Furushima & Manabe 0.57 0.44 0.32 0.08 0.46 0.20 0.15 0.68 0.68 0.50 0.26
0.31 0.84 0.96 0.14 0.27 0.97 0.58 0.71 0.06 0.72 0.03
2011). Chen et al. (2006) provided a review of DACE methods, including 0.25 0.26 0.70 0.84 0.48 0.98 0.87 0.99 0.99 0.19 0.10
the adaptation to DACE based optimization. In the conventional approach 0.54 0.56 0.23 0.31 0.44 0.48 0.22 0.95 0.20 0.42 0.32
of DACE, an experimental design is employed to structure a series of 0.18 0.02 0.56 0.40 0.26 0.84 0.17 0.37 0.66 0.09 0.29
computer experimental runs. This arrangement facilitates the creation of a 0.66 0.15 0.88 0.62 0.10 0.96 0.07 0.16 0.37 0.30 0.03
ep

statistical "metamodel" that serves as an approximation for the 0.06 0.80 0.57 0.82 0.42 0.70 0.71 0.94 0.74 0.80 0.17
0.03 0.18 0.77 0.74 0.52 0.40 0.02 0.30 0.02 0.49 0.44
performance output simulated by an intricate computer model. The 0.27 0.79 0.90 0.38 0.33 0.44 0.30 0.57 0.10 0.78 0.75
metamodel is a mathematical surrogate that can be employed to study the
simulated system more efficiently. In DACE based optimization, the
computer model is an optimization algorithm instead of the traditional
computer simulation model. Specifically, in this paper, we develop the
following DACE based optimization approach using the steps below.
Pr

1. Using an experimental design (binned LH), a set of sample points,

each representing a system design in the design parameter space, is
generated.
2. The performance (revenue) of each system design point is then
determined by fixing the system design variables and solving the
MILP to obtain the solution to a control sub problem.
3. A multivariate adaptive regression splines (MARS) model is fit to the
experimental design obtained by the binned LH experimental design

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4572175
 Fig. 5 . Computing revenues solving control problem
As depicted in Figure 5, each row serves as input for the control sub-problem
formulation. This problem is addressed using the MILP presented in equation (35),
resulting in the derivation of the respective revenues denoted as 𝑅𝑒𝑣(𝑁𝑠).

2.4.3 MARS Model

ed
To forecast the revenue for the EV charging station system, the second-stage
model employs calibration of a MARS statistical model (Friedman 1991, Chen
1999). This specific model demonstrates notable proficiency in predicting revenue
by considering the count of available charging slots across diverse locations. MARS
is adapted to the experimental design extracted through binned LH, along with the
correlated revenues resulting from solving the second-stage control problem (35).

iew
The resultant fitted model, as shown in equation (36), serves as a predictor for
revenue.

𝑅𝑒𝑣(𝑁𝑠) = 𝛽0 + ∑𝐾 ) (36)
Fig. 3 . Step Function 𝑘=1 𝛽𝑘𝐵𝐹𝑘(𝑁𝑠

Table 2
20 points Binned LH Design (Partial) In our paper, we fit two different MARS models, one with basis interaction
Ns1 Ns2 Ns3 Ns4 Ns5 Ns6 Ns7 Ns8 Ns9 Ns10 Ns11 terms and the other with no interaction.
5 0 0 0 0 0 6 3 0 1 0
7 0 4 5 0 3 1 6 6 1 3 2.4.4 First-Stage EV System Master Problem

ev
0 0 0 0 0 0 0 0 8 4 0 ∗ ∗
1 6 1 3 0 7 0 8 8 9 2
The system of charging stations (𝑥 , 𝑁𝑠 ) is obtained by maximizing
0 4 3 10 8 7 0 0 3 9 6 profit using the optimization problem given by (37),
0 5 0 0 1 10 2 0 5 8 4
7 2 0 0 10 0 0 0 9 2 1 𝑍𝑀𝐼𝐿𝑃(𝑥∗, 𝑁𝑠∗) = max 𝑅𝑒𝑣(𝑁𝑠) - 𝑐𝑗𝑥𝑗 ― 𝑁𝑐𝑗𝑁𝑠𝑗 (37)
10 6 8 6 10 0 9 2 0 8 7
2 0 0 0 0 3 0 0 0 0 10
0 0 0 2 0 8 0 0 4 0 0
s.t. Equations (30) and (32)
∗ ∗ ∗ ∗

r
0 0 4 0 0 0 2 0 0 0 6 To evaluate the quality of (𝑥 , 𝑁𝑠 ), 𝑍𝑀𝐼𝐿𝑃(𝑥 , 𝑁𝑠 ) is then calculated using
2 0 0 0 0 0 0 4 4 1 0 equation (38).
0 7 10 0 0 10 2 5 0 5 0
0 0 4 7 0 10 8 10 10 0 0
𝑍𝑀𝐼𝐿𝑃(𝑥∗, 𝑁𝑠∗) = 𝑚𝑎𝑥 ∑𝑡∈𝑇 ∑𝑗∈𝐽 [(𝑀𝑡(𝑔―
𝑡𝑗 + 𝐵𝑡𝑗 ― 𝑔𝑡𝑗) + 𝑟𝑡𝑁𝑑𝑡𝑗)] ― ∑𝑗∈𝐽(𝑐𝑗𝑥𝑗 +𝑁𝑐𝑗𝑁𝑠𝑗)
― +
1
0
4
0
0
0
2
0
0
7
0
6
0
2
8
2
6
8
0
0
3
7
5
0
0
0
0
0
1
0
0
7
10
5
0
0
0
0
0
5
0
0
10
0
0
9
0
2
0
4
0
5
0
0
0
0
0
7
0
6
0
0
0
0
0
5
er ∗ ∗

DACE approach, respectively.

∗
∗
s.t. Equations (2) – (34), and 𝑥𝑗 = 𝑥 , 𝑁𝑠𝑗 = 𝑁𝑠

∗
∗

The obtained 𝑍𝑀𝐼𝐿𝑃(𝑥 , 𝑁𝑠 ) and (𝑥 , 𝑁𝑠 ) are the profit and solution of the
(38)
pe
3. System Design Experiments
3.1 MILP Experiments
For experimental purposes, all simulations were conducted on a
workstation equipped with an Intel Core i7 CPU @2.60 GHz, featuring 2
physical cores, 4 logical processors, and 8 GB RAM. The optimization
procedures were executed using CPLEX 12.6.3 and GUROBI 9.5. The data
utilized for this investigation encompasses wind generation (𝑊𝑡) (National
ot

Renewable Energy Laboratory 2012), solar generation (𝑆𝑡) (Miller & Lumby
2012), market price (𝑀𝑡) (Electric Reliability Council of Texas 2002), and
demand profiles (𝑑𝑡) (Khosrojerdi et al. 2012). The fixed retail price (rt) of
electricity is 10.24 cents per kilowatt-hour per time periods (U.S.
Department of Energy 2012). The maximum (v) and minimum (u) battery
tn

Fig. 4 . Percentage of Stations opened levels are 3.6 MWh and 720 kWh per slot. The charging rate (cr) and
discharging rates (dc) are 600 kW and 75 kW per slot. The capacity of every
2.4.2 Second-Stage Control Problem slot (sc) is 18.75 kWh. In our study, the storage efficiency (e) is assumed to
For each system of charging stations (𝑥, 𝑁𝑠 ) in the experimental be 79.8% (Wetz, 2010). For convenience, we assumed that a station (j) more
𝜑)
design, the corresponding second-stage control problem revenue 𝑅𝑒𝑣(𝑁𝑠) than a 20-mile ( radius from a hotspot (i) is unable to fulfill the demand
is determined using MILP as shown in equation (35). of that hotspot. The cost to open a slot (Ncj) at a station (j) is assumed to be
10% of the operational cost of the station (cj). The maximum allowable
rin

𝑅𝑒𝑣(𝑁𝑠) = 𝑚𝑎𝑥 ∑𝑡∈𝑇 ∑𝑗∈𝐽 [(𝑀𝑡(𝑔― 𝑡𝑗 + 𝐵𝑡𝑗 ― 𝑔𝑡𝑗) + 𝑟𝑡𝑁𝑑𝑡𝑗)]

― +
(35)
number of slots (ℶ) that can be opened is 10. Given market price fluctuations
every 15 minutes (t), our formulation encompasses a daily control problem
𝑁𝑠 = 𝑁𝑠. spanning 96 15-minute periods. In instances where the demand for a given
s.t. Equations (2) – (34) and
time period surpasses capacity, it is assumed that 50% (recapture rate (𝜙))
of the customers are willing to wait for service in a subsequent time slot,
while the remaining customers are lost.
Table 3 below provides a depiction of the optimal operational station
ep

configuration, the count of open slots, and the individual profits generated
by each station under varying operating cost scenarios. Additionally, the
table includes the computational time if the simulation is completed within
a 6-hour time limit.
Table 3
Number of Slots Per Charging Stations vs. Different Cost Scenarios
Pr

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4572175
 ed
iew
Fig. 7. Major stations: different cost vs. number of slots
We employed CPLEX to address the MILP for a range of operating In Figure 7, the distribution of opened slots across various cost scenarios is
cost values. It has been noted that the best integer solution typically presented. The illustration highlights that at an operational cost (c) of $0, Fort Worth
materializes within a span of 20 minutes to 1 hour and 55 minutes, although and Dallas exhibit a substantial slot count of 6 and 5, respectively, while Denton and
verifying its optimality necessitates approximately 5 additional days of Garland closely follow with slot counts of 4 and 3, respectively. Interestingly, Fort
computational effort. Consequently, a maximum runtime of 6 hours is Worth and Dallas maintain their high slot counts despite cost escalation, whereas

ev
applied to all scenarios except the one involving an operational cost (c) of Garland and Denton experience a reduction in slots as costs rise. This visual
$100 per day. This specific cost value is selected as the baseline, as it representation mirrors real-world dynamics, where, beyond a certain threshold, the
mirrors a probable operating expense for a charging station compared to feasibility of opening higher-cost slots diminishes due to insufficient demand. In
the other scenarios. The foundational station cost of $0 per day is observed essence, when the cost reaches $200 per day, the generated profits from Garland and
to render all stations operational, yielding a higher count of opened slots Denton no longer cover their costs, resulting in their closure.
and overall profit relative to the other scenarios. As operational costs

r
escalate, more stations remain closed, and the quantity of slots per station 3.2 DACE Based System Design Optimization Experiments
diminishes correspondingly. However, during the transition from a cost of Table 4
$70 to $100, the number of slots (Ns) in Garland increases from 2 to 3. This Comparisons of the DACE MILP objective solutions
alteration occurs because Garland absorbs the demand previously served 𝑍𝑀𝐼𝐿𝑃(𝑥∗, 𝑁𝑠∗)
by Rockwall (which remains closed at an operational cost of $100 per day).
Based upon our demand assumptions, there is no need to activate 8 or
more stations to meet demand. Conversely, opening fewer than 3 stations
would be inadequate to fulfill demand. As depicted in Table 3, operational
costs of $60, $70, and $100 yield 5, 5, and 4 opened stations, respectively.
For the $100 per day cost scenario, it took CPLEX approximately 4 days
er Software for
MARS Design

ARESLAB
ARESLAB
SPM
SPM
Interaction
allowed or not

Yes
No
Yes
No
Testing
R2

97.0
97.7
97.2
98.7
CPLEX CP
/Couenne (% Diff)
2406.8 (3.1)
2670.0 (0.7)
2414.0 (2.8)
2678.5 (0.4)
MINOS (% Diff)

2566.8 (4.6)
2674.6 (4.3)
2555.8 (5.0)
2629.5 (2.2)
pe
In all our experiments utilizing the DACE based system design
and 23 hours of processing time to attain an optimal solution. The optimal optimization, a baseline operational cost (c) of $100 per station per day is
solution's objective value stands at $2689.38 (𝑍𝑀𝐼𝐿𝑃), with operational upheld. We generate four MARS revenue models utilizing two software
stations encompassing Fort Worth, Dallas, Garland, and Denton. Their tools: MATLAB 8.6 with the ARESLAB toolbox (Jekabsons 2016) and
respective daily profits are $2072.21, $293.96, $201.76, and $121.45. Salford Predictive Modeler (SPM) 8.0 (Minitab 2016). Each software
Furthermore, the count of opened slots at each of these locations is 5, 4, 3, package calibrates one MARS model with basis interaction terms and
and 1, respectively. another with no interactions. For the SPM non-interaction MARS model, we
In addition to solving the baseline scenario (operational cost of $100) set the maximum basis functions to 100 and consider the minimum
with CPLEX, we also solve it with GUROBI with the lp file generated from observation between knots as 1. Subsequently, the optimized MARS models
ot

CPLEX OPL. Given that the generation of lp file is negligible, GUROBI are produced using three distinct software programs: CPLEX CP Optimizer
took 2 hours to provide the same optimal solution as CPLEX. (CPLEX 2015), AMPL 11.2 (Fourer et al. 1990) with the Couenne solver
(Belotti et al. 2009), and MINOS solver (Murtagh et al. 2006). The charging
station systems derived from CPLEX CP and Couenne are identical, thereby
generating 8 distinct systems through the DACE approach, as depicted in
tn

Table 4. The coefficient of determination (R-squared) for each MARS model

is computed, as displayed in the table. Additionally, the percentage
∗ ∗
difference (% Diff) between the objective solutions 𝑍𝑀𝐼𝐿𝑃(𝑥 , 𝑁𝑠 ) and
𝑍𝑀𝐼𝐿𝑃 is enclosed in parentheses in the table. Initial analysis indicates that
the DACE approach utilizing the non-interaction MARS metamodel from
the SPM software exhibits superior performance, with a percentage
rin

∗ ∗
difference of 𝑍𝑀𝐼𝐿𝑃(𝑥 , 𝑁𝑠 ) of 0.4% and has the highest testing R-squared.
In this study, charging station systems generated through the DACE
approach with non-interaction metamodels exhibit slightly higher accuracy
than those with interaction terms. This implies that minimal demand shifts
occur due to the stations being widely dispersed. Consequently, the
distribution of demand (e.g., equations (4) and (5))) and the allocation of
wind power across stations (e.g., equation (27) exert limited influence on the
ep

Fig. 6 . Demand distribution at an operational cost of $100 solution in this specific case.
∗ ∗
Figure 6 portrays the temporal distribution of demand(𝐷𝑡𝑗) across The system design build (𝑥 , 𝑁𝑠 ) derived from our optimal model
stations from time periods 1 to 96. Notably, the total demand distribution (CPLEX CP/Couenne, MINOS) is further dissected and juxtaposed against
for Fort Worth takes the lead, Dallas follows as the second highest, trailed the MILP, as demonstrated in Table 5 below.
by Garland and Denton. In terms of daily consumption per station, the Table 5
values are 4.61 MWh, 3.89 MWh, 1.97 MWh, and 1.12 MWh, respectively. Number of Slots (MILP vs. DACE)
Pr

Number of slots per opened Stations (cost $100)

Furthermore, the demand is lowest between time periods 12 and 20,
Fort Worth Dallas Garland Denton Total
encompassing the early morning hours from 3 am to 5 am. Subsequently, MILP 5 4 3 1 13
a gradual upswing unfolds, culminating in peak demand between time CPLEX CP/Couenne 4 5 2 2 13
periods 52 to 68, spanning the afternoon from 1 pm to 5 pm. This MINOS 3 5 2 2 12
heightened demand is succeeded by a gradual decline. It is noticeable that each system design builds features identical
operational stations, positioned in Fort Worth, Dallas, Garland, and Denton.
Furthermore, the cumulative count of open slots is 13 for MILP, CPLEX CP,
and AMPL – Couenne approaches, while the AMPL – MINOS solution

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4572175
involves 12 open slots. An optimized model using mixed-integer linear programming (MILP) is
developed to determine optimal EV charging station locations, the quantity of slots
3.3 Interpretable Profit Functions to activate at each station, and the resulting overall profit. The result indicates that
Fort Worth should have the highest number of slots, trailed by Dallas, Garland, and
Denton. However, this method's drawback lies in its extended computational time,
taking several hours to complete, despite giving an optimal solution. To address this

ed
limitation, a two-stage framework and a system design optimization approach rooted
in DACE (Design and Analysis of Computer Experiments) are introduced for
solving the EV charging station network conundrum. In this investigation, the
DACE approach reveals that systems without interaction terms yield superior results
compared to those with interaction terms, implying minimal demand shifts due to
the considerable station separation. Furthermore, the DACE strategy generates

iew
highly interpretable profit functions, facilitating the analysis of marginal profits in
relation to the number of slots open at each station. These profit functions allow
decision makers to optimize profit under different cost scenarios without the use of
a commercial solver. Prominent basis functions are linked to Fort Worth, Dallas,
Garland, and Denton. Remarkably, the DACE approach streamlines the solution
process, requiring only about 18 minutes to achieve a solution within 1% of
optimality, in contrast to the several hours taken by the MILP approach.
In terms of future prospects, there is a plan to explore solving the problem with
stochastic input variables related to wind and solar power generation, as well as
market pricing fluctuations.

ev
Fig. 8. Profit Functions at Four charging Stations Declaration of Competing Interest
One of the major benefits of the DACE based system design optimization The authors affirm the absence of any identifiable conflicting
approach is that the MARS models with no interaction yield highly financial interests or personal affiliations that might have seemed to exert
interpretable profit functions. These functions allow decision makers to an impact on the findings presented in this manuscript.
analyze the marginal profits as a function of the number of slots opened at

r
each station. Graphs portraying the relationship between marginal profits Acknowledgements
and open slots are presented in Figure 8. Prominent basis functions are Partial funding for this study is provided through grants ECCS-
linked with Fort Worth, Dallas, Garland, and Denton. All other basis 1128871, ECCS 1128826, and ECCS-1938895 from the National Science
functions connected to different stations possess coefficients of zero in the Foundation. Our sincere appreciation goes to Amirhossein Khosrojerdi for
projected revenue function, indicating that, as per MARS analysis, these
stations hold negligible marginal profits. Consequently, the DACE
optimization phase maintains them in a closed state. Notably, it is evident
that the optimal number of slots for the Fort Worth station is 4, whereas for
Dallas, the optimal count is 5. Similarly, for Garland and Denton, the
er contributing the DFW EV demand profiles.

References

Ariyajunya, B., Chen, Y., Chen, V. C. P., Kim, S. B., & Rosenberger, J. (2021). Addressing state space
pe
∗ multicollinearity in solving an ozone pollution dynamic control problem. European Journal of
optimal count is 2 slots per station. These findings mirror the outcomes (𝑥 Operational Research, 289(2), 683–695. https://doi.org/10.1016/j.ejor.2020.07.014.
∗
, 𝑁𝑠 ) derived from CPLEX CP and AMPL – Couenne. Furthermore, the Arslan, O., & Karaşan, O. E. (2016). A Benders decomposition approach for the charging station
optimal profits achieved by Fort Worth, Dallas, Garland, and Denton are location problem with plug-in hybrid electric vehicles. Transportation Research Part B:
Methodological, 93, 670–695. https://doi.org/10.1016/j.trb.2016.09.001.
$904.55, $1543.95, $35.67, and $194.71 respectively, summing up to Awasthi, A., Venkitusamy, K., Padmanaban, S., Selvamuthukumaran, R., Blaabjerg, F., & Singh,
∗ ∗
$2678.88, mirroring the 𝑍𝑀𝐼𝐿𝑃(𝑥 , 𝑁𝑠 ) results obtained from the non- A. K. (2017). Optimal planning of electric vehicle charging station at the distribution system
interaction MARS model using SPM software, as presented in Table 4. using hybrid optimization algorithm. Energy, 133, 70–78.
Observe that by using these profit functions, we can construct an optimal https://doi.org/10.1016/j.energy.2017.05.094.
Battistelli, C., Baringo, L., & Conejo, A. (2012). Optimal energy management of small electric
solution without the use of a commercial optimization solver. energy systems including V2G facilities and renewable energy sources. Electric Power
ot

Systems Research, 92, 50–59. https://doi.org/10.1016/j.epsr.2012.06.002.

3.4 CPU Time Comparisons Belotti, P., Lee, J., Liberti, L., Margot, F., & Wächter, A. (2009). Branching and bounds
Table 6 tightening techniques for non-convex MINLP. Optimization Methods & Software, 24(4-5),
597-634. https://doi.org/10.1080/10556780903087124.
CPU Time Comparisons (MILP vs. DACE) Biesinger, B., Hu, B., Stubenschrott, M., Ritzinger, U., & Prandtstetter, M. (2017). Optimizing
Task Time charging station locations for electric car-sharing systems. Evolutionary Computation in
tn

Binned LHS Design 1 sec Combinatorial Optimization (157–172). http://dx.doi.org/10.1007/978-3-319-55453-2_11.

Revenue Function (250 Training and 75 Testing – 3.3 sec average) – GUROBI 18 min Brandstätter, G., Kahr, M., & Leitner, M. (2017). Determining optimal locations for charging
stations of electric car-sharing systems under stochastic demand. Transportation Research
Interaction No interaction Part B: Methodological, 104, 17–35. https://doi.org/10.1016/j.trb.2017.06.009.
Process Times Process Times Cervellera, C., Chen, V. C. P., & Wen, A. (2006). Optimization of a Large-Scale Water Reservoir
SPM 12 sec SPM 6 sec Network by Stochastic Dynamic Programming with Efficient State Space Discretization.
CPLEX CP 1 min 20 sec CPLEX CP 34 sec European Journal of Operational Research, 171(3), 1139–1151.
DACE - Total 19 min 32 sec DACE - Total 18 min 40 sec https://doi.org/10.1016/j.ejor.2005.01.022.
rin

Chang, S., Li, H., & Nahrstedt, K. (2014, December). Charging facility planning for electric
Original MILP (Gurobi) - 2 hours vehicles. 2014 IEEE International Electric Vehicle Conference (IEVC) (1-7). IEEE.
To further substantiate the utilization of the DACE based system design Chawal, U. (2018). Optimizing a System of Electric Vehicle Charging Stations (Doctoral
optimization approach, the process run times of the interaction and non-interaction dissertation). https://rc.library.uta.edu/uta-ir/handle/10106/27378. Accessed May 3, 2023.
Chen, C., & Hua, G. (2014). A New Model for Optimal Deployment of Electric Vehicle Charging
models are computed and tabulated in Table 6. After generating the binned LH and Battery Swapping Stations. International Journal of Control and Automation, 7(5), 247–
experimental design via MATLAB, the revenue data is collected employing 258.
GUROBI, culminating in a total runtime of 18 minutes. Once the revenue data is Chen, V. C. P. (1999). Application of Orthogonal Arrays and MARS to Inventory Forecasting
assembled, it is employed to construct the MARS model through SPM software. Stochastic Dynamic Programs. Computational Statistics and Data Analysis, 30(3), 317–341.
ep

https://doi.org/10.1016/s0167-9473(98)00084-x.
Ultimately, optimization of the MARS model is carried out utilizing CPLEX CP, Chen, V. C. P., Ruppert, D., & Shoemaker, C. A. (1999). Applying Experimental Design and
with the resultant outcomes provided. The swiftest comprehensive process time is Regression Splines to High-Dimensional Continuous-State Stochastic Dynamic Programming.
recorded as 18 minutes and 40 seconds, attributed to the non-interaction model. Operations Research, 47(1), 38–53. https://doi.org/10.1287/opre.47.1.38.
Due to the efficiency with which this MARS model discerns station revenues, the Chen, V. C. P. (2001). Measuring the Goodness of Orthogonal Array Discretizations for Stochastic
Programming and Stochastic Dynamic Programming. SIAM Journal on Optimization, 12(2), 322–
DACE approach can swiftly reoptimize across diverse cost scenarios, obviating 344. https://doi.org/10.1137/s1052623498332403.
the need for reacquiring responses from the binned LH design. Specifically, Chen, V. C. P., Günther, D., & Johnson, E. L. (2003). Solving for an Optimal Airline Yield
Pr

variations in operational cost (c) lead only to shifts in the profit functions depicted Management Policy via Statistical Learning. Journal of the Royal Statistical Society Series C:
in Figure 8, while alterations in the slot-opening cost (Nc) solely induce tilting, Applied Statistics, 52(1), 19–30. https://doi.org/10.1111/1467-9876.00386.
thus permitting optimization across distinct cost scenarios without resorting to a Chen, V. C., Tsui, K. L., Barton, R. R., & Meckesheimer, M. (2006, April). A review on design,
modeling and applications of computer experiments. IIE Transactions, 38(4), 273–291.
commercial optimization solver. In comparison to the original 2-hour https://doi.org/10.1080/07408170500232495.
computational requirement for the MILP, the DACE approach exhibits enhanced CPLEX, IBM ILOG. (2015). International Business Machines Corporation v12.6.3 (Version
computational feasibility. 12.6.3). http://www-01.ibm.com/software/commerce/optimization/cplex-optimizer/.
Electric Reliability Council of Texas. (2002). ERCOT History.
http://www.ercot.com/about/profile/history/. Accessed August 2, 2016.
4. Conclusion Fan, H., Tarun, P. K., Chen, V. C. P., Shih, D. T., Rosenberger, J. M., Kim, S. B., & Horton, R. A.

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4572175
 (2018). Data-Driven Optimization for Dallas-Fort Worth Airport Deicing Activities. Annals of January 2012). http://www.eia.gov/cneaf/electricity/epm/epm_sum.html. Accessed August
Operations Research, 263(1–2), 361–384. https://doi.org/10.1007/s10479-017-2747-1. 12, 2016.
Fourer, R., Gay, D. M., & Kernighan, B. W. (1990). A Modeling Language for Mathematical Vazifeh, M. M., Zhang, H., Santi, P., & Ratti, C. (2019). Optimizing the deployment of electric
Programming. Management Science, 36(5), 519–554. vehicle charging stations using pervasive mobility data. Transportation Research Part A:
https://doi.org/10.1287/mnsc.36.5.519. Policy and Practice, 121, 75–91. https://doi.org/10.1016/j.tra.2019.01.002.
Friedman, J. H. (1991). Multivariate adaptive regression splines. The annals of statistics, 19(1), Wetz, D. (2010). Energy Storage Needs and Options. Renewable Energy Sources, The University
1-67. https://doi.org/10.1214/aos/1176347963. of Texas at Arlington.

ed
Furushima, T., & Manabe, K. (2011). Superplastic micro-tubes fabricated by dieless drawing Wu, T. H., Pang, G. K., Choy, K. L., & Lam, H. Y. (2015, June). An optimization model for a
processes. Superplastic Forming of Advanced Metallic Materials (327–360). Elsevier. battery swapping station in Hong Kong. 2015 IEEE Transportation Electrification
http://dx.doi.org/10.1533/9780857092779.3.327. Conference and Expo (ITEC) (1-6). IEEE.
Gerding, E., Stein, S., Robu, V., Zhao, D., & Jennings, N. R. (2013, May). Two-sided online Yang, L., Chen, J., Li, W., & Wen, Z. (2021). Research on Optimizing the Location and Capacity
markets for electric vehicle charging. Proceedings of the 2013 international conference on of Electric Vehicle Charging Stations. Lecture Notes of the Institute for Computer Sciences,
Autonomous agents and multi-agent systems. Social Informatics and Telecommunications Engineering, 76–90.
Gorbunova, A., & Anisimov, I. (2020). Optimizing the Location of Urban Charging Stations: https://doi.org/10.1007/978-3-030-77569-8_6.
Case study of the City of Tyumen, Russian Federation. WIT Transactions on the Built Yang, Z., Chen, V. C. P., Chang, M. E., Sattler, M. L., & Wen, A. (2009). A Decision-Making

iew
Environment. https://doi.org/10.2495/ut200021. Framework for Ozone Pollution Control. Operations Research, 57(2), 484–498.
GUROBI Optimization, LLC. (2022). GUROBI Optimization Reference Manual v9.5 (Version https://doi.org/10.1287/opre.1080.0576.
9.5). https://www.gurobi.com. Yudovina, E., & Michailidis, G. (2015). Socially Optimal Charging Strategies for Electric
He, F., Yin, Y., & Zhou, J. (2015). Deploying public charging stations for electric vehicles on Vehicles. IEEE Transactions on Automatic Control, 60(3), 837–842.
Urban Road Networks. Transportation Research Part C: Emerging Technologies, 60, 227– https://doi.org/10.1109/tac.2014.2346089.
240. https://doi.org/10.1016/j.trc.2015.08.018. You, P., Yang, Z., Chow, M. Y., & Sun, Y. (2015). Optimal cooperative charging strategy for a
Jekabsons, G. (2016). ARESLab: Adaptive regression splines toolbox for Matlab/Octave, smart charging station of electric vehicles. IEEE Transactions on Power Systems, 31(4),
2011. URL http://www. cs. rtu. lv/jekabsons. 2946-2956.
Kang, Q., Wang, J., Zhou, M., & Ammari, A. C. (2016). Centralized Charging Strategy and Zhang, C., Peng, K., Zhang, X., Jiang, Y., Liu, Y., & Cai, Y. (2023). A robust optimal dispatching
Scheduling Algorithm for Electric Vehicles Under a Battery Swapping Scenario. IEEE strategy of distribution networks considering fast charging stations integrated with
Transactions on Intelligent Transportation Systems, 17(3), 659–669. photovoltaic and energy storage. Frontiers in Energy Research, 11,

ev
https://doi.org/10.1109/tits.2015.2487323. https://doi.org/10.3389/fenrg.2023.1126295.
Khosrojerdi, A., Allen, J., Lee, W. J., & Mistree, F. (2012, September). Designing an Electric
Charging Station for Plug-In-Hybrid-Electric-Vehicles in the Face of Uncertain
Demand. 12th AIAA Aviation Technology, Integration, and Operations (ATIO) Conference
and 14th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference.
Kulvanitchaiyanunt, A., Chen, V. C., Rosenberger, J., Sarikprueck, P., & Lee, W. J. (2015,
May). A linear program for system level control of regional PHEV charging stations. 2015
IEEE Industry Applications Society Annual Meeting. (2046-2052). IEEE.

r
Lambert, F. (2016). Norway keeps electric vehicle tax exemption until 2020, positions itself to
stay EV leader. Electrek. https://electrek.co/2016/11/09/norway-keeps-electric-vehicle-tax-
exemption-until-2020-positions-itself-to-stay-ev-leader/. Accessed September 16, 2017.
Lamontagne, S., Carvalho, M., Frejinger, E., Gendron, B., Anjos, M. F., & Atallah, R. (2022).
Optimising Electric Vehicle Charging Station Placement using Advanced Discrete Choice
Models. ArXiv:2206.11165 [Math]. http://arxiv.org/abs/2206.11165.
Lin, C.-F., LeBoulluec, A. K., Zeng, L., Chen, V. C. P., & Gatchel, R. J. (2014). A Decision-Making
Framework for Adaptive Pain Management. Health Care Management Science, 17(3), 270–
283. https://doi.org/10.1007/s10729-013-9252-0.
Lin, W., & Hua, G. (2015, July). The flow capturing location model and algorithm of electric
vehicle charging stations. 2015 International Conference on Logistics, Informatics and
Service Sciences (LISS) (1-6). IEEE.
er
pe
Ma, J., & Zhang, L. (2018). A Deploying Method for Predicting the Size and Optimizing the
Location of an Electric Vehicle Charging Stations. Information, 9(7), 170.
https://doi.org/10.3390/info9070170.
Mak, H. Y., Rong, Y., & Shen, Z. J. M. (2013). Infrastructure planning for electric vehicles
with battery swapping. Management science, 59(7), 1557-1575.
https://doi.org/10.1287/mnsc.1120.1672.
MATLAB. (2016). The Math Works, Inc., 2016. (Version 2016a)
https://www.mathworks.com/.
Miller, A., & Lumby, B. (2012). Utility scale solar power plants: a guide for developers and
investors. Guidelines book written for IFC. World Bank Group, New Delhi, India.
ot

Minitab, LLC. (2016) Salford Predictive Modeler v8.0 (Version 8.0). https://www.salford-
systems.com/products/.
Murtagh, B. A., & Saunders, M. A. (1977). MINOS. A large-scale nonlinear programming
system (for problems with linear constraints). User’s guide. Defense Technical Information
Center. http://dx.doi.org/10.21236/ada044906.
National Renewable Energy Laboratory. (2012). National Solar Radiation Data Base.
tn

http://rredc.nrel.gov/solar/old_data/nsrdb/. Accessed July 26, 2016.

Park, C. J., Lee, J., Park, G. L., & Hyun, J. S. (2014). Development of Reservation
Recommendation Algorithms for Charging Electric Vehicles in Smart-Grid
Cities. International Journal of Smart Home, 8(1), 113–
122. https://doi.org/10.14257/ijsh.2014.8.1.12.
Pilla, V. L., Rosenberger, J. M., Chen, V. C. P., & Smith, B. (2008). A Statistical Computer
Experiments Approach to Airline Fleet Assignment. IIE Transactions, 40(5), 524–537.
https://doi.org/10.1080/07408170701759734.
Sacks, J., Welch, W. J., Mitchell, T. J., & Wynn, H. P. (1989). Design and analysis of computer
rin

experiments. Statistical science, 4(4), 409-423.

Sadeghi Ahangar, S., Abazari, S. R., & Rabbani, M. (2022). A region-based model for
optimizing charging station location problem of electric vehicles considering disruption -
A case study. Journal of Cleaner Production, 336, 130433.
https://doi.org/10.1016/j.jclepro.2022.130433.
Shih, D. T., Kim, S. B., Chen, V. C. P., Rosenberger, J. M., & Pilla, V. L. (2014). Efficient Computer
Experiment-Based Optimization through Variable Selection Annals of Operations Research,
216(1), 287–305. https://doi.org/10.1007/s10479-012-1129-y.
ep

Siddappa, S., Günther, D., Rosenberger, J. M., & Chen, V. C. P. (2007) Refined Experimental
Design and Regression Splines Method for Network Revenue Management. Journal of Revenue
and Pricing Management, 6(3), 188–199. https://doi.org/10.1057/palgrave.rpm.5160086.
Soltani, N. Y., Kim, S. J., & Giannakis, G. B. (2014). Real-time load elasticity tracking and
pricing for electric vehicle charging. IEEE Transactions on Smart Grid, 6(3), 1303-1313.
Sule, N. V., Chen, V., & Sattler, M. (2011). A Decision-Making Framework for Assessing Control
Strategies for Ground Level Ozone. Atmospheric Environment, 45(28), 4996–5004.
https://doi.org/10.1016/j.atmosenv.2011.05.070.
Pr

Tan, J., & Lin, W. H. (2014, October). A stochastic flow capturing location and allocation
model for siting electric vehicle charging stations. 17th international IEEE conference on
intelligent transportation systems (ITSC) (2811-2816). IEEE.
Tang, Z. C., Guo, C. L., Hou, P. X., Fan, Y. B., & Jia, D. M. (2013). Optimal Planning of
Electric Vehicle Charging Stations Location Based on Hybrid Particle Swarm
Optimization. Advanced Materials Research, 724–725, 1355–1360.
https://doi.org/10.4028/www.scientific.net/amr.724-725.1355.
Tsai, J. C. C., Chen, V. C. P., Beck, M. B., & Chen, J. (2004). Stochastic Dynamic Programming
Formulation for a Wastewater Treatment Decision-Making Framework. Annals of Operations
Research, 132(1–4), 207–221. https://doi.org/10.1023/b:anor.0000045283.86576.62.
U.S. Department of Energy. (2012). Electric Power Monthly March 2012 (with Data for

This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=4572175