IEEE TRANSACTIONS ON SMART GRID, VOL. 14, NO.

6, NOVEMBER 2023 4621

Distributionally Robust Model Predictive Control

for Smart Electric Vehicle Charging Station
With V2G/V2V Capability
Hoang Tien Nguyen , Graduate Student Member, IEEE, and Dae-Hyun Choi , Member, IEEE

Abstract—This paper proposes a distributionally robust model the power distribution grid operation. Furthermore, EVCSs
predictive control (DRMPC) for energy management of a vehicle- still rely on power from coal-fired power plants to meet the
to-grid (V2G)/vehicle-to-vehicle (V2V)-enabled smart electric demands of EV charging, which cannot reduce the greenhouse
vehicle charging station (EVCS) with a photovoltaic (PV) system
and an energy storage system. The proposed DRMPC method gas emissions and carbon pollution completely [2].
aims to reduce the operational cost of the EVCS while ensur- Motivated by these challenges, smart EVCSs with a solar
ing the desired charging demands of electric vehicle (EV) users photovoltaic (PV) system and an energy storage system (ESS)
under uncertainties in electricity buying/selling prices, PV gen- are rapidly becoming promising charging facilities. Smart
eration outputs, and future EV charging demands. To cope with EVCSs can decrease their operational costs and maintain
these uncertainties, the proposed method includes the follow-
ing three features: i) tractable reformulation of the worst-case stable distribution grid operations by reducing the power
expected buying cost and selling revenue using a Wasserstein consumed from the grid through utilization of pollutant-free
metric and duality theory, ii) determination of a distribution- PV and ESS powers while providing charging services to
ally robust bound on the random PV generation output using EVs [3], [4]. In addition, vehicle-to-grid (V2G) and vehicle-
its support information, and iii) a scenario-based approach to to-vehicle (V2V) technologies allow the EVs to supply their
predicting the future EV charging demand. To improve com-
putational efficiency, a penalty method is proposed to relax the stored power back to the grid and exchange power with
complementarity constraints, while still ensuring nonsimultane- other EVs via the bidirectional chargers of smart EVCSs,
ous charging and discharging of EVs under the derived sufficient respectively, thereby enabling stable power grid operation and
conditions. Numerical examples using a real-world operational economical EV charging [5]. These advantages of smart EVCS
dataset of the EVCS are provided to demonstrate the effectiveness suggest the need for developing an optimal energy manage-
of the proposed DRMPC method under uncertain environments
in terms of the EVCS cost saving via V2G/V2V capability, data ment algorithm for efficient scheduling of V2G/V2V-enabled
utilization, and computational complexity. smart EVCSs equipped with the PV system and ESS.
However, uncertainties in electricity buying/selling price,
Index Terms—Electric vehicle charging station, distribution-
ally robust optimization, model predictive control, uncertainty, PV generation output, and heterogeneous EV charg-
vehicle-to-grid, vehicle-to-vehicle. ing/discharging for smart EVCS may cause maloperations in
smart EVCS via abnormal ESS and EV charging/discharging,
thereby increasing the operational cost of smart EVCS.
I. I NTRODUCTION Furthermore, under uncertain smart EVCS operation, the
charging/discharging scheduling of a large number of EVs
OMPARED to gasoline vehicles, electric vehicles (EVs)
C are becoming essential for clean power distribution
system operation owing to their significant reductions in green-
with V2G/V2V functionality may increase the computational
complexity of the smart EVCS energy management signif-
icantly. The main objective of this study is to develop a
house gas emissions and carbon pollution during movement
computationally efficient optimization framework that reduces
and power charging from electric vehicle charging stations
the operational cost of the PV/ESS-integrated smart EVCS via
(EVCSs) [1]. However, simultaneous charging of a large num-
V2G/V2V while coping with the aforementioned uncertainties.
ber of EVs via the EVCS may have detrimental impacts on
Several studies have formulated optimization problems
Manuscript received 4 October 2022; revised 10 January 2023 and 6 March wherein smart EVCS is optimally scheduled to minimize its
2023; accepted 27 March 2023. Date of publication 31 March 2023; date operational cost while meeting the EV charging preferences.
of current version 23 October 2023. This work was supported in part Multi-stage optimization frameworks consisting of day-ahead
by the Basic Science Research Program through the National Research
Foundation of Korea (NRF) funded by the Ministry of Education under Grant and real-time EV scheduling were developed in which the
2022R1F1A1062888, and in part by NRF funded by the Korea Government operational cost of smart EVCS and the peak-to-average ratio
(MSIT) under Grant 2021R1A4A1031019. Paper no. TSG-01480-2022. of the system are reduced by utilizing available energy of PV
(Corresponding author: Dae-Hyun Choi.)
The authors are with the School of Electrical and Electronics system [6] and both PV system and ESS [7] in the smart
Engineering, Chung-Ang University, Seoul 156-756, South Korea (e-mail: EVCS, respectively. A deep reinforcement learning method
[email protected]; [email protected]). for energy management of multiple PV/ESS-integrated smart
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TSG.2023.3263470. EVCSs was proposed in [8], where the EVCSs exchange the
Digital Object Identifier 10.1109/TSG.2023.3263470 surplus energy of ESSs with each other to reduce operational
1949-3053 
c 2023 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: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
4622 IEEE TRANSACTIONS ON SMART GRID, VOL. 14, NO. 6, NOVEMBER 2023

costs. In [9], the cost degradation model of the ESS along capability under uncertain electricity buying/selling
with the levelized cost of PV power was integrated in the prices, PV generation outputs, and future EV charging
smart EVCS scheduling problem to reflect a realistic operation demands.
of the smart EVCS. Much of the recent research has con- 2) We present tractable reformulations for two objective
sidered optimization approach using the V2G/V2V capability functions including the worst-case expected buying cost
of EVs to reduce the charging costs of conventional EVCS and selling revenue in the DRMPC problem. The orig-
and maintain stable power distribution system operation. This inal DRMPC problem becomes a finite-dimensional
ranges from a distributed model predictive control (MPC) for convex problem through these reformulations, which
real-time voltage regulation [10], adaptive controller of EV can be solved by off-the-shelf optimization solvers. In
to stabilize the distribution grid frequency via V2G [11], and addition, we formulate a distributionally robust chance
V2V energy sharing framework for economical EVCS opera- constrained problem (DRCCP) integrated with physical
tion [12]. More recently, a mixed-integer linear programming- limits (i.e., supports) for the random PV generation to
based MPC method was proposed in [13], in which the smart determine the reasonable lower bound of the random PV
EVCS charging cost is reduced by utilizing PV power and generation output. Since only the lower bound of PV
V2G capability. However, none of previous studies incor- generation output is employed in the DRMPC problem,
porated multiple uncertainties in stochastic electricity price, the computational complexity of the DRMPC problem
PV generation output, and EV charging/discharging in the is small and fixed even with the increasing number of
optimization framework, which can degrade the scheduling PV generation output samples.
performance of smart EVCS under uncertain environments. 3) We propose a data-driven scenario-based approach
Traditionally, robust optimization (RO) [14] and stochastic that predicts uncertain future EV charging/discharging
optimization (SO) [15] have been employed to handle uncer- demands on the prediction horizon of the DRMPC. A
tainties in power system operations. However, SO and RO key part of this approach is to incorporate a scenario-
have the following limitations, respectively: i) having prior based future EV charging/discharging constraint using
knowledge of the true distribution of uncertain data and ii) historical data into the DRMPC model to improve the
calculating too conservative solution based on the worst-case performance of the DRMPC problem.
scenario. To resolve the drawbacks of SO and RO, distribution- 4) We adopt a penalty method to reduce the computation
ally robust optimization (DRO) has been developed recently. time of the DRMPC problem while guaranteeing nonsi-
Compared to SO and RO, the DRO requires no knowledge multaneous optimal charging and discharging of EVs by
of the true distribution of uncertain data and calculates a relaxing their non-convex complementarity constraints.
less-conservative robust solution based on an ambiguity set We derive two sufficient conditions with which nonsi-
comprising potential distributions of uncertain data [16]. There multaneous optimal charging and discharging of the EVs
are two types of ambiguity sets using moment information are ensured under the penalty method.
(e.g., mean and variance of the distribution) and statistical dis- Section II describes a system model of the smart EVCS
tance (e.g., Wasserstein metric). Since the distributions cannot along with the Wasserstein metric and notations used in
be described using only the mean and variance, Wasserstein- this work. The proposed DRMPC problem is formulated in
based DRO has emerged as a prominent approach with Section III. The reformulation of the DRMPC problem and
advantages such as asymptotic consistency, tractable reformu- relaxation of the EV complementarity constraints are presented
lation, and finite sample guarantee [17]. The Wasserstein-based in Section IV. Section V presents some numerical examples,
DRO has been adopted to handle uncertainties in various smart and the conclusions are noted in Section VI.
grid applications (e.g., economic dispatch [18], peer-to-peer
energy trading [19], and optimal power flow [20]).
In summary, previous studies on energy management of II. P RELIMINARIES
smart EVCS with V2G/V2V capability posed several chal- A. Notation
lenges. First, the uncertainties in the electricity buying/selling Let 
x denote a random variable with support  = [x, x],
prices and PV generation outputs were not modeled explic- where x and x represent the lower and upper bounds of  x,
itly in the optimization problem, which deteriorated the smart respectively. The symbol  xm denotes the mth sample of  x.
EVCS operational performance. Second, the future EV charg- The expectation with respect to the probability distribution
ing demands in the MPC prediction horizon influence optimal P is given as EP . The space M() denotes the probability
scheduling of smart EVCS; however, previous studies do not space that includes all probability distributions P supported
consider them in the MPC optimization problem. Third, an on . The operator ·  indicates an arbitrary norm in Rn .
increase in the number of binary variables associated with Let Pch(dch),x,y represent the charging (discharging) power of
charging/discharging of a larger number of EVs can increase component x from (to) component y.
the solving time of the optimization problem significantly.
This study aims to address these challenges and the main
contributions are presented as follows: B. Wasserstein Metric
1) We develop a Wasserstein-based distributionally robust To measure the distance between the probability distribu-
MPC (DRMPC) framework for optimal scheduling of tions of the uncertainties, we use the following Wasserstein
the PV/ESS-integrated smart EVCS with V2G/V2V metric.

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
NGUYEN AND CHOI: DRMPC FOR SMART EVCS WITH V2G/V2V CAPABILITY 4623

Fig. 2. Illustration of the power flows among the grid, PV system, ESS, and
Fig. 1. PV/ESS-integrated smart EVCS with V2G/V2V functionality. current/future EVs in the smart EVCS.

Definition 1 (Wasserstein Metric [17]): The Wasserstein V2G function. The current and future EVs also consume power
metric of two probability distributions denoted by P1 , P2 ∈ from the PV system (Pch,evc,pv , Pch,evf,pv ) and ESS (Pch,evc,ess ,
M() is defined as follows: Pch,evf,ess ). The ESS charges power from the grid (Pch,ess,grid )
 and PV system (Pch,ess,pv ) when the buying price is low and
   
dW (P1 , P2 ) := inf  ξ2  d
ξ1 −  ξ1 , d
ξ2 (1) PV generation is available, respectively. The ESS discharges
2 its stored power to the EVs and grid (Pdch,ess,grid ) to increase
where  is a joint distribution of two random variables,  ξ1 the revenue of the EVCS. The surplus PV generation power
and ξ2 with marginal distributions P1 and P2 , respectively. (Ppv,grid ) is sold to the grid. All the power flows considered
Using the notion of a Wasserstein ball of radius  centered in this study are nonnegative.
at the empirical distribution 
PN := N1 N m=1 δξm based on N
historical samples 
ξm (where δξm is the unit point mass at 
ξm ), III. O PTIMIZATION P ROBLEM F ORMULATION
an ambiguity set PN is defined as follows: In this section, we show formulation of the DRMPC-based
 
PN = P ∈ M() : dW P,  PN ≤  . (2) optimization problem to schedule the optimal operation of
the smart EVCS with V2G/V2V functionality while cop-
With a confidence level 1 − β to ensure that the exact ing with uncertainties in buying/selling prices, PV generation
distribution is within the ambiguity set, the radius  is outputs, and future EV charging demands. The goal of the
calculated by [20] DRMPC optimization problem is to minimize the multi-
  objective functions (5)–(8) while satisfying the operational
1 1
(N) = D ln . (3) constraints (9)–(33) of the smart EVCS.
N 1−β
Here, D is a constant that can be obtained by solving the A. Objective Function
following optimization problem For each current EV i ∈ Etc and future EV i ∈ Et,j f at time t
    with k ∈ Kt = {1, 2, . . . , Ht } and variable prediction horizon

2 1 
N
D = min  1 + ln

eρξm −
μ1
2
, (4) Ht (i.e., maximum departure time of the EVs in Etc ∪ Et,j f ),
ρ≥0 ρ N the objective function of the DRMPC problem is defined as
m=1
follows:
μ is the sample mean of 
where  ξ.
min J1 − J2 + ω3 J3 , (5)
⎡ ⎤
C. System Model of Smart EVCS With PV System and ESS  buy buy
J1 = sup EP ⎣ 
πk Pk t⎦, (6)
We consider a situation in which the EVs perform P∈PN buy k∈Kt
V2G/V2V services at a PV/ESS-integrated smart EVCS as ⎡ ⎤
shown in Fig. 1. Fig. 2 describes the power flows among the 
grid, PV system, ESS, and EVs in the smart EVCS. In this fig- J2 = inf EP ⎣ 
πksell Psell
k t⎦, (7)
P∈PN sell
ure, the current and future EVs represent the scheduled EVs k∈Kt
at the current and future time steps on the prediction hori- 
 evc

evc,des 2
J3 = Edi ,i − Ei 
zon of the MPC, respectively. For V2V, the current and future
i∈Etc
EVs exchange their powers with the current EVs (Pch,evc,evc )
and between future EVs (Pch,evf,evf ), respectively. Furthermore, 1    evf

evf,des 2
+ Ed ,i,j − Ei,j  . (8)
the current EVs charge their powers from the future EVs Nd i
d f j∈N i∈Et,j
(Pch,evc,evf ), and vice versa (Pch,evf,evc ). The current and future
EVs not only consume power from the grid (Pch,evc,grid , In (6) and (7), J1 and J2 are the expectations of the total buy-
Pch,evf,grid ) via the grid-to-vehicle (G2V) function, but also ing cost and selling revenue of the smart EVCS with respect to
supply power back to the grid (Pdch,evc,grid , Pdch,evf,grid ) via the the worst-case probability distribution of the stochastic buying

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
4624 IEEE TRANSACTIONS ON SMART GRID, VOL. 14, NO. 6, NOVEMBER 2023

ch,ess,grid ch,ess,pv
and selling prices ( π buy , 
π sell ) within the Wasserstein ambi- the grid (Pk ) and PV system (Pk ) and ii) ESS
guity set defined in (2), respectively. The parameter N buy(sell) discharging power to the grid (Pk
dch,ess,grid
) and current/future
represents the number of samples of the random buying (sell- EVs (Pch,evc,ess , Pch,evf,ess ). Constraint (14) presents the SOE
k k
ing) price. In (8), J3 is a regularizer that helps meet the desired dynamics of the ESS, which is expressed in terms of the
energy demand of the current EV i ∈ Etc and future EV i ∈ Et,j f ,
SOE at time k (Ekess ), charging/discharging power (Pch,ess ,
k
which is defined as the difference between the state of energy Pdch,ess ) with charging/discharging efficiency (η ch,ess , ηdch,ess ),
(SOE) of each EV i at departure time di (Edevc i ,i
, Edevf
i ,i,j
) and its k
and scheduling unit time t. The SOE of the ESS is lim-
evc,des evf,des
desired SOE (Ei , Ei,j ). Note that two terms (J1 − J2 ited by (15). The limits of the ESS charging and discharging
and J3 ) have a tradeoff relationship between the decrease of powers are described by (16) and (17), where bess is a
k
the EVCS operational cost and the satisfaction of the desired binary decision variable (18) that determines the charging and
SOE for EVs in terms of a weight ω3 in (5). This study aims to discharging status of the ESS.
place a higher priority on the satisfaction of the desired SOE
for EVs than on the minimization of the EVCS operational
D. Current EV Constraints
cost. Therefore, the value of the weight ω3 is selected with a
large number. In addition, Eievc,des and di of current EV i are For k ∈ Kt , the constraints of the current EV i ∈ Etc are as
assumed to be provided by the EV user when it is plugged follows:
evf,des
in a charger. Ei,j and di of future EV i are obtained from  ch,evc ch,evc,grid ch,evc,pv
Pk,i = Pk + Pch,evc,ess
k + Pk
the historical operational data of the EVCS, and a scenario set
i∈Etc
N d including N d EV prediction scenarios is used to forecast
the future EVs. + Pch,evc,evc + Pch,evc,evf , (19)
 k k
dch,evc,grid
Pdch,evc
k,i = Pk + Pch,evc,evc
k + Pch,evf,evc
k , (20)
B. EVCS Buying/Selling Power Constraints i∈Etc
For k ∈ Kt , the power exchanged between the EVCS and Pdch,evc t
k,i
grid through the transformer at the substation is limited by the evc
Ek+1,i = Ek,i
evc
+ ηich,evc Pch,evc
k,i t− , (21)
following constraints: ηidch,evc
evc
buy ch,evc,grid ch,evf,grid ch,ess,grid Eevc ≤ Ek+1,i
evc
≤ Ei , (22)
Pk = Pk + Pk + Pk , (9) i
ch(dch),evc ch(dch),evc
dch,evc,grid dch,evf,grid dch,ess,grid pv,grid 0 ≤ Pk,i ≤ Pi , k ≤ di , (23)
k = Pk
Psell + Pk + Pk + Pk , (10)
ch(dch),evc
buy
≤P
trans
, Psell
trans Pk,i = 0, k > di , (24)
Pk k ≤P . (11)
ch,evc dch,evc
buy Pk,i Pk,i = 0. (25)
In (9), Pk is the buying power of the EVCS from the grid,
which is the sum of the charging powers of current/future Constraint (19) denotes the aggregated charging powers
ch,evc,grid ch,evf,grid ch,ess,grid
EVs (Pk , Pk ) and ESS (Pk ). In (10), of all the current EVs in the EVCS, which is the sum of
ch,evc,grid
Psell
k is the selling power of the EVCS to the grid, which is the charging power consumed from the grid (Pk ) via
ch,evc,pv
the sum of the discharging powers of the current/future EVs G2V function, ESS (Pch,evc,essk ), PV system (P k ), and
dch,evc,grid dch,evf,grid dch,ess,grid
(Pk , Pk ), ESS (Pk ), and surplus PV other current/future EVs (Pch,evc,evc , Pch,evc,evf
) via the V2V
pv,grid k k
power generation (Pk ). The buying and selling powers process. Constraint (20) expresses the aggregated discharg-
trans
are limited by the capacity of the transformer (P ) in (11). ing powers of all current EVs associated with the V2G/V2V
functionality, which is the sum of the discharging power
C. ESS Constraints (Pdch,evc,grid ) to the grid via the V2G process and power
For k ∈ Kt , the operational constraints of the ESS in the exchanged (Pch,evc,evc
k , Pch,evf,evc
k ) with the other current/future
smart EVCS are as follows: EVs via the V2V process. Constraint (21) indicates the SOE
dynamics of each current EV i, which is expressed in terms
ch,ess,grid ch,ess,pv
Pch,ess
k = Pk + Pk , (12) of the SOE at time k (Ek,i evc ), charging/discharging power

Pdch,ess =
dch,ess,grid
Pk + Pch,evc,ess + Pch,evf,ess , (13) (Pch,evc
k,i , Pdch,evc
k,i ) with charging/discharging efficiency (ηich,evc ,
k k k
dch,evc
P dch,ess
t ηi ), and scheduling unit time t. The SOE of the EV is
ess
Ek+1 = Ekess + ηch,ess Pch,ess t − k dch,ess , (14) limited by (22). During a parking time (k ≤ di ), the charging
k
η ch(dch),evc
ess (discharging) power (Pk,i ) of the EV i is constrained
Eess ≤ Ek+1
ess
≤E , (15) by (23). Constraint (24) ensures that the EV stops charging
0 ≤ Pch,ess ≤ P
ch,ess ess
bk , (16) and discharging after it departs from the EVCS (k > di ). The
k
dch,ess   complementarity constraint (25) enforces nonsimultaneous
0 ≤ Pdch,ess
k ≤P 1 − bess
k , (17) charging and discharging of the EV.
bess
k ∈ {0, 1}. (18)
Constraints (12) and (13) encode the charging (Pch,ess
k ) and E. Future EV Constraints
discharging (Pdch,ess
k ) powers of the ESS, respectively. The for- In general, accurate prediction of future EV charg-
mer and latter consist of the i) ESS charging power from ing/discharging demands on the MPC prediction horizon

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
NGUYEN AND CHOI: DRMPC FOR SMART EVCS WITH V2G/V2V CAPABILITY 4625

is a challenging problem owing to uncertainties in the 1 − α under the worst-case probability distribution P within a
arrival/departure times, desired SOE, and battery capacities Wasserstein ball. Here, α and N pv are the violation probabil-
of the heterogeneous future EVs along with the complex ity and the number of samples of the random PV generation
V2G/V2V process. To resolve this, we adopt a simple yet output, respectively.
effective scenario-based approach that requires no distribu- Remark 1: The formulated DRMPC problem has the fol-
tion of the EV operational characteristics. This approach lowing drawbacks. First, the DRMPC problem becomes
leverages and embeds the EV charging/discharging scenar- infinite-dimensional because of two reasons: i) the worst-
ios in the DRMPC model to characterize the future EVs case expected objective functions (6) and (7) with respect
charging/discharging demands. to an infinite-dimensional space of probability distributions
Let N d = {1, 2, . . . , N d } denote a set of EV scenarios and ii) the chance constraints (33) associated with the infi-
obtained from the historical operational data of the EVCS, mum over a set of probability distributions. Evidently, solving
where N d is the number of scenarios. The historical operational the infinite-dimensional DRMPC problem is a challenging
data provides information about arrival/departure time, desired task. Second, the non-convex complementarity constraints (25)
SOE, and battery capacity of EVs at the smart EVCS. The and (32) of the EVs increase the computational time of the
scenario-based constraints for the future EVs are as follows: DRMPC problem significantly. In the next section, we present
1   ch,evf ch,evf,grid ch,evf,pv the proposed solution approach to overcome these drawbacks.
d
Pk,i,j = Pk + Pch,evf,ess
k + Pk
N
j∈N i∈Et,j
d f

IV. S OLUTION A PPROACH

+ Pch,evf,evf
k + Pch,evf,evc
k , (26)
A. Reformulation of Objective Functions J1 and J2
1   dch,evf dch,evf,grid
d
Pk,i,j = Pk + Pch,evf,evf
k + Pch,evc,evf
k . For the tractable finite-dimensional convex problem refor-
N dj∈N i∈Et,j
f mulations of the worst-case expected objective functions J1
(27) and J2 in (6) and (7), the result in Proposition 1 [16] using
the strong duality theory is adopted in our DRMPC model.
The left-hand side of (26) and (27) represents the aver- Proposition 1: Suppose that a random vector  ξ ∈ Rn is
age charging and discharging demands of all future EVs with supported on a closed and convex set  ⊆ R and that the
n
respect to a total of N d scenarios, respectively. These aver- Wasserstein ambiguity set PN is constructed from the sample
age values are considered as their predictions, which are set {
ξ 1,
ξ 2, . . . ,
ξ N }. If the objective function l(
ξ ) is upper
equal to the charging and discharging powers of the future semi-continuous, the worst-case expectation supP∈PN EP [l(ξ )]
EVs interacting with the grid, PV system, ESS, and other is equivalent to
future/current EVs. The other constraints (28)–(32) of the 
future EVs are formulated as constraints (21)–(25), similar to infλ≥0,sm ∈R λ + N1 N sm
   m=1  (34)
those of the current EVs except for the subscript j of the vari- s.t. supξ ∈ l 
ξ − λ ξ −ξ 1 ≤ sm .
able for the scenario index and arrival time ai,j of the future
EV i under scenario j. Note that if l(ξ ) is convex, for any fixed λ, the supremum
Pdch,evf t in (34) can only be obtained at the vertexes ξ , ξ , or  ξ m.
ch,evf ch,evf k,i,j
evf
Ek+1,i,j = Ek,i,j
evf
+ ηi,j Pk,i,j t − dch,evf
, (28) Given a random buying price  πk
buy
with the closed
ηi,j buy buy
and convex support [π k , π k ] and sample set
evf
i,j ≤ Ek+1,i,j ≤ E i,j ,
Eevf evf buy buy buy
(29) πk,1 , 
{ πk,2 , . . . , 
πk,N buy }, the worst-case expected buying cost
ch(dch),evf ch(dch),evf J1 is reformulated using Proposition 1 as follows:
0 ≤ Pk,i,j ≤ Pi,j , ai,j ≤ k ≤ di,j , (30)
ch(dch),evf ⎡ ⎤
Pk,i,j = 0, k < ai,j , k > di,j , (31) 
sup EP ⎣ t⎦
buy buy
J1 = 
πk Pk
Pch,evf dch,evf
k,i,j Pk,i,j = 0. (32) P∈PN buy k∈Kt
⎧ ! "
buy buy N buy buy
⎪ J
⎪ 1 = inf k∈Kt λk k + N buy
1
m=1 sk,m
⎪
buy buy
F. PV System Constraints ⎪ λk ≥0,sk,m ∈R
! "
⎪
⎨ s.t. π buy Pbuy t + λbuy π buy −  buy buy
To address the uncertainty of the random PV generation πk,m ≤ sk,m ,
= k k k ! k " (35)
output, we formulate the following chance constraint (33): ⎪
⎪ buy buy
π k Pk
buy buy
t − λk π k − 
buy buy
πk,m ≤ sk,m ,
⎪
⎪
 ⎪
⎩ buy buy buy
inf P Pk
ch,ess,pv
+ Pk
ch,evc,pv 
πk,m Pk t ≤ sk,m .
P∈PN pv
 However, Proposition 1 cannot be directly adopted to the
≤
ch,evf,pv pv,grid pv
+ Pk + Pk Pk ≥ 1 − α. (33) worst-case expectation J2 owing to its converse expression. To
tackle this challenge, we reformulate J2 into a tractable con-
The chance constraint (33) implies that the PV power sup- vex problem using the Wasserstein metric and strong duality
ch,ess,pv ch,evc,pv ch,evf,pv pv,grid
port (i.e., sum of Pk , Pk , Pk , and Pk ) to theory, as illustrated in Proposition 2.
the ESS, current/future EV, and grid must be less than or equal Proposition 2: Given the stochastic selling price 
πksell with
to the PV generation output, with a minimum probability of the closed and convex support [π k , π k ] and sample set
sell sell

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
4626 IEEE TRANSACTIONS ON SMART GRID, VOL. 14, NO. 6, NOVEMBER 2023

{
πk,1
sell , 
πk,2
sell , . . . , 
πk,N
sell }, the worst-case expected selling rev-
sell (N pv is the number of samples of the random PV generation
enue J2 is reformulated as output). wk,m , rk,m
max , r min , z
k,m k,m and vk are auxiliary variables.
⎡ ⎤ pv
 k is the radius of a PV-generation-output-related Wasserstein
P⎣
J2 = inf E 
πksell Psell
k t⎦ ball at time k. Problem 1 is solved only when PV genera-
P∈PN sell
k∈Kt tion is available during the daytime. Furthermore, Problem 1
⎧ ! "
⎪ N sell sell is a bilinear programming problem, which can be effectively
⎪ J2 = supλsell k k + N sell
−λsell sell 1
⎨ k ≥0,sk,m ∈R
sell k∈Kt m=1 sk,m
! " solved with commercial solvers such as GUROBI and IPOPT.
= s.t. π sell t − λsell π sell
k − πk,m
sell sell ≥ ssell , pv
⎪ k Pk Remark 2: The solution Pk of Problem 1 is employed
⎪
⎩
k k,m
 sell Psell t ≥ ssell .
πk,m k k,m to establish the constraint (38), which is embedded in the
pv
(36) DRMPC problem. Note that only the lower bound Pk is
used to compute the optimal EV charging/discharging sched-
Proof: See Appendix A. ule instead of a large number of PV generation output samples.
Thus, the computational complexity of the DRMPC problem
B. Reformulation of PV Chance Constraint is small and fixed even with the increasing number of PV
The chance constraint (33) of PV generation output can be generation output samples.
converted to a computationally tractable form using a direct Remark 3: The physical limits of random PV generation
DRO method [21]. However, this direct method incurs a heavy outputs in the support-integrated DRCCP (Problem 1) restrict
computational burden with the increasing number of historical the computed lower bounds within their physical limits,
samples. This motivates us to propose the following solu- thereby ensuring a realistic solution of the DRMPC frame-
tion approach that addresses PV chance constraint with high work. By contrast, the support-free method (Appendix C)
computational performance. ignores the physical limits of the PV generation outputs and
The chance constraint (33) is rewritten using a lower-bound hence generates the unrealistic lower bounds of PV generation
variable Pk of 
pv pv
Pk as follows: outputs.
# pv pv $
inf P Pk ≤  Pk ≥ 1 − α, (37)
P∈PN pv C. Relaxation of EVs Complementarity Constraints
where Similar to the constraints (16)–(18) of the ESS, the non-
ch,ess,pv ch,evc,pv ch,evf,pv pv,grid pv convex complementarity constraints (25) and (32) of the EVs
Pk + Pk + Pk + Pk ≤ Pk . (38)
are relaxed to constraints with binary variables using which the
To obtain a tight lower bound for the random PV generation optimization problem can be solved through a mixed-integer
output on the entire prediction horizon, we solve the following programming (MIP) method [23]. However, the MIP method
DRCCP: may yield a significantly higher computation time of the
 pv DRMPC problem as the number of binary variables increases
max Pk s.t. (37). (39)
pv
Pk owing to the large number of EVs. To resolve this, we pro-
k∈Kt
pose a penalty method [24], [25] that reduces the computation
Note that in the tractable reformulation of (37) the physi- time of the DRMPC problem significantly while ensuring
cal limits  ] of the random variable 
pv pv,min pv,max pv
Pk ∈ [Pk , Pk Pk , the optimal solution along with the nonsimultaneous charg-
namely support of the random variable, are explicitly reflected. ing/discharging of the EVs. A key component of the proposed
The support can be constructed using the historical PV gen- penalty method is the removal of the complementarity con-
eration output data along with the nominal power Ppv,nom of straints (25) and (32) and addition of the following penalty
the PV system. Finally, using the result in Theorem 2 [22], function J4 (40) with weight ω4 to the objective function of
the support-integrated DRCCP is reformulated as Problem 1. the DRMPC problem.
Problem 1 (Distributionally Robust Bound Determination): ⎛ ⎞
 ⎜  dch,evc   dch,evf ⎟
 pv J4 = ⎝ Pk,i + Pk,i,j ⎠. (40)
max Pk k∈Kt i∈Etc
pv
Pk j∈N d i∈Et,j
f
k∈Kt

N pv Sufficient conditions that guarantees the performance of the
pv
s.t. αN vk − pv
zk,m ≥ k N pv , proposed penalty method are provided in Proposition 3.
m=1 Proposition 3: Complementarity constraints (25) and (32)
! " ! "
can be replaced by the penalty function J4 with a positive
−Pk +  −
pv pv pv,max pv
Pk,m wk,m − Pk max
Pk,m rk,m
! " weight ω4 added to the objective function of the DRMPC
−
pv,min pv problem when two conditions are satisfied:
+ Pk min
Pk,m rk,m ≥ vk − zk,m , π buy Nπ
  1) N1π N j=1 πk,j ≥ N π
1
j=1 πk,j , and
sell
 min 
−wk,m − rk,m
max
+ rk,m  ≤ 1, 2) ηich,evc dch,evc
ηi ch,evf dch,evf
< 1 and ηi ηi < 1.
1
wk,m ≥ 0, rk,m
max
≥ 0, rk,m
min
≥ 0, zk,m ≥ 0. Proof: See Appendix B.
Remark 4: The two sufficient conditions in Proposition 3
Here, 
pv
Pk,m is the mth sample of the PV generation output are readily satisfied in practice because the buying electricity
at time k belonging to the sample set {
Pk,1 , 
Pk,2 , . . . , 
pv pv pv
Pk,N pv } price is generally higher than the selling price [26], [27] and

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
NGUYEN AND CHOI: DRMPC FOR SMART EVCS WITH V2G/V2V CAPABILITY 4627

Fig. 4. Comparison of the operational costs of the DRMPC method with

and without: a) future EV predictions and b) V2G/V2V capabilities.

The capacities of the PV system and transformer were set

trans
to Ppv,nom = 100 kW and P = 150 kW, respectively. For
the ESS, its minimum and maximum capacities were set to
ess
Eess = 20 kWh and E = 200 kWh, respectively, with a
ch(dch),ess
Fig. 3. Simulation profiles for one day: a) actual/predicted scenarios nominal charging (discharging) power of P = 50 kW.
of the buying price, b) actual/predicted scenarios of the selling price, The maximum power of each charger in the EVCS was set
c) actual/predicted scenarios of PV generation, and d) four scenarios with to 13.2 kW. The charging and discharging efficiencies of the
different numbers of EVs.
ESS and EV batteries were set to 0.95. The parameter val-
the charging and discharging efficiencies of the EV batteries ues of the distributionally robust bound problem (Problem 1)
are typically less than 100%. were defined as N pv = 10, α = 0.6. For Problem 2, to sat-
Finally, the proposed DRMPC problem can be formulated isfy the desired SOE of the EV users with the highest priority,
as Problem 2: the weight ω3 of the objective function J3 was set to a large
Problem 2 (DRMPC): value (ω3 = 100). The weight ω4 for the penalty function J4
was selected to have a small value (ω4 = 10−4 ) so as not
min J1 − J2 + ω3 J3 + ω4 J4
to influence the other three objectives. The number of histor-
s.t. Buying/selling power&ESS constraints: (9)- (11), (12)- (18), ical scenarios for future EV prediction was set to N d = 10.
Current&future EVs contraints: (19)- (24), (26)- (31), All parameters for the historical scenarios were obtained from
PV constraints: (38), the operational dataset of the EVCS at Caltech in [28]. The
sizes of the sample sets for the buying and selling prices were
Auxiliary constraints: (35), (36).
identically set to N buy = N sell = 20 with a confidence level
of 1 − β = 0.95. The proposed DRMPC approach was car-
V. N UMERICAL E XAMPLES ried out every 15 min. All optimization problems were tested
using Python 3.6 with the GUROBI solver 9.1.2 on an com-
A. Simulation Setup
puter with an AMD Ryzen 7 2700X Eight-Core Processor at
The performance of the proposed DRMPC approach was 3.7 GHz and 32 GB RAM.
analyzed using the real-world operational data of the EVCS at
Caltech [28]. The conventional EVCS at Caltech was equipped
with 54 chargers that are connected to the power distribu- B. Benefits of Future EV Prediction and V2G/V2V Capability
tion grid through a three-phase transformer. To build a smart Fig. 4(a) demonstrates the impact of the proposed scenario-
EVCS model, the PV system and ESS were added to the based future EV prediction (26), (27) on the performance of
Caltech EVCS. To evaluate the computational performance the DRMPC model for four scenarios in terms of the normal-
of the proposed approach, data from three consecutive days ized cost. Here, the normalized cost represents the operational
were used to construct a scenario. The profiles of the buy- cost of each DRMPC model with and without future EV
ing and selling prices were generated based on the predicted prediction, which are normalized by that of the benchmark
values of the hourly real-time prices [29] with the prediction DRMPC model with accurate future EV prediction, respec-
errors following a Gamma distribution with a shape param- tively. Note from Fig. 4(a) that the DRMPC model with the
eter of 2 and rate parameter of 1. Figs. 3(a) and (b) show proposed future EV prediction has an increased cost of only
the actual/predicted buying and selling prices, respectively. 0.4–1.7% compared to the benchmark model. Meanwhile, the
The profiles of the PV generation outputs were obtained from increase of the DRMPC model without future EV prediction
an open dataset [30]. The prediction errors were assumed to is 2.1–4.8%, which is 2.6–5.3 times more than that with future
follow a normal distribution, where the mean is zero and EV prediction.
variance increases from 0 to 30 on the prediction horizon. Fig. 4(b) compares the normalized cost of the DRMPC
Fig. 3(c) shows the actual/predicted PV generation outputs. model without and with V2G/V2V functionality. Here, the
Four scenarios with different numbers of EVs at the EVCS normalized cost indicates the operational cost of the DRMPC
are depicted in Fig. 3(d). models without V2G and/or V2V, which is normalized by

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
4628 IEEE TRANSACTIONS ON SMART GRID, VOL. 14, NO. 6, NOVEMBER 2023

TABLE I
C OMPARISON OF C OMPUTATION T IME B ETWEEN THE SAA, RO-SAA,
DRO-MIP AND DRO-PENALTY M ETHODS W ITH R ESPECT
TO THE N UMBER OF P RICE S AMPLES

Fig. 5. Comparison of the average operational costs of the SAA, RO-SAA,

and proposed DRO methods in terms of the number of price samples.

that of the DRMPC model with both V2G and V2V. Note computation times of the proposed DRO-Penalty method
from Fig. 4(b) that compared to the DRMPC model with- are higher than those of the SAA and RO-SAA methods.
out V2G and V2V, the DRMPC model with only V2G Nevertheless, the computation times of the proposed method
has a lower cost of 1.1–14.1% along with further reduc- are still acceptable in the DRMPC model with a 15-min
tion of the cost 8.6–17.5% owing to the V2V in addition scheduling time unit. Additional observation from this table
to the V2G function. Therefore, we conclude from Figs. 4(a) is that the computation times of the proposed DRO-Penalty
and (b) that the proposed DRMPC method with future EV method are much less than those of the DRO-MIP method and
prediction and V2G/V2V functionality can effectively decrease the gap of computation time between these two methods gener-
the operational cost of the smart EVCS. ally becomes larger as the number of price samples increases.
This observation demonstrates that the proposed DRO-Penalty
method is computationally efficient compared to the DRO-MIP
C. Performance Comparison With SO and RO Methods method.
In this subsection, the out-of-sample performance of the
proposed DRMPC approach was compared with those of the D. Penalty Method for Complementarity Constraints of EVs
SO and RO methods through Monte-Carlo simulations using This subsection demonstrates the performance of the
3000 samples of the buying and selling prices. For the SO proposed penalty method associated with the complementarity
model, the sample average approximation (SAA) method was constraints of the EVs in terms of the computational com-
adopted by setting the radius of the Wasserstein ball to zero plexity of the DRMPC model. We verify from Fig. 6(a) that
buy
(k = ksell = 0 [17]). The RO model was developed in accor- no simultaneous charging and discharging of the four ran-
dance with the method in [31]; however, this previous work domly selected EVs occur in the proposed penalty method.
was limited to handle only the total buying cost (J1 ) [32]. Fig. 6(b) shows the numerical complementarity gap in terms
In this simulation, the RO method corresponds to the RO- of the penalty weight ω4 , which is defined as follows:
SAA approach in which the total selling revenue (J2 ) is
reformulated using the SAA method. Fig. 5 compares the Numerical Complementarity Gap
    + ch,evc dch,evc ,
total operational cost of the EVCS among the three methods = Pk,i Pk,i + Pch,evf Pdch,evf
k,i,j k,i,j
in terms of the buying/selling price sample (N buy , N sell ). As t∈T k∈Kt j∈N d i∈Etc ∪E f
t,j
expected, the results show that the RO-SAA method yields a
higher cost than the SAA and DRO methods. This is because which represents the sum of the multiplication of simultane-
the RO-SAA method uses only the lower and upper bounds ous charging and discharging of all current and future EVs
of the uncertainties to make decisions, thereby yielding too i ∈ Etc ∪Et,j
f with respect to the prediction horizon K during the
t
conservative solution. Another observation is that the total entire simulation period T . The complementarity gap should
operational costs of the SAA and DRO methods decrease with become zero to ensure the normal operation of EV batteries.
increasing price samples. This is because the probability dis- Fig. 6(b) shows the values of the numerical complementarity
tribution of the uncertainties is approximated more accurately gap with varying penalty weight ω4 . Note from this figure
owing to more samples of the uncertainties. In particular, it that the complementarity gap is zero when the value of ω4 is
is observed that the cost of the DRO method is lower than greater than or equal to 0.0001. Fig. 6(c) compares the total
that of the SAA method in the range of samples [5, 150]. operational cost between the MIP and penalty methods with
This justifies the outperformance of the proposed data-driven varying penalty weight ω4 . Note from this figure that the total
DRO method over the SAA method from the perspective of operational cost of the penalty method equals to that of the
the number of leveraged data while coping with uncertainties. MIP method at ω4 = 0.0001 and the former becomes greater
Table I compares the computation times of the proposed than the latter when ω4 > 0.0001. Based on this simulation
DRMPC method (DRO-Penalty) and three methods (i.e., SAA, result, we select the value of ω4 as 0.0001. Fig. 6(d) compares
RO-SAA, and DRO-MIP) with respect to the number of the computation time of the conventional MIP and proposed
selling/buying price samples. Note from this table that the penalty methods. Note from this figure that the computation

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
NGUYEN AND CHOI: DRMPC FOR SMART EVCS WITH V2G/V2V CAPABILITY 4629

Fig. 7. Distributionally robust bound of the PV generation output: a) without

support and b) with support.

the DRCCP problem to avoid unreasonable lower bounds that

may cause highly conservative solutions.

F. Impact of EV Battery Degradation Cost on the DRMPC

This subsection investigates the impact of EV battery degra-
dation cost on the performance of the proposed DRMPC
framework. The following battery degradation cost is added
into the objective function of the DRMPC problem.
⎡
Fig. 6. Performance validation of the DRMPC approach with the proposed  
penalty method: a) charging and discharging powers of four EVs, b) numerical
 ⎢  ch,evc ch,evc Pdch,evc k,i
complementarity gap with increasing ω4 , c) total operational cost of the MIP J5 = ⎣ ηi Pk,i + dch,evc σi t
and penalty methods with increasing ω4 and d) computational time of the k∈Kt i∈Etc ηi
MIP and penalty method with increasing EVs. ⎤
 dch,evf 
1   Pk,i,j ⎥
+ d ηich,evf Pch,evf
k,i,j + dch,evf σi,j t⎦,
time of the penalty method increases gradually whereas the N
j∈N d
ηi
fi∈Et,j
computation time of the MIP method increases dramatically
as the number of EVs increases. This is because a larger num- where σi and σi,j are the battery degradation prices for current
ber of EVs in the MIP method leads to a greater number of EV i ∈ Etc and future EV i ∈ Et,j
f at scenario j, respectively. In

binary variables in the EV constraints, thereby yielding a sig- the battery degradation cost formula, the battery degradation
nificantly higher computation time. This significant increase in price σ of each current/future EV can be computed based on
the computation time is aggravated when the number of sce- the daily battery wear cost (dc), average daily battery wear
narios for the future EV prediction increases in the DRMPC (dw), and the global wear coefficient (Kw ) as follows:
model. Considering an MPC scheduling time unit of 15 min, dc
Fig. 6(d) confirms that the proposed penalty-based DRMPC σ = ,
dw/Kw
method is practical for real-time charging scheduling of a large
where Kw is determined according to the capacity loss char-
number of EVs.
acteristics of the battery; dc ($/day) is calculated based on
the capital cost and the salvation value of the battery at
E. Distributionally Robust Bound of PV Generation Output the end of cycle life along with the daily discount rate; dw
In this subsection, the merit of the support in the (kWh/day) is calculated using daily amounts of energy con-
DRCCP problem (Problem 1) was verified by comparing the sumed in driving and discharge process through the V2G/V2V.
support-integrated DRCCP with the support-free method [33] The detailed derivation of the battery degradation model and
pv
(Appendix C). The lower bounds Pk of the random PV gen- the procedure of the aging price calculation can be referred
eration outputs with different violation probabilities α are to [34]. For simplicity, we use an identical degradation price
depicted in Fig. 7(a) and (b), corresponding to without and (σi = σi,j = 0.003 $/kWh) for all EVs in our simulation study.
with support, respectively. Note from these figures that an The performance of the DRMPC with the battery degrada-
increasing α results in a higher lower bound of the PV gener- tion cost was analyzed using the same simulation environment
ation output because the lower bound is allowed to be violated for the DRMPC without the battery degradation cost. Similar
more frequently. Fig. 7(a) shows unrealistic lower bounds with to the results without the degradation cost as shown in
negative values when α = 0.2 and 0.4 owing to no support Figs. 4(a) and (b), Figs. 8(a) and (b) show that both the
presence in the DRCCP problem. In contrast, Fig. 7(b) shows proposed future EV prediction method and the V2G/V2V
that all lower bounds have non-negative values within the sup- functionality can reduce the operational cost of the EVCS
port range in the support-integrated DRCCP problem. This significantly even when the degradation cost is considered
demonstrates the necessity of integration of the support into in the DRMPC model. Fig. 9 compares the out-of-sample

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
4630 IEEE TRANSACTIONS ON SMART GRID, VOL. 14, NO. 6, NOVEMBER 2023

demands, scenario-based constraints that reflect the random

future EV demands are embedded in the DRMPC problem.
Lastly, a penalty method based on relaxation of the EV com-
plementarity constraints is presented to significantly reduce the
computational time of the DRMPC problem while guarantee-
ing nonsimultaneous optimal charging and discharging of the
EVs under the derived sufficient conditions. Numerical exam-
ples show that under uncertain environments, the proposed
DRMPC method can achieve savings of the EVCS operational
cost via V2G/V2V, improve data utilization efficiency over the
Fig. 8. Comparison of the operational costs of the DRMPC method including
SAA method, and reduce the computational time over the MIP
the battery degradation cost with and without: a) future EV predictions and method.
b) V2G/V2V capabilities.
A PPENDIX A
P ROOF OF P ROPOSITION 2
Denote the objective function l( ξ) =  πksell Psell
k t, where

ξ = πk is supported on  = [ξ , ξ ]. For notational simplicity,
sell

the dependence of l on the decision variable Psell k is omitted.

We use the 1-norm in the Wassestein metric (1) and rewrite
the worst-case expectation of l( ξ ) using (1) as
 /    
#  $  
inf EP l  ξ = /  l ξ P dξ
inf,P 
P∈PN s.t. |ξ − 
ξ | d
ξ , dξ ≤
2
 /    
infPm ∈M() N1 N  
= / m=1  l ξ  Pm d ξ
N 
s.t. N m=1  |ξ − ξ̂m |Pm dξ ≤ ,
1 
Fig. 9. Comparison of the average operational costs of the SAA, RO-SAA,
and proposed DRO methods including the battery degradation cost in terms
of the number of price samples. where Pm (ξ ) = P(ξ |
ξ = ξ̂m ). The second equality results
from the law of total probability [17]. Then, the dual problem
is formulated as
#  $
performance of the proposed DRO method with that of the inf EP l 
ξ
SAA and RO-SAA methods when all three methods consider P∈PN
N 
the degradation cost. Fig. 9 shows similar results shown in 1     
Fig. 5 without the degradation cost: i) the operational cost of = inf sup l ξ Pm d
ξ
Pm ∈M() λ≥0 N 
the RO-SAA method is the highest due to its conservative m=1
 N 

solution and ii) the operational costs of the DRO and SAA 1   
methods decrease with the increasing number of price sam- +λ ξ − ξ̂m |Pm d
| ξ −
N 
ples and the cost of the DRO method is lower than that of the m=1
N 
SAA method in the range of samples [5, 150]. These results 1  !  "  
demonstrate the efficiency of the proposed DRMPC frame- ≥ sup inf lξ + λ|
ξ − ξ̂m | Pm d
ξ − λ
λ≥0 Pm ∈M() N 
work in terms of the EVCS operational cost even when the m=1
battery degradation cost is considered. (41)
1 
N !  "
= sup −λ + inf l 
ξ + λ|
ξ − ξ̂m | . (42)
VI. C ONCLUSION λ≥0 N 
ξ ∈
m=1
This study proposes a DRMPC framework (Problem 2) Given the convex and closed set  and the convex and con-
to schedule optimal operation of a PV/ESS-integrated smart tinuous function l(
ξ ), the inequality (41) becomes an equality
EVCS by minimizing its operational cost and satisfying the by the strong duality result in [35]. Using auxiliary variables
charging demands of the current/future EVs under uncertain- sm , we can rewrite (42) as
ties in buying/selling electricity prices, PV generation outputs, 
#  $
P 
supλ≥0,sm ∈R ! −λ + N1 Nm=1"sm
and future EV charging demands. Two tractable reformu- inf E l ξ =   (43)
lations for the worst-case expected buying cost and selling P∈PN  
s.t. infξ ∈ l ξ + λ|ξ − ξ̂m | ≥ sm
revenue are presented using the Wasserstein metric and dual- ⎧ N
⎪ ! "m ∈R −λ
! + N" m=1 sm
1
⎪ supλ≥0,s
ity theory to cope with buying/selling price uncertainties. The ⎨
randomness of PV power generation is captured by determin- = s.t. l ξ − λ ξ − ξ̂m ≥ sm , (44)
⎪
⎪ ! "
ing the distributionally robust bound (Problem 1) of the PV ⎩ l ξ̂m ≥ sm .
generation output integrated with the support of its physical
limit, which consequently prevents calculation of an irra- Problem (44) is derived from (43) because l(
ξ ) is a positive
tional bound. To handle the uncertainty in future EV charging linear function in 
ξ and hence the infimum of the left-hand

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
NGUYEN AND CHOI: DRMPC FOR SMART EVCS WITH V2G/V2V CAPABILITY 4631

◦ ch,evc,pv ◦ dch,evc,grid
side in (43) occurs at only ξ or ξ̂m . Solving (44) for all time With Pk > 0, Pk > 0, complementary slack-
steps k ∈ Kt is equal to Proposition 2. ch,evc,pv
ness (47) implies λk
dch,evc,grid
= λk = 0. Therefore, we
obtain
A PPENDIX B pv,grid
P ROOF OF P ROPOSITION 3 μch,evc
k + μdch,evc
k = −λk ≤0
Karush-Kuhn-Tucker (KKT) conditions for a convex ⇒ μch,evc
k + ηich,evc ηidch,evc μdch,evc
k < 0. (53)
problem {min f0 (x) s.t. fi (x) ≤ 0, hi (x) = 0} with primal pv
◦
optimal x and dual optimal (λ, μ) are written as The second inequality follows as μch,evc k = −λk ≤ 0
◦ ◦ and the second sufficient condition of Proposition 3 (ηich,evc
Primary feasibility: fi x ≤ 0, hi x = 0, (45) ηidch,evc < 1). By (53), we have a contradiction in (49)
Dual feasibility: λi ≥ 0, (46) because the left-hand side of the equation cannot be equal
◦
Complementary slackness: λi fi x = 0, (47) to zero. This implies that simultaneous EV charging and
◦  ◦  ◦ discharging is suboptimal.
Stationary: ∇f0 x + λi ∇fi x + μi ∇hi x = 0. (48) ◦
Case 2 (Pch,evc,ess
k > 0): Without loss of generality, we
i i
assume that there is only EV i at the EVCS and time step
Since the current and future EVs have similar con- ◦ ◦ ◦ dch,evc,grid ◦
k, and hence Pch,evc,ess
k = Pch,evc
k,i and Pk = Pdch,evc
k,i .
straints, we consider only the current EVs in this proof. ◦ ch,evc ◦ dch,evc

Let μsell denote the dual variable of the power balance • Pk,i ≥ Pk,i : By (21), we compute another feasible
k
constraint (10). Dual variables μch,evc , μdch,evc are defined point with Pch,evck,i and Pdch,evc
k,i = 0 that provides the same
k k
for the charging/discharging power constraints of the current energy to EV i:
EVs (19), (20), respectively. The SOE constraints (21), (22) ◦
◦ Pdch,evc
k,i t
of the EVs are then rewritten as follows: ηich,evc Pch,evc t− = ηich,evc Pch,evc t
  k,i
ηidch,evc
k,i

k−1
ch,evc ch,evc Pdch,evc
n,i t ◦
Ei − E0,i −
evc evc
ηi Pn,i t − dch,evc ≤ 0, ◦ ch,evc Pdch,evc
n=1 ηi ⇔ Pch,evc
k,i = Pk,i −
k,i
  ηich,evc ηidch,evc

k−1
P dch,evc
n,i t evc ◦ ch,evc ◦ dch,evc
evc
E0,i + ηich,evc Pch,evc
n,i t − dch,evc − Ei ≤ 0, < Pk,i − Pk,i ≤ Pi
ch,evc
. (54)
n=1 ηi
dch,ess,grid
Let Pch,evc,ess and Pk denote the ESS charg-
which have dual variables λsoe,evc1 k,i and λsoe,evc2
k,i , respectively. k
ch(dch),evc1 ch(dch),evc2 ing/discharging power corresponding to the feasible point.
The dual variables λk,i , λk,i correspond to the Then, we obtain
lower and upper charging (discharging) constraints of the cur- ◦ ◦
rent EVs in (23). Since the power flows in this study are Pch,evc,ess
k = Pch,evc
k,i < Pch,evc
k,i − Pdch,evc
k,i
(.)
non-negative, λk denotes the dual variable associated with ◦ ◦ dch,evc,grid
(.) = Pch,evc,ess
k − Pk,i . (55)
constraint Pk ≥ 0.
Using the KKT conditions and objective evaluation, we To support the objective evaluations later, we assume that the
prove the non-simultaneous charging and discharging of EVs feasible point and assumed optimal solution have the same
by contradiction. Assume that at some time k ∈ {1, 2, . . . , Ht } selling power Psell
k , which thus yielding
and for some EVs i ∈ Etc , simultaneous charging and discharg- ◦ dch,evc,grid ◦ dch,ess,grid
◦ ch(dch),evc dch,ess,grid
ing, i.e., 0 < Pk,i ≤ Pi
ch(dch),evc
, is the optimal solution Pk = Pk + Pk . (56)
of the DRMPC problem. Thus, the complementary slackness
From (13), (55), and (56), we verify that the power limit (17)
conditions imply λch,evc1 = λdch,evc1 = 0. dch,ess
k,i k,i of the ESS is still satisfied (i.e., Pdch,ess k <P ).
Owing to space limitations, we consider only the case with
Now it is sufficient to compare the objective value of
Pch,evc,evc = 0. The case with Pch,evc,evc > 0 can be derived
k k
dch,evc,grid the feasible point Pch,evc , Pdch,evc and the optimal solu-
similarly. For Pch,evc,evc
k = 0, we obtain Pk > 0 ◦ ch,evc ◦ dch,evc
k,i k,i

by (20). We will consider each power in (19) with non-zero tion Pk,i , Pk,i . While J1 , J2 and J3 are the same with
values to obtain the following three cases: both points, (Pch,evc k,i , Pdch,evc
k,i ) yield smaller values for J4 .
◦ ch,evc,pv ◦ ch,evc ◦ dch,evc
Case 1 (Pk > 0): Stationary conditions (48) associ- Therefore, we conclude that Pk,i , Pk,i is suboptimal.
ch,evc dch,evc ch,evc,pv dch,evc,grid pv,grid ◦ ch,evc ◦ dch,evc ◦
ated with Pk,i , Pk,i , Pk , Pk , and Pk • Pk,i < Pk,i : This case also yields that Pch,evc k,i ,
◦ dch,evc
yield Pk,i is suboptimal in a similar manner.
! " ◦ ch,evc,grid
− μch,evc + ηich,evc ηidch,evc μdch,evc + ηich,evc ηidch,evc ω4 Case 3 (Pk > 0): This case means that the EVs
k k
exchange power with the grid bidirectionally at the same time.
+ ηich,evc ηidch,evc λdch,evc2
k,i + λch,evc2
k,i = 0, (49) Without loss of generality, we assume that there is only one
pv ch,evc,pv
μch,evc
k + λk − λk = 0, (50) EV i ◦at the EVCS. ◦ ◦ ◦
dch,evc dch,evc,grid • Pch,evc
k,i ≥ Pdch,evc
k,i : Similar to (54), for Pch,evc k,i , Pdch,evc
k,i ,
μk − μk − λk
sell
= 0, (51) we find another feasible◦ point that◦ provides the same energy
pv pv,grid
−μk + λk − λk
sell
= 0. (52) to the EV: Pch,evc k,i < Pch,evc
k,i − Pdch,evc
k,i which implies that

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
4632 IEEE TRANSACTIONS ON SMART GRID, VOL. 14, NO. 6, NOVEMBER 2023

ch,evc,grid ◦ ch,evc,grid ◦ dch,evc,grid

Pk,i < Pk,i − Pk,i . With the first sufficient [5] C. Liu, K. Chau, D. Wu, and S. Gao, “Opportunities and challenges of
condition of Proposition 3, we obtain vehicle-to-home, vehicle-to-vehicle, and vehicle-to-grid technologies,”
Proc. IEEE, vol. 101, no. 11, pp. 2409–2427, Nov. 2013.
0 1 0 1 0 1 0 1
  [6] R. Wang, P. Wang, and G. Xiao, “Two-stage mechanism for massive
sup EP πk ≥ EP πk ≥ EP πksell ≥ inf EP πksell ,
buy buy
electric vehicle charging involving renewable energy,” IEEE Trans. Veh.
P∈P P∈P
Technol., vol. 65, no. 6, pp. 4159–4171, Jun. 2016.
from which we can compare the objectives of the optimal and [7] Q. Yan, B. Zhang, and M. Kezunovic, “Optimized operational cost
reduction for an EV charging station integrated with battery energy
feasible points as follows: storage and PV generation,” IEEE Trans. Smart Grid, vol. 10, no. 2,
0 1 0 1 pp. 2096–2106, Mar. 2019.
buy ◦ ch,evc,grid ◦ dch,evc,grid
sup EP πk Pk,i − inf EP πksell Pk,i [8] M. Shin, D.-H. Choi, and J. Kim, “Cooperative management for
P∈P P∈P PV/ESS-enabled electric vehicle charging stations: A multiagent deep
0 1 0 1 reinforcement learning approach,” IEEE Trans. Ind. Informat., vol. 16,
P buy ◦ ch,evc,grid buy ◦ dch,evc,grid
≥ sup E πk Pk,i − sup EP πk Pk,i no. 5, pp. 3493–3503, May 2020.
P∈P
0 1 P∈P [9] K. Chaudhari, A. Ukil, K. N. Kumar, U. Manandhar, and
> sup EP πk Pk,i
buy ch,evc,grid S. K. Kollimalla, “Hybrid optimization for economic deployment of
. ESS in PV-integrated EV charging stations,” IEEE Trans. Ind. Informat.,
P∈P vol. 14, no. 1, pp. 106–116, Jan. 2018.
ch,evc,grid [10] J. Hu, C. Ye, Y. Ding, J. Tang, and S. Liu, “A distributed MPC to
Thus, the feasible point Pk,i has a smaller objective exploit reactive power V2G for real-time voltage regulation in distribu-
function
◦
than the◦
assumed optimal point. tion networks,” IEEE Trans. Smart Grid, vol. 13, no. 1, pp. 576–588,
• Pch,evc
k,i < Pdch,evc
k,i : Similarly, we can obtain the same Jan. 2022.
[11] C. Mu, W. Liu, and W. Xu, “Hierarchically adaptive frequency control
conclusion from for an EV-integrated smart grid with renewable energy,” IEEE Trans.
0 1 0 1 Ind. Informat., vol. 14, no. 9, pp. 4254–4263, Sep. 2018.
buy ◦ ch,evc,grid ◦ dch,evc,grid
sup EP πk Pk,i − inf EP πksell Pk,i [12] M. Shurrab, S. Singh, H. Otrok, R. Mizouni, V. Khadkikar, and
P∈P P∈P
0 1 H. Zeineldin, “An efficient vehicle-to-vehicle (V2V) energy sharing
P sell dch,evc,grid framework,” IEEE Internet Things J., vol. 9, no. 7, pp. 5315–5328,
> inf E πk Pk,i . Apr. 2022.
P∈P
[13] G. R. C. Mouli, M. Kefayati, R. Baldick, and P. Bauer, “Integrated
Thus, simultaneous charging and discharging of the EV in this PV charging of EV fleet based on energy prices, V2G, and offer of
reserves,” IEEE Trans. Smart Grid, vol. 10, no. 2, pp. 1313–1325,
case is suboptimal. This completes the proof. Mar. 2019.
[14] Z. Chen, L. Wu, and Y. Fu, “Real-time price-based demand response
management for residential appliances via stochastic optimization
A PPENDIX C and robust optimization,” IEEE Trans. Smart Grid, vol. 3, no. 4,
F ORMULATION OF THE D ISTRIBUTIONALLY pp. 1822–1831, Dec. 2012.
ROBUST B OUND W ITHOUT S UPPORT [15] M. Bazrafshan and N. Gatsis, “Decentralized stochastic optimal power
flow in radial networks with distributed generation,” IEEE Trans. Smart
The DRCCP (39) for the PV chance constraints is reformu- Grid, vol. 8, no. 2, pp. 787–801, Mar. 2017.
lated using support-free DRO method [33] as follows: [16] R. Gao and A. Kleywegt, “Distributionally robust stochastic optimization
 pv with Wasserstein distance,” Math. Oper. Res., early access, Aug. 2022,
max Pk doi: 10.1287/moor.2022.1275.
pv [17] P. Mohajerin Esfahani and D. Kuhn, “Data-driven distributionally robust
Pk
k∈Kt optimization using the Wasserstein metric: Performance guarantees and
pv

N
pv
tractable reformulations,” Math. Program., vol. 171, no. 1, pp. 115–166,
s.t. αN vk −
pv
zk,m ≥ k N pv , Sep. 2018.
[18] J. Zhai, Y. Jiang, Y. Shi, C. N. Jones, and X.-P. Zhang, “Distributionally
m=1 robust joint chance-constrained dispatch for integrated transmission-
− Pk + 
pv pv
Pk,m + Mqk,m ≥ vk − zk,m , distribution systems via distributed optimization,” IEEE Trans. Smart
  Grid, vol. 13, no. 3, pp. 2132–2147, May 2022.
M 1 − qk,m ≥ vk − zk,m , [19] J. Li, M. E. Khodayar, J. Wang, and B. Zhou, “Data-driven distribution-
qk,m ∈ {0, 1}, zk,m ≥ 0, ally robust co-optimization of P2P energy trading and network operation
for interconnected microgrids,” IEEE Trans. Smart Grid, vol. 12, no. 6,
pp. 5172–5184, Nov. 2021.
where M is a big-M coefficient. [20] C. Duan, W. Fang, L. Jiang, L. Yao, and J. Liu, “Distributionally
robust chance-constrained approximate AC-OPF with Wasserstein met-
ric,” IEEE Trans. Power Syst., vol. 33, no. 5, pp. 4924–4936,
R EFERENCES Sep. 2018.
[1] M. Boucher, “Transportation electrification and managing traffic con- [21] W. Xie, “On distributionally robust chance constrained programs with
gestion: The role of intelligent transportation systems,” IEEE Electrific. Wasserstein distance,” Math. Program., vol. 186, no. 1, pp. 115–155,
Mag., vol. 7, no. 3, pp. 16–22, Sep. 2019. Mar. 2021.
[2] M. Yang, L. Zhang, Z. Zhao, and L. Wang, “Comprehensive ben- [22] A. Arrigo, C. Ordoudis, J. Kazempour, Z. De Grève, J.-F. Toubeau,
efits analysis of electric vehicle charging station integrated photo- and F. Vallée, “Wasserstein distributionally robust chance-constrained
voltaic and energy storage,” J. Clean. Prod., vol. 302, pp. 1–12, optimization for energy and reserve dispatch: An exact and physically-
Jun. 2021. bounded formulation,” Eur. J. Oper. Res., vol. 296, no. 1, pp. 304–322,
[3] U. Datta, A. Kalam, and J. Shi, “Smart control of BESS in PV inte- Jan. 2022.
grated EV charging station for reducing transformer overloading and [23] J. Hu, J. E. Mitchell, J.-S. Pang, K. P. Bennett, and G. Kunapuli,
providing battery-to-grid service,” J. Energy Stor., vol. 28, pp. 1–10, “On the global solution of linear programs with linear complemen-
Apr. 2020. tarity constraints,” SIAM J. Optim., vol. 19, no. 1, pp. 445–471,
[4] A. Sierra, C. Gercek, K. Geurs, and A. Reinders, “Technical, finan- Jan. 2008.
cial, and environmental feasibility analysis of photovoltaic EV charging [24] X. Hu and D. Ralph, “Convergence of a penalty method for mathematical
stations with energy storage in China and the United States,” IEEE J. programming with complementarity constraints,” J. Optim. Theory Appl.,
Photovolt., vol. 10, no. 6, pp. 1892–1899, Nov. 2020. vol. 123, no. 2, pp. 365–390, Nov. 2004.

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.
NGUYEN AND CHOI: DRMPC FOR SMART EVCS WITH V2G/V2V CAPABILITY 4633

[25] K. Garifi, K. Baker, D. Christensen, and B. Touri, “Convex relaxation Hoang Tien Nguyen (Graduate Student Member,
of grid-connected energy storage system models with complementar- IEEE) received the B.Eng. degree (Highest Hons.)
ity constraints in DC OPF,” IEEE Trans. Smart Grid, vol. 11, no. 5, in electrical engineering from the Hanoi University
pp. 4070–4079, Sep. 2020. of Science and Technology, Hanoi, Vietnam, in
[26] Y. M. Ding, S. H. Hong, and X. H. Li, “A demand response energy 2020. He is currently pursuing the M.S. degree
management scheme for industrial facilities in smart grid,” IEEE Trans. in electrical and electronics engineering with
Ind. Informat., vol. 10, no. 4, pp. 2257–2269, Nov. 2014. Chung-Ang University, Seoul, South Korea. His
[27] P. Sharma, M. Kolhe, and A. Sharma, “Economic analysis of a building research interests lie in safe learning, data-driven
integrated photovoltaic system without and with energy storage,” IOP optimization, and control of power systems.
Conf. Ser. Mater. Sci. Eng., vol. 605, no. 1, Aug. 2019, Art. no. 012013.
[28] Z. J. Lee, T. Li, and S. H. Low, “ACN-data: Analysis and applications
of an open EV charging dataset,” in Proc. 10th ACM Int. Conf. Future
Energy Syst. (e-Energy), 2019, pp. 139–149.
[29] A.-H. Mohsenian-Rad and A. Leon-Garcia, “Optimal residential load
control with price prediction in real-time electricity pricing environ-
ments,” IEEE Trans. Smart Grid, vol. 1, no. 2, pp. 120–133, Sep. 2010.
[30] “Open data sets.” Accessed: Aug. 15, 2022. [Online]. Available: https://
site.ieee.org/pes-iss/data-sets Dae-Hyun Choi (Member, IEEE) received the
[31] D. Bertsimas and M. Sim, “Robust discrete optimization and network B.S. degree in electrical engineering from Korea
flows,” Math. Program., vol. 98, no. 1, pp. 49–71, Sep. 2003. University, Seoul, South Korea, in 2002, and the
[32] H. T. Nguyen and D.-H. Choi, “Decentralized distributionally robust M.Sc. and Ph.D. degrees in electrical and com-
coordination between distribution system and charging station opera- puter engineering from Texas A&M University at
tors in unbalanced distribution systems,” IEEE Trans. Smart Grid, early College Station, College Station, TX, USA, in
access, Sep. 28, 2022, doi: 10.1109/TSG.2022.3210232. 2008 and 2014, respectively. He is currently an
[33] Z. Chen, D. Kuhn, and W. Wiesemann, “Data-driven chance con- Associate Professor with the School of Electrical
strained programs over Wasserstein balls,” Oper. Res., Jul. 2022, and Electronics Engineering, Chung-Ang University,
doi: 10.1287/opre.2022.2330. Seoul, South Korea. From 2002 to 2006, he was a
[34] H. Farzin, M. Fotuhi-Firuzabad, and M. Moeini-Aghtaie, “A prac- Researcher with Korea Telecom, Seoul, where he
tical scheme to involve degradation cost of lithium-ion batteries in worked on designing and implementing home network systems. From 2014
vehicle-to-grid applications,” IEEE Trans. Sustain. Energy, vol. 7, no. 4, to 2015, he was a Senior Researcher with LG Electronics, Seoul, where he
pp. 1730–1738, Oct. 2016. developed home energy management systems. His research interests include
[35] A. Shapiro, “On duality theory of conic linear problems,” in Semi- power system state estimation, electricity markets, the cyber-physical secu-
Infinite Programming. Boston, MA, USA: Springer, 2001, pp. 135–165. rity of smart grids, and the theory and application of cyber-physical energy
[Online]. Available: https://link.springer.com/chapter/10.1007/978-1- systems. He received the Best Paper Award at 2012 IEEE Third International
4757-3403-4_7 Conference on Smart Grid Communications in Tainan City, Taiwan.

Authorized licensed use limited to: Pukyong National Univ. Downloaded on October 25,2024 at 14:46:27 UTC from IEEE Xplore. Restrictions apply.