IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 21, NO.

2, APRIL 2024 2125

EV Charging Scheduling Under Demand Charge:

A Block Model Predictive Control Approach
Lei Yang , Student Member, IEEE, Xinbo Geng, Member, IEEE, Xiaohong Guan , Life Fellow, IEEE,
and Lang Tong , Fellow, IEEE

Abstract— This paper studies the online scheduling of electric the charging scheduling under demand charge over multiple
vehicle charging by a service provider subject to a demand charge charging stations.
in a distribution system. Demand charge imposes a penalty on the
peak power consumption over each billing period, representing a Index Terms— Demand charge, demand side management,
substantial cost for the service provider with a large number of online scheduling, charging of electric vehicles, model predictive
clients. Because the demand charge is calculated at the end of the control (MPC).
billing period, it poses challenges in real-time scheduling when
energy demand forecasts are inaccurate, resulting in either overly
conservative power consumption or substantial demand charge. I. I NTRODUCTION
We propose a block model predictive control approach that
decomposes the demand charge into a sequence of stage costs.
Optimality conditions on demand patterns are also presented and
analyzed. Numerical simulations demonstrate the efficacy of the
W E CONSIDER the problem of scheduling electric vehi-
cle (EV) charging in an EV charging facility by a
commercial service provider. The EV charging model accounts
proposed approach. for the stochastic arrival of EVs and their charging demands,
the need to constrain total charging power imposed by the
Note to Practitioners—This paper addresses a significant prac-
tical problem of minimizing the demand charge on the real-time distribution grid, and the costs of EV charging arising from
scheduling of deferrable demands. In particular, we consider a the energy and the demand charge set by the tariff of a local
setting where a commercial electric vehicle (EV) charging service utility. At a higher level, such a scheduling problem falls in
provider has to manage the online scheduling of a large number the category of stochastic dynamic programming, for which
of arriving EVs at a charging facility subject to a maximum the optimal solution suffers from the curse of dimensionality,
charging power constraint and a tariff with the demand charge.
A major practical challenge is to balance the tradeoff between making it unrealistic for practical implementations.
maximizing profit in scheduling as much EV charging as possible The importance of a carefully designed EV charging sched-
and the need to minimize penalty on the peak charging power. ule is well understood, and there is extensive literature on
We propose a model predictive control strategy that decomposes various approaches, each capturing certain aspects of the
the overall demand charge into a sequence of terminal costs. Also EV charging problem; see a brief review in Section I-A.
addressed is the practical constraint arising from the mismatched
EV charging decision period and the power measurement period Among the most significant factors are the need to exploit
used to compute the demand charge. Using real data collected the flexibility of charging demands, avoid peak-demand hours
at the Adaptive Charging Network (ACN) testbed in simulations, during which the energy cost is high, incorporate colocated
the proposed approach yields 8-12% improvement in operational renewables, and deal with dynamically varying electricity
profit over existing benchmarks, while it has yet been tested in prices.
actual charging systems. In the future research, we will address
This paper focuses on two nontrivial but less studied
aspects of centralized EV charging in a distribution system
Manuscript received 16 December 2022; accepted 20 February 2023. Date by a profit-seeking service provider: the stochastic arrival of
of publication 31 March 2023; date of current version 9 April 2024. This
article was recommended for publication by Associate Editor H.-J. Kim and charging demands and the high cost of the demand charge.
Editor L. Moench upon evaluation of the reviewers’ comments. The work of Without a computationally tractable optimal scheduling solu-
Lei Yang and Xiaohong Guan was supported by the National Key Research tion, we aim to develop a suboptimal model predictive control
and Development Program of China under Grant 2016YFB0901900. The
work of Xinbo Geng and Lang Tong was supported in part by the U.S. (MPC) strategy that exploits the structure of EV charging char-
National Science Foundation under Award 1816397 and Award 1809830. acteristics and incorporates the demand charge as a sequence
(Corresponding author: Lei Yang.) terminal cost associated with the optimization at each stage.
Lei Yang is with the Faculty of Electronic and Information Engineer-
ing, Xi’an Jiaotong University, Xi’an 710049, China, and also with the The centralized EV charging in a distribution system allows
School of Electrical and Computer Engineering, Cornell University, Ithaca, us to remove or deemphasize some of the complications
NY 14850 USA (e-mail: [email protected]). commonly considered in the literature, some of which can be
Xinbo Geng and Lang Tong are with the School of Electrical and
Computer Engineering, Cornell University, Ithaca, NY 14850 USA (e-mail: incorporated under the general MPC framework, while others
[email protected]; [email protected]). become unnecessary for the problem at hand. Specifically,
Xiaohong Guan is with the Faculty of Electronic and Information we assume that the marginal cost of electricity is deterministic
Engineering, Xi’an Jiaotong University, Xi’an 710049, China (e-mail:
[email protected]). and the demand charge price fixed as part of a long-term
Digital Object Identifier 10.1109/TASE.2023.3260804 contract with the distribution utility. The behind-the-meter
1545-5955 © 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: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 2126 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 21, NO. 2, APRIL 2024

renewable is not considered. While such resources can be objective. EMPC has been applied to centralized EV charging
incorporated in the MPC-based dynamic optimization, their problem in [29] and [30]. The EMPC models considered
significance rests primarily in the scheduling performance in these contributions differ considerably from our paper’s
rather than in developing an MPC-based solution. approach. Halvgaard et al. [29] formulated one of the earliest
EMPC approaches to EV charging that minimizes charging
costs involving a fixed set of EVs over a scheduling interval.
A. Related Work
More recent work of Engel et al. [30] considers the cen-
The literature on EV charging scheduling is vast. Here we tralized co-optimization of EV charging and building energy
restrict our review to centralized scheduling of a large number management systems. Missing in these formulations are the
EV chargings by a profit-seeking operator. arrival and departure dynamics of EV demands and stochastic
A significant line of contributions to the EV charging lit- charging completion deadlines.
erature follows a baseline of charging optimizations involving Demand charge can be a significant part of its cost for
a fixed set of EVs [1], [2], [3], [4], [5], [6], [7], [8], [9]. a profit-maximizing operator of a large EV charging ser-
Most work under such formulations applies naturally to the vice. Because demand charge is determined by the maximum
residential EV charging problems, where distributed imple- power consumption (typically measured over multiple charg-
mentations are often important. While these approaches can ing periods), it is nontrivial to capture such cost within each
be adapted to the centralized EV charging problem considered receding-horizon MPC optimization. In [8], Lee et al. proposed
in this paper, they have very different objectives from that an MPC method that decomposed the demand charge into
of a commercial EV charging service provider. Some of multiple convex problems, where heuristics were introduced
the missing ingredients in these models are the stochastic to weight the demand charge to influence the performance on
arrival/departure of charging requests, charging demands, and the overall profit. Although not designed for the EV charging
charging completion deadlines. problem, the work of Kumar et al. [31] and the recent paper of
Our approach presented in this paper follows the online Risbeck and Rawlings [32] also bear significant relevance to
job-scheduling formulation where charging demands (jobs) our work. Similarly to [8], the approach of Kumar et al. [31]
arrive and depart sequentially with uncertainty. One line of to incorporating the demand charge in MPC is heuristic by
approaches is to cast the problem as the classical (deter- setting a time-varying weight at each optimization problem.
ministic) deadline scheduling problem [10], [11], [12], [13], Risbeck and Rawlings, on the other hand, proposed a princi-
[14], [15] where jobs with completion deadlines are centrally pled approach that explicitly incorporated the demand charge
scheduled in real-time. In the context of EV charging, jobs as a terminal cost of each EMPC optimization in [32]. Their
are EVs with stochastic demands, processors are EV charging approach, however, is designed with respect to a reference tra-
connecting ports, and the operator is the scheduler. The simple jectory that can only be derived from perfect demand forecasts,
and sometimes optimal deterministic scheduling schemes such and the maximum consumption that sets the demand charge
as the earliest deadline first (EDF) and the least-laxity-first must be computed within a single decision interval. In practice,
(LLF) algorithms remain to be the benchmarks for compar- the EV charging decision interval can be significantly shorter
isons. These approaches treat job arrivals as deterministic and than the period over which the maximum demand charge is
arbitrary. computed, for which no existing techniques explicitly account,
The scheduling of EV charging under stochastic arrivals to our best knowledge.
starts with a Markov decision process (MDP) formulation of
the scheduling problem [16], [17], [18], [19], [20], [21], [22],
[23], [24], [25]. The approach presented in this paper follows
B. Summary of Results and Contributions
the problem formulation of these approaches. Recognizing that
the MDP solutions are intractable in practice, efforts have We formulate a centralized EV charging problem as an
been made to discover and exploit the structural properties of MDP where our model captures the arrival dynamics of EVs,
the problem to achieve suboptimal (sometimes asymptotically stochastic charging demands, and stochastic completion dead-
optimal) performance [21], [22], [23], [24], [25]. In partic- lines. A significant contribution is incorporating the demand
ular, when the charging demand arrival is light relative to charge in the MDP formulation and deriving a novel block
overall system charging capacity, it is shown in [23] that, economic MPC solution, referred to as BMPC, to deal with
without the demand charge, the Whittle’s index policy [26] the demand charge that must be computed over multiple
is asymptotically optimal. Outside the light traffic regime scheduling intervals. A key ingredient of BMPC is to trans-
and without demand charge, the Less Laxity First with Later form the terminal demand charge into a sequence of stage
Deadline (LLF-LD) scheduling algorithm proposed in [27] terminal costs of each BMPC optimization by tracking the
exhibits near-optimal performance in simulations. peak consumption and assessing the impact of the demand
MPC, more precisely receding-horizon economic MPC charge in each stage. Although the idea of tracking the peak
(EMPC) [28], is a computationally tractable solution that often demand was considered in [33], [34], and [32], BMPC differs
performs well in practice when dealing with optimization from existing techniques in terms of the specific stage costs
problems involving dynamic systems and economic objectives. used in the BMPC optimization and how the peak consumption
Such a model fits naturally to the large-scale EV charging is computed over multiple scheduling intervals. To support an
problem for a service provider with a profit-maximizing MPC approach for EV charging applications, we show that the

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: EV CHARGING SCHEDULING UNDER DEMAND CHARGE: A BLOCK MPC APPROACH 2127

MPC solution is optimal when the traffic arrivals are clustered

and sparse.
Extensive numerical simulations were performed based on
the real dataset collected from the Adaptive Charging Network
(ACN) testbed, which has provided an open database of
EV charging since February 2016 and successfully operated
and delivered over 1000 MWh to avoid tons of greenhouse
gases [8]. We compared BMPC with four benchmark algo-
rithms, including MPC with scaled demand charge (MPC-
scaled), EMPC [32], MPC without demand charge (MPC-w/o)
and LLF-LD [27]. Numerical results demonstrate that BMPC
achieves the best performance in most cases. In particular, Fig. 1. Schematic of EV-DC at a public facility.
BMPC can achieve 8% more total reward than the second-best
approach when EV charging requests are stochastic. Compared
with MPC-w/o and LLF-LD, BMPC can obtain more than A1) At each interval,2 a charger can only work on one
12% total reward on average, even under 80% forecast accu- EV, and each EV can receive service from only one
racy. charger at any given time.
The remainder of this paper is organized as follows. A2) Each charger has a constant charging rate R at each
We formulate the EV-DC problem as an MDP in Section II. interval t ∈ T . The charging decision at interval t
In Section III, we develop BMPC as the online scheduling for the i-th charger is a binary variable u i,t , i =
algorithm and demonstrate the numerical results in Section 1, . . . , N , with 1 activating and 0 deactivating the
IV. Finally, we conclude the paper in Section V. charging port.
A3) The EV arriving at the i-th charger at the beginning
of interval t0 reveals random Di,t0 (total amount of
II. EV C HARGING U NDER D EMAND C HARGE energy to be completed) and Ti,t0 (deadline). An EV
will be automatically removed at the end of deadline.
We consider the problem of scheduling of EV charging
A4) The operator receives a per unit reward π r and
under demand charge (EV-DC) at a public facility [16], [18],
pays a time-varying charging cost πte if it serves
[23], [35], [36], where EVs with charging demands arrive
an EV at interval t. For simplicity, we assume that
stochastically, each with a random amount of energy request
parameters π r and {πte }t=0 T −1
are deterministic. The
and specified deadline for completion [23].
proposed approaches can be easily extended towards
stochastic settings.
A5) If the total charging demand of EV i is not completed
A. Nominal Model Assumptions before it leaves, then a penalty occurs at unit price q.
Consider an EV charging facility with N chargers (charging We also assume that the penalty price is greater than
ports) as illustrated in Fig. 1. EVs arriving at the charging the largest charging cost over the whole horizon, i.e.,
facility are assigned randomly to one of the available chargers. q ≥ πte , ∀t ∈ T .
We assume that, upon arrival, the EV reveals to the operator Remark 1: Comments and justifications for these assump-
its charging demand and deadline for completion.1 tions are in order. Assumption A1 is standard and holds for
The operator faces a deadline scheduling problem, aimed most existing charging stations [39]. The on-off charging
at completing as many EV charging jobs as possible by their model assumed in A2 is a reasonable approximation to the
deadlines. The reward for the operator is the revenue from actual charging profile3 [40], and it can be extended to cases
serving EV demands. The cost, on the other hand, comes from with continuously varying charging rates [8]. A3 assumes that
the electricity consumed in EV charging, the demand charge random charging requests and completion time are known
imposed by the distribution utility, and the penalty when the upon arrival, which is necessary for quality-differentiated EV
charging demands are not fulfilled. The operator also faces charging services. Some commercial and business EV charging
the constraint that only a finite number of chargers can be facilities impose the duration of charging [37], [38]. A4 is
activated simultaneously due to transformer constraints from the standard assumption that the price of charging (per kWh)
the distribution circuit. is the same for all the EVs [41]. The marginal cost of EV
We consider a finite scheduling horizon T , indexed by charging (mostly from the energy cost for the service provider)
t ∈ T := {0, 1, · · · , T − 1}. Some of the key details and is constant for EVs processed within the same interval t.
assumptions of EV-DC are outlined below. Generalizations to the time-varying case such as [42] are
possible. Because there are physical constraints (e.g., from
1 For EV charging at public parking facilities at shopping and office
2 Throughout the paper, we use “stage” to index optimization and “interval”
complex, a customer may not stand by until the charging is completed.
It is therefore natural that a deadline for completion is included in the to index time (decision epoch).
charging demand so that the customer can be away and return at the time 3 In practice, the charging rate may vary depending on the SoC, especially
of completion [37], [38]. when the battery is nearly fully charged.

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 2128 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 21, NO. 2, APRIL 2024

power system transformers) that limits the amount of power

delivery at any given time, the EV charging service provider
may not be able to complete charging requests by the deadline
when the demand is high. In such cases, we assume in
A5 that the service provider must compensate the customer
with the penalty price q to enhance the quality of charging
services [43], [44], where the soft deadline can be hardened
by setting the non-completion penalty much higher than the Fig. 2. Temporal structure of ℓ scheduling intervals and one measurement
window under the monthly demand charge (DC). The average power con-
charging cost. Thus it is always optimal (i.e., reward maxi- sumption over a measurement window is calculated as Rℓ (ut + · · · , ut+ℓ−1 ).
mizing) to finish as many EV requests as possible.

B. Problem Formulation
4) Demand Charge: Based on the pricing of the demand
EV loads are flexible demands such that their services can
charge, the average power consumption over ℓ consecutive
be delayed. By shifting part or all of the demands, EV-DC
scheduling intervals is calculated for each non-overlapping
has strong inter-temporal dependencies. We present in section
measurement window.5 The maximum value among these
a Markov decision process (MDP) model for EV-DC [21],
measurement windows for the demand charge is represented
[27], subject to various constraints including the maximum
by variable ψ and calculated as follows:
charging rate, time-varying electricity price and a monthly
demand charge. t+ℓ−1 N
1) Exogenous Stochastic Input ξ : The input of EV-DC R X X
ψ = max′ u i,τ , (4)
model is a vector random process that models the arrivals of t∈T ℓ τ =t i=1
EV demands at individual chargers. The occupancy of each
charger is an on-off process with the charger being occupied
where T ′ = {0, ℓ, 2ℓ, . . . , T − ℓ} denotes the set of the
for the duration of the EV charging deadline and being idle for
beginning interval of each measurement window. Thus the
the duration of a Bernoulli process based on with parameter
demand charge is computed as C(ψ) = π d ψ, where π d is
αi set by the overall arrival rate of the EV demands.4 At
the unit price of the demand charge.
the beginning of an occupied period of charger i, say at t0 ,
Remark 2: For the computation of the demand charge,
an EV arrives with random energy demand Di,t0 and random
practically the length of a measurement window may not
deadline Ti,t0 . Thus the input process at charger i is given by
match the control resolutions, e.g., operating a charger at every
ξi,t = (Di,t0 , Ti,t0 ) for t = t0 , . . . , Ti,t0 . When the charger is
1-5 minutes for fast charging rather than 15 minutes [45].
idle, ξi,t = (0, 0). With probability αi , ξi,t = (0, 0) transitions
Therefore, we introduces an integral parameter ℓ representing
to ξi,t+1 = (Di,t+1 , Ti,t+1 ).
the total number of scheduling intervals contained in a mea-
2) System State and State Evolution: The state of charger
surement window, i.e., the length of a measurement window
i at interval t is given by a tuple xi,t = (ri,t , τi,t ), where ri,t
is identical to ℓ intervals, and its value obviously relies on the
represents the remaining demand to be served by deadline Ti,t
specific application settings, as shown in Fig. 2.
at charger i and τi,t = Ti,t − t the lead time to the EV’s
5) Reward: The reward collected from all the EVs at
deadline at interval t. Hence, the system state is modeled as
( interval t is given by
xi,t − (u i,t , 1) if τi,t ) > 1,
xi,t+1 = (1)
ξi,t if τi,t ≤ 1.
 
N
X X
G(xt , ut , ξ t ) = R1(π r − πte ) u i,t − q (ri,t − u i,t )
Note that when the charger is free, its state is (0, 0). When
i=1 i∈Jt
there is no EV arriving at charger i, the state of the charger
remains at (0, 0). (5)
3) Constraints: The total amount of power used for charg-
ing at one interval is limited by where 1 denotes the length of a scheduling interval and Jt
N the set of EVs that will leave at interval t + 1, i.e., Jt := {i :
τi,t = 1}.
X
u i,t ≤ M, t ∈ T , (2)
i=1 6) MDP Formulation: The objective of EV scheduling is
where M denotes the maximum number of simultaneous to find the optimal control policy {µ∗t }t∈T to maximize the
chargers allowed by the maximum power constraint of the expected total reward in the presence of the demand charge.
local transformer (M < N ). The charger cannot be activated At each interval t, a control law maps states to controls:
when no EV is connected:
ut = µt (xt , ξ t ). (6)
u i,t ≤ xi,t , i ∈ {1, . . . , N }, t ∈ T . (3)
4 The assumption on the Bernoulli process is made for the theoretical
analysis shown in Section III-C, whereas the algorithm proposed in this paper 5 For demand side management applications of power systems, the length
does not require specific EV arrival process. Thus ACN dataset can be used of a measurement window is often 15 or 30 minutes. In this paper we fix the
for the numerical tests shown in Section IV. value at 15 minutes.

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: EV CHARGING SCHEDULING UNDER DEMAND CHARGE: A BLOCK MPC APPROACH 2129

Given an initial state x0 , the EV-DC can be formulated as an formulated as:

integer linear program: " #
X
"T −1 # max Eξ G(xt , ut , ξ t ) − C(φT )
{µt }t∈T
t∈T
X
max Eξ G(xt , ut , ξ t ) − C(ψ)
{µt }t∈T
t=0
s.t. (1)(2)(3)(6)(8),
s.t. (1)(2)(3)(4)(6), (7) (9)

Remark 3: The model defined in (7) can be easily extended With the amended state, it is much easier to apply the
to a much broader class of deferrable load scheduling under idea of MPC-based approaches on the reformulated problem
demand charge problems beyond EV-DC, such as applications (9). Generally, MPC considers the optimization problem of
mentioned in [46] and [33]. The main difficulties of dealing a shorter horizon {t, · · · , t + W } and utilizes a predicted
with the demand charge in the MDP framework come from the trajectory {ξ̂ k }t+W
k=t
−1
. As a result, BMPC solves the following
mismatch of different timescales. In particular, three timescales deterministic problem for a forecast window of length W with
coexist in the formulated model (See Fig. 2): the current state xt and φt at t ∈ T ′ :

(1) the control ut = µt (xt , ξ t ) happens at every t ∈ T = t+W

X−1
G xk , uk , ξ̂ k − H(φt+W )


{0, 1, · · · , T − 1}; max (10a)
{uk }t+W
k=t
−1
(2) the peak average consumption that sets the demand charge k=t
(
is calculated at the end of each measurement window at xi,k − (u i,k , 1) if τi,k > 1,
t = {ℓ, 2ℓ, · · · , T }; s.t. xi,k+1 = (10b)
ξ̂ i,k if τi,k ≤ 1,
(3) the demand charge is imposed at the terminal T .
N
X
Because the reward G(·) is realized at every interval t whereas u i,k ≤ M, (10c)
the demand charge C(ψ) is levied at the end of the entire i=1
control horizon, it is challenging to balance the trade-off u i,k ≤ xi,k , i ∈ {1, . . . , N } (10d)
between the immediate reward and the uncertain demand
φk updates as in (8), (10e)
charge set at the end of the scheduling horizon.
k = t, . . . , t + W − 1,

III. B LOCK M ODEL P REDICTIVE C ONTROL where H(φt+W ) is the BMPC end-of-horizon cost (to be spec-
ified in Section III-B). The main difference between BMPC
We present a variation of the standard MPC scheduling to and the nominal MPC is the block structure. Instead of moving
address two demand charge related issues. The first arises from from t to t + 1, BMPC moves one block (ℓ intervals) at each
the mismatch between the EV charging control intervals and stage, i.e., from t to t + ℓ. The optimal controls of (10) in
the demand charge measurement periods. Instead of realizing a the first block {u∗t , · · · , u∗t+ℓ−1 } will be implemented. Others
single decision in each decision epoch in the traditional MPC, {u∗t+ℓ , · · · , u∗t+W −1 } are only advisory.
BMPC realizes a block of decisions. The second arises from Remark 4: Different from those EMPC methods [28] that
the need of appropriating the terminal demand charge cost to aim to minimize the gap between a controlled system and
each stage decision. The proposed framework is termed Block a reference in process control engineering, BMPC aims to
MPC because the rolling window moves a block of ℓ intervals maximize total reward for EV charging applications where
at a time. stability and tracking error are not of concern.
Remark 5: Generally, the complexity of BMPC grows
exponentially with the number of unfinished jobs, since the
A. Block Model Predictive Control Under Demand Charge decision variables are integers. However, the state-of-art com-
mercial solvers such as Gurobi and CPLEX work efficiently
To address the timescale mismatch issues arising from the to handle integer programs with over thousands of variables
demand charge (see Remark 2), we introduce an additional [47], [48], and the computation time of the proposed approach
system state φt at every interval t, which represents the highest is in an acceptable range, as shown in Section IV.
average consumption over an ℓ-sized measurement window The BMPC approach is summarized as Algorithm 1. Two
until interval t. The new state variable φt evolves according factors affect the performance of BMPC: an end-of-horizon
to cost H(φt+W ), and an initial guess on the maximum aver-
 ( Pt PN ) age consumption φ0 . The choice of the end-of-horizon cost
 max φ , τ =t−ℓ+1 i=1 u i,τ

if t + 1 ∈ T ′ H(φt+W ) lies at the heart of BMPC solution to EV-DC.
t
φt+1 = ℓ Detailed discussions and comparisons are in Section III-B. The
φt initial guess φ0 can be derived by the mainstream techniques

otherwise,

(8) for load forecasting at the distribution level [49], [50], such as
regression or neural networks by using the historical demands
Note that φt+1 ≥ φt for every t ∈ T . Therefore, given the as the training samples. The influence on this value is also
system state x0 and φ0 , the EV-DC (7) can be equivalently discussed in Section III-C.

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 2130 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 21, NO. 2, APRIL 2024

Algorithm 1 BMPC Under Demand Charge justified. In this case, we have

1: Initialization: Initialize system with x0 , φ0 , W . X X
2: for t ∈ {0, ℓ, · · · , T − ℓ} do C(φT ) = C(φ0 ) + H(φt+W ) = C(φ0 ) + C(φt+1 − φt ),
3: Observe the system current state xt , the up-to-date t∈T ′ t∈T
(14)
highest average consumption φt and the EV arrival ξ t ;
4: Forecast the future EV arrivals {ξ̂ k }t+W −1
k=t+1 over t + 1 to which enables us to define a revised reward function that
t + W − 1; considers the cost of demand charge at each t ∈ T :
5: Solve the BMPC Problem (10);
V(xt , ut , ξ t , φt ) := G(xt , ut , ξ t ) − C(φt+1 − φt ). (15)
{u∗k }k+W
k=t
−1
= BMPC(xt , φt , {ξ̂ k }t+W
k=t
−1
) (11)
It is clear that (13) is a direct result of formulating BMPC
using (15). The equation above reveals that (13) embeds
6: Take the first block (ℓ intervals) of controls from the the demand charge cost, which occurs at the end of control
solution to (10): U∗t = (u∗t , . . . , u∗t+ℓ−1 ) horizon, into each stage as decomposed in (15). Therefore,
7: for j ∈ {t, t + 1, . . . , t + ℓ − 1} do (13) effectively avoids the inferior performance by directly
8: Observe the system state x j and the input ξ j ; using the demand charge structure such as (12a) or (12b).
9: Take u∗j = µ∗j (x j , ξ j ) as the decision at j and update
the system state to x j+1 according to (1);
10: Update φ j according to (8). C. Optimality of BMPC
11: end for MPC rarely achieves optimality in general. To justify and
12: end for support the proposed solution, we consider the particular case
when the proposed BMPC happens to be optimal. Specifically,
we assume that EV arrivals are clustered during operation
B. BMPC End-of-Horizon Cost H(φt+W ) hours; many EVs sometimes arrive close to one another,
and no EV arrives between clusters.6 Such characteristics are
Intuitively, good choices of the end-of-horizon cost supported by the real-data set collected at the ACN testbed.
H(φt+W ) should reflect the amortization of the demand charge To formalize the “clustered arrival characteristics”, we intro-
in the current interval t. Some primitive forms could be: duce the notion of the decoupling interval.
Definition 1: An interval t d ∈ T is a decoupling interval if
H(φt+W ) := C(φt+W ), t ∈ T ′ , (12a) (xi,t d , ξi,t d ) = (0, 0) for all i ∈ {1, . . . , N }.
W For the EV-DC problem, a decoupling interval corresponds
H(φt+W ) := C(φt+W ), t ∈ T ′ . (12b)
T to the time period when the total charging demands are zero,
However, these two choices perform poorly in practice because i.e., energy demands of all the parking EVs have been finished
(12a) imposes the demand charge over the entire control before t d and no new ones arrive at t d . We also have several
horizon T on the rolling window. When W ≪ T , the demand assumptions:
charge C(φt+W ) would dominate the total reward of the rolling A7) The length of a scheduling interval of (7) is the same
window. As a result, the solution in this setting will often be as the measurement window, i.e., ℓ = 1.
so conservative that the operator would rather sacrifice most of A8) For ∀t ∈ T , the forecasts of the exogenous variables
t+W −1
the reward than incur a large demand charge cost. In addition, {ξ k }k=t+1 are accurate within the forecast window.
(12a) fails to capture the fact that demand charge is only posed For a given arrival trajectory ξ with an initial state x0 , let
for the peak consumption. JξBMPC (x0 , φ0 ) be the total reward produced by Algorithm 1,
A slightly better choice is (12b), which splits the demand Jξ∗ (x0 ) the optimal reward of (7) and ψξ∗ the corresponding
charge equally among T intervals. This choice essentially optimal maximum average consumption. Then we present the
assumes that the states of EV charging loads within each following optimality conditions of BMPC:
measurement window are almost identical, which is often not Proposition 1: Under Assumption A7 and A8, for every
true in practice. arrival trajectory ξ and initial condition x0 , BMPC achieves
We propose a more judicious choice optimal reward, i.e., JξBMPC (x0 , φ0 ) = Jξ∗ (x0 ), if
(i) each stage optimization includes at least one decou-
H(φt+W ) := C(φt+W − φt ), t ∈ T ′ . (13) pling interval,
(ii) the initial estimate of the maximum consumption is
The rationale behind (13) is twofold. First, it is clear that
perfect, i.e., φ0 = ψξ∗ .
φt+W ≥ φt always holds true according to (8). If the peak
consumption of the current rolling window {t, · · · , t + W } is The proof of Proposition 1 is given in Appendix A.
no higher than the previous one (φt+W = φt ), no additional Remark 7: We take Proposition 1 not as a theoretical guar-
cost should be considered, i.e., H(φt+W ) = 0. Additional antee in practice but as a theoretical ideal for EV charging
cost occurs only when the peak consumption increases, i.e., in practice as the two conditions for optimality are strong; a
φt+W > φt . justification of these conditions is in order.
Remark 6: For EV-DC where the demand charge cost func- 6 Theoretically, under the general Markovian arrival model, there are always
tion is linear, the BMPC end-of-horizon cost can be further periods when no job arrives.

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: EV CHARGING SCHEDULING UNDER DEMAND CHARGE: A BLOCK MPC APPROACH 2131

TABLE I
PARAMETER S ETTINGS

schedules for the demands between t and t pd , whereas the

rest {ut pd , . . . , ut+ℓ−1 } are sub-optimal since only part of the
demand information has been considered. If the next BMPC
starts at t +ℓ, the decisions of the corresponding block are still
sub-optimal, hence BMPC would never reach the optimum
under ℓ > 1. To mitigate this effect, we only need to take
decisions {ut , . . . , ut pd −1 } and start the next BMPC at t pd .
Thus, with the optimality conditions in Proposition 1 satisfied,
Fig. 3. Demonstration of optimality of BMPC under ℓ > 1. BMPC also produces an optimal solution under ℓ > 1.

IV. N UMERICAL R ESULTS

To evaluate the online algorithm proposed in Section III,
The first condition on traffic pattern implies that the arrival we used a large-scale EV charging dataset, collected from a
traffic is bursty and clustered, which has been observed smart charging facility at Caltech on the ACN testbed.7 The
in practice [51] and justified by the ACN dataset (see relevant parameters were summarized in Table I. We assumed
Section IV-C). The specific condition in (i), however, depends that the time-varying electricity prices were deterministic
on the window size used in BMPC. In practice, the window and collected from the Electric Reliability Council of Texas
size is chosen a priori, which may not guarantee (i). In such (ERCOT).8 Note that we set πr > maxt∈T πte and q > πr
cases, one can create virtual decoupling intervals by elimi- to encourage chargers to serve as many charging requests as
nating some jobs from scheduling and paying for the penalty possible rather than cause large penalty. All the simulations
for those unscheduled jobs. The optimal selection of jobs for were conducted by a computer with i7-10700 CPU and 16 GB
elimination requires further study and is outside the scope of RAM using MATLAB and Gurobi.
this paper.
The second condition on the perfect initial estimation of
A. Aggregated EV Loads
ψξ∗ cannot generally hold. In practice, ψξ∗ is estimated using
historical data and is bound to have errors. Empirically, we find To investigate the impact of the demand charge under
in our experiments that the performance of BMPC is not different volumes of traffic, we generated a series of charging
sensitive to the estimation error, and the loss of total reward loads based on the original dataset, by aggregating the EVs
diminishes when the total scheduling horizon is long. In our arriving during the same time period on each day into one day.
simulations, the performance loss could be negligible when For example, 2-day aggregation level represents the one-day
the estimation errors are below 10% (see Section IV-D). EV loads that are aggregated from a two-day ACN dataset,
Remark 8: Proposition 1 enables us to operate BMPC by as shown in Table II. To quantify the traffic of these aggregated
less number of stage optimizations for EV-DC problems. data, we adopted the following metric:
Suppose there are total of P (P ≥ 1) decoupling intervals
P P
t0 ∈Ta i∈{1,...,N } Di,t0
in an arrival trajectory ξ , indexed by t pd , p = 1, . . . , P. The Intensity = × 100%, (16)
RMT
first condition indicates that the schedules before and after any
t pd are only coupled by φt pd . The second condition gives the where Ta denotes the horizon of the aggregation level a.
optimal threshold on the maximum average consumption for We increased the average intensity (over 30 trajectories) of the
all the BMPC problems, i.e., φt pd = ψξ∗ , ∀ p. Therefore, when original traffic from 5% to 29%. In the subsequent sections,
all the simulations were conducted based on these aggregated
BMPC starts at the beginning of each t pd , it always produces
d one-day data with different intensities. Furthermore, to balance
the optimal solution over t pd to t p+1 − 1, thus the next BMPC
d the weight of the demand charge, we scaled down the monthly
problem can start directly at t p+1 but not necessarily t + 1. demand charge price9 to a single day and varied it from 0.2
Remark 9: To guarantee the optimality of BMPC when the
system has the setting ℓ > 1, A7 can be further relaxed by 7 The details of this facility can be found in [52] and the data is available

slightly modifying Algorithm 1, as demonstrated in Fig. 3. at http://ev.caltech.edu

8 We adopted Day-ahead Market (DAM) prices. Available at
Suppose BMPC is running at t from an optimal state x∗t and
http://www.ercot.com/mktinfo/dam
we take the decisions of the first block (the blue frame). 9 For simplity, we assume fixed monthly demand charge price that ranges
Clearly, {ut , . . . , ut pd −1 } produced by BMPC are the optimal from 6 $/kW to 18 $/kW [34].

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 2132 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 21, NO. 2, APRIL 2024

TABLE II larger gaps (near 40% at most) and showed better performance
AGGREGATION OF EV L OADS when the aggregation level increased. This was due to the less
flexibility of scheduling so that the demand charge tended to
be close to each other.
Another factor that affects the performance of BMPC is
the parameter M that limits the number of simultaneously
$/kW to 0.6 $/kW, where the length of a measurement window activated charging. According to (16), decreasing the value of
was still fixed at 15 minutes. M is equivalent to increasing the traffic intensity. In Fig. 4b,
we showed that the variation of M had the minimal influence
on the gap to optimality of BMPC.
B. On Gap to Optimality
Then we fixed the traffic intensity at 17% and demonstrated
For the optimal offline charging solution, an integer program the average performance gaps of BMPC and other methods
that defined the deterministic EV-DC was solved to compute with the variation of the demand charge prices, as shown in
the upper bound of the total reward by using realizations of Fig. 4b. EMPC [32] obtained the optimal reward over all the
the EV loads from each arrival trajectory ξ , as shown below: demand charge prices, while the results of BMPC were also
T −1
X close, showing only 1% gap when the demand charge price
J (ξ ) = max
∗
G(xt , ut , ξ t ) − C(ψξ ) reached 0.6 $/kW. The gap produced by MPC-scaled had a
{ut }t∈T
t=0 steady growth from 0 to 35% since its fixed weight of the
s.t. (1)(2)(3)(4). (17) demand charge could not accommodate high demand charge
prices. For the methods that do not consider demand charge,
By solving (17), we also obtained the optimal demand charge.
MPC-w/o and LLF-LD both experienced rapid increase as the
Furthermore, three related MPC approaches and one index
demand charge price was over 0.4 $/kW, hitting over 100%
rule were included as benchmarks in our comparison studies:
due to the large demand charge costs.
MPC with scaled demand charge (MPC-scaled), EMPC [32],
2) Multi-Resolution Case (ℓ = 3): In practice, the reso-
MPC without demand charge (MPC-w/o) and LLF-LD [27]
lution of control can be significantly finer than that of the
(see Appendix B). Then we ran each algorithm and computed
demand charge measurement. For example, the measurement
its total reward gap to the upper bound (in percentage) as
window size can be 15 minutes whereas the EV charging
the performance measure. For a given scenario, we simulated
decisions can be made at the one to five minute resolution,
all the methods over 30 trajectories and reported the average
i.e., ℓ = 3-15. The results presented in this section are from
performance.
simulations with ℓ = 3, i.e., 1 = 5 minutes.
Fig. 5a illustrated the performance of BMPC with the
C. Performance Under Perfect Forecast benchmarks over varied intensity of EV loads at a fixed
This section validates the performance of BMPC and other demand charge price 0.3 $/kW. Although Proposition 1 did
benchmarks under the ideal scenario, where we assume all the not hold when ℓ = 3, the average gap produced by BMPC
clients use reservation apps to offer accurate information on was close to the optimal at light traffic regime and less than
EV arrivals, charging demands and deadlines. 1% even the traffic level reached 29%-intensity. On the other
1) Single-Resolution Case (ℓ = 1): We first considered the hand, EMPC [32] showed sub-optimal solutions (between 2%
case when the resolution of measurement window matched and 10%) compared to the ℓ = 1 case, which resulted from the
that of the decision, i.e. ℓ = 1 (1 = 15 minutes). In this mismatching of the timescales of peak consumption from
case, BMPC operated at the same timescales as the other the reference trajectory (see (31) and (32) in Appendix B).
MPC-based benchmarks and the forecast window length was The performance of the rest of the methods, MPC-scaled,
set to W = 4 hours. MPC-w/o and LLF-LD, all had improvements when the EV
Fig. 4a showed the performance of BMPC with other loads became heavier. Due to the consideration of the demand
methods under different intensities of EV loads (x-axis) at charge, MPC-scaled outperformed MPC-w/o and LLF-LD by
a fixed demand charge price 0.3 $/kW, where the average 4% and 2% on average, respectively.
gaps to the upper bound over all the sampled trajectories The influence of the demand charge prices was also com-
(y-axis) were compared. The best method was EMPC [32], pared in Fig. 5b, where the traffic intensity was also fixed at
which actually reached the upper bound (0% gaps at all 17% as the previous section. BMPC was the best method to
traffic levels) because it tracked the optimal reference tra- handle the increase of the demand charge price, making only
jectory. BMPC also produced the optimal solutions before 1.5% gap at 0.6 $/kW. The performance of EMPC gradually
the intensity of the EV loads reached 17%, where all the deviated from the optimum compared to the ℓ = 1 case and
optimality conditions of BMPC could be met and therefore peaked at 6%. MPC-scaled had less than 2% performance gap
validated Proposition 1. Then slightly bigger gaps were shown when the demand charge price was below 0.2 $/kW, whereas
when the EV loads went more intensive, since the number it caused larger gaps (over 10%) as the demand charge was
of the decoupling intervals became less and the optimality more expensive. As for MPC-w/o and LLF-LD, they both
conditions could not be held at 17%-intensity and the above. showed a fast degradation on the total reward (at around 30%),
Consequently, a 0.6% gap rose up at 29%-intensity. The other resulting from the ignorance of the increasing demand charge
methods (MPC-scaled, MPC-w/o and LLF-LD) had much costs.

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: EV CHARGING SCHEDULING UNDER DEMAND CHARGE: A BLOCK MPC APPROACH 2133

Fig. 4. Average performance gap to the upper bound under perfect forecast Fig. 5. Average performance gap to the upper bound under perfect forecast
with the system setting ℓ = 1. The average computation time of the BMPC with the system setting ℓ = 3. The average computation time of the BMPC
problems is 0.66s. problems is 0.86s.

TABLE III
C ASES FOR S IMULATIONS U NDER F ORECAST E RRORS
D. Performance Under Forecast Errors

This section investigates the performance of BMPC and the

benchmarks under forecast errors. Here we mainly focuses
on the multi-resolution case (1 = 5 minutes), which fits the
practical settings but has been less addressed by the existing
works.
Although the reservation mechanism can be adopted, some
clients may not follow their reservations. For an EV i that Table IV reported the average performance of BMPC and
actually arrives at t0 with Di,t0 , we assume a probability the benchmarks under forecast errors δr = δa = δe = δm =
δr (0 < δr < 1) that the reservation information on this EV 0.1, where the reward and demand charge gap were both
is not accurate. If so, its arrival time and the energy demand compared. BMPC consistently obtained the largest reward over
reported on the reservation apps were uniformly sampled from all the cases (between 2.09% and 5.07% reward gap), and
[t0 (1−δa ), t0 (1+δa )] (0 < δa < 1) and [Di,t0 (1−δe ), Di,t0 (1+ it always captured the optimal demand charge even though
δe )] (0 < δe < 1), respectively. Meanwhile, the forecast on the starting from a smaller value. When the traffic was less intense
initial value φ0 for BMPC was set as (1−δm )ψξ∗ (0 < δm < 1). with a low demand charge price, i.e., Case I, MPC-scaled was
As shown in Table III, we set up four cases to demonstrate the second best method with 4.70% reward gap, whereas its
the influence of the forecast errors under different demand demand charge was identical to MPC-w/o and LLF-LD. This
intensities and demand charge prices. indicates that such method had more revenue from charging

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 2134 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 21, NO. 2, APRIL 2024

TABLE IV
AVERAGE P ERFORMANCE OF BMPC AND B ENCHMARKS U NDER 10% F ORECAST E RRORS (δr = δa = δe = δm = 0.1)

TABLE V
AVERAGE P ERFORMANCE OF BMPC AND B ENCHMARKS U NDER 20% F ORECAST E RRORS (δr = δa = δe = δm = 0.2)

the EVs. However, LLF-LD outperformed MPC-scaled by W/T on the demand charge was only able to handle low
around 1% when more demands were involved (see Case III) demand charge prices. EMPC [32], however, was dramatically
since more EVs would be reported inaccurately. When the affected by the increase of the forecast error in both reward and
demand charge price was high, i.e., Case II and Case IV, demand charge, causing nearly 20% reward gap in Case IV.
EMPC [32] was the second best method (6.69% and 12.51% Although MPC-w/o still produced the same demand charge as
reward gap respectively) with around 22% demand charge LLF-LD, their reward gap was getting larger to over 10% since
gap. It was found that the reward from EMPC [32] was more forecast errors were imposed to the forecast window.
more sensitive to the intensity levels rather than demand
charge prices, which might result from the predictions over the V. C ONCLUSION
entire trajectories (see Appendix C). It was also worth noting
We consider the problem of EV-DC to maximize the reward
that the demand charge was the major factor to determine
from charging services at a charging station. Due to the
the performance of MPC-w/o and LLF-LD, where MPC-w/o
difficulties of multiple timescales posed by the demand charge
showed the least performance in each case and could reach
pricing, we propose the BMPC algorithm with a special end-
over 18% reward gap.
of-horizon cost to incorporate the demand charge at each stage
Table V reported the performance of all the methods when
optimization, and the optimality conditions of the proposed
all the forecast errors increased to 0.2. Generally, the reward
method are presented and analyzed based on characterized
gaps from MPC-based methods became larger than the previ-
demand patterns. Through the simulations on the ACN testbed,
ous case, while LLF-LD was not affected since its performance
our proposed approach shows advantageous performances
did not depend on any predictions. BMPC continued to be the
compared to the benchmark methods, highlighting the signifi-
best method in all the cases except 1.33% more reward gap
cant impact of demand charge on the EV charging scheduling.
than LLF-LD in Case III. This would result from the fact
that BMPC took a block of controls based on the inaccurate
information for the near future. To be more specific, BMPC A PPENDIX A
committed to the charging actions {u∗t , u∗t+1 , · · · , u∗t+ℓ−1 } after P ROOF OF P ROPOSITION 1
solving (10) at t. The subsequent actions from u∗t , which Proof: We first show that the original EV-DC problem
were optimal for the predicted EV trajectory, would become (7) can be decomposed by t d . For a given trajectory ξ and x0 ,
sub-optimal for the realized EV profile. Such impact to lose We rewrite (7) in the deterministic form:
the schedule revenue would become more dominant when the T −1
forecast error was larger under heavier traffics. On the other
X
max Jξ (x0 ) = G(xt , ut , ξ t ) − C(ψξ )
hand, although BMPC did not follow the optimal demand {ut }t∈T
t=0
charge when φ0 was 20% less than the optimal peak, it pro- s.t. (1)(2)(3)(4) (18)
duced the closest gap between 7.69% and 25.00% among all
the benchmarks. We denote T d = {t pd , p = 1, . . . , P} as the set of the
Similarly to Table IV, MPC-scaled had good performance decoupling intervals over trajectory ξ . For each t pd where
with just 0.08% and 3.02% less reward than BMPC in Case I (xt pd , ξ t pd ) = (0, 0), the system state at t pd + 1 is not relevant
and Case III, respectively, and thus showed that the fix weight to the previous states, i.e., (1) is decoupled. Since the energy

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: EV CHARGING SCHEDULING UNDER DEMAND CHARGE: A BLOCK MPC APPROACH 2135

requests before t pd cannot be deferred to intervals after t pd + 1, SP(BMPC) − S ′ (0) = {t1d , . . . , W − 1}:
the EV-DC problem (7) can be decomposed into P + 1 seg-
ments: S(0) = {0, . . . , t1d −1}, S( p) = {t pd , . . . , t p+1
d
−1} (1 ≤ W
X −1

p < P), and S(P) = {t P , . . . , T − 1}. Let the optimal reward

d max Jξ ,S ′ (0) (x′t d , φt1d ) = G(x′k , u′k , ξ̂ k ) (22a)
u ∈S (0)
′ ′ 1

of (7) be Jξ∗ (x0 ) and the optimal solution u∗ , then we have: k=t1d

− C(φt+W − φt1d )
P−1
X s.t. (10b)(10c)(10d)(10e) (22b)
Jξ∗ (x0 ) =Jξ∗,S(0) (x0 ) + Jξ∗,S( p) (xt pd ) + Jξ∗,S(P) (xt Pd )
p=1 td td
   According to A8, {ξ̂ k }k=0
1 1
= {ξ t }t=0 , i.e., the scheduler knows
− C max ψξ ,S(0) , max ψξ ,S( p) , ψξ ,S(P) ,
∗ ∗ ∗ the accurate information to solve (21). The second optimality
1≤ p<P condition shows that φ0 = ψξ∗ , which indicates that φ0 ≥
(19) ψξ∗,S(0) . Without loss of generality, two cases are discussed.
1) ψξ∗ Happens in S(0): We have φt∗d = φ0 = ψξ∗ = ψξ∗,S(0)
where 1
since the scheduler would not necessarily rise the consumption
t1d −1 level beyond the optimal one. Otherwise, more penalty from
demand charge would reduce the total reward of the segment
X
Jξ∗,S(0) (x0 ) = G(xt , u∗t , ξ t ), (20a)
t=0 S(0). In other words, the optimal reward of (21) is the identical
d
t p+1 −1 to Jξ∗,S(0) (x0 ).
Jξ∗,S( p) (xt pd ) =
X
G(xt , u∗t , ξ t ), 1 ≤ p < P, (20b) 2) ψξ∗ Happens Out of S(0): It is obvious that φ0 >
ψξ ,S(0) . This provides a higher threshold on the maximum
∗
t=t pd
consumption of the segment S(0), and indicates that the last
T −1
X term of (21a) is not effective. Hence, the scheduler would
Jξ∗,S(P) (xt P ) = G(xt , u∗t , ξ t ), (20c) not take demand charge into consideration so that the optimal
t=t P
reward of (21) is also identical to the Jξ∗,S(0) (x0 ), where the
N
X latter also ignores the demand charge in the one-shot solution.
ψξ∗,S(0) = max ∗
u i,t , (20d)
t∈S(0) Consequently, for both cases shown above, we have the
i=1
N
optimal reward of (21) Jξ∗,S(0) (x′0 , φ0 ) = Jξ∗,S(0) (x0 ), and the
ψξ∗,S( p) = max
X
∗
u i,t , 1 ≤ p < P, (20e) optimal solution u′∗ 0 of (21) must be within the optimal set
t∈S( p)
i=1
U ∗ , i.e., u′∗
0 ∈U .
∗

N Then the system updates to x′1 and we have x′1 ∈ X ∗ and

φ1 = φ0 = ψξ∗ . Then we assume t1d > 1 (this would not
X
ψξ∗,S(P) = max ∗
u i,t . (20f)
t∈S(P)
i=1 make any difference if t1d = 1). Similarly to the above, the
new forecast window still contains t1d so that the new BMPC
where (20d), (20e) and (20f) are the results based on Assump- problem can also be decomposed into two sub-problems as
tion 1. Let ψξ∗ be the optimal maximum consumption of (7), follows:
which is clearly the maximum of (20d), (20e) and (20f). SP(BMPC) − S(0, 0) = S(0) \ {0}:
We also denote X ∗ and U ∗ as the optimal system state and
solution set of the problem (7), respectively. t1d −1
X
Then we show that for each segment above, the correspond- max Jξ ,S(0,0) (x 1 , φ1 ) = ′
G(x′k , u′k , ξ̂ k ) (23a)
u′ ∈S(0,0)
ing BMPC (10) reaches an optimal solution. Suppose BMPC k=1
operates at t = 0. The first optimality condition is equivalent − C(φt1d − φ1 )
to W ≥ max(t1d , maxk=1,...,P−1 (tk+1 − tk ), T − t Pd ), i.e., the s.t. (10b)(10c)(10d)(10e) (23b)
length of the forecast window is no shorter than the largest
gap of adjacent decoupling intervals. Thus, the current BMPC SP(BMPC) − S ′ (0): see (22)
can be decomposed into two sub-problems by t1d , as shown The last term of (23a) follows the same rule as dis-
below10 cussed above. Based on the Principle of Optimality, we have
SP(BMPC) − S(0): Jξ∗,S(0) (x′1 , φ1 ) the optimal reward from t = 1 to t1d − 1 and
thus u′∗ ∗
1 ∈ U . As BMPC rolls to the subsequent stage starting
t1d −1 d
X at t1 − 1, BMPC always produces the optimal solutions, thus
max
′
Jξ ,S(0) (x′0 , φ0 ) = G(x′k , u′k , ξ̂ k ) − C(φt1d − φ0 ) we have
u ∈S(0)
k=0
(21a) t1d −1
X
s.t. (10b)(10c)(10d)(10e) (21b) G(x′t , u′∗
t , ξ t ) = Jξ ,S(0) (x0 ),
∗
(24)
t=0
x′0 = x0 , (21c)
where the left hand side is the total reward produced by BMPC
10 Here to distinguish from (7), we use and x′ u′
to represent the system from t = 1 to t1d − 1. And then BMPC starts at t = t1d and
state and the decision variable used in all the BMPC problems, respectively. repeats as above until the last segment S(P). Consequently,

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 2136 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 21, NO. 2, APRIL 2024

we have At t ∈ T with the system state xt and φt , EMPC can be

d
t p+1 −1 formulated as
X
G(x′t , u′∗
t , ξ t ) = Jξ ,S( p) (xt pd ), ∀1 ≤ p < P,
∗
(25) max JtEMPC
t=t pd {uk }t+W
k=t
−1
(
T −1 xi,k − (u i,k , 1) if τi,k > 1,
s.t. xi,k+1 = , (33a)
X
G(x′t , u′∗
t , ξ t ) = Jξ ,S(P) (xt Pd ).
∗
(26) ξi,k
ref
if τi,k ≤ 1.
t=t Pd
N
X
and u i,k ≤ M, (33b)
i=1
φT = φT −1 = · · · = φ0 = ψξ∗ . (27) u i,k ≤ xi,k , i ∈ {1, . . . , N } (33c)
According to (19)(20) and (24)-(27), we have N
X
φk+1 = max(φk , u i,k ) (33d)
T −1
X i=1
JξBMPC (x′0 , φ0 ) = G(x′t , u′∗
t , ξt) − C(φT ) = Jξ∗ (x0 ), (28)
k = t, . . . , t + W − 1,
t=0

and thus BMPC produces an optimal solution. □ t+W ,

xt+W = xref (33e)
Note that the temporal structure is also the same as the nominal
A PPENDIX B MPC.
B ENCHMARKS 3) MPC Without Demand Charge (MPC-w/O): MPC-w/o
follows the framework of the nominal MPC, that moves
1) MPC With Scaled Demand Charge (MPC-Scaled):
only one interval at each time, i.e., solving the following
MPC-scaled imposes a fraction of the demand charge in each
optimization problem consisting of W intervals at every t ∈ T
rolling-window optimization, and such form is adopted in [31].
starting at current state xt :
At each t ∈ T , MPC-scaled solves the following problem:
t+W
X−1
t+W
X−1 max JtMPC-w/o := G(xk , uk , ξ̂ k )
max JtMPC-scaled := G(xk , uk , ξ̂ k )
{uk }t∈T k=t
k=t
N
!! s.t. (10b)(10c)(10d).
W X
− C max u i,k (34)
T k∈[t,t+W −1]
i=1
Note that MPC-w/o does not take demand charge into consid-
s.t. (10b)(10c)(10d). (29)
eration, which is another major difference from BMPC.
Unlike BMPC, only the optimal control u∗t will be imple- 4) LLF-LD [27]: LLF-LD is an online algorithm for
mented and the temporal structure of MPC-scaled is identical deferrable load scheduling, which prioritizes tasks with less
to the nominal MPC. laxity at each interval. Laxity, as defined in [27], is the
2) EMPC [32]: EMPC [32] aims to track a predetermined difference between a server’s lead time and its remaining
reference schedule with the consideration of demand charge processing time, reflecting the maximum number of intervals
for the setting that ℓ = 1. Here we present a slight modification that a task can tolerate before the time it has to be continuously
of EMPC [32] so that it applies to cases when ℓ > 1. processed to avoid non-completion penalty. For the EV-DC
Let (xref , uref , ξ ref ) be an arbitrarily known reference trajec- problem in Section II, the laxity of an EV at charger i at t
tory over T (see Appendix C for the reference trajectory). For is τi,t − ri,t . If the laxity of two tasks are the same, then it
each t ∈ T , the objective of EMPC [32] is defined as prioritizes the one with later deadline. Note that the available
total charging power at each interval for LLF-LD is always
t+W
X−1 M R. Based on the LLF-LD rule, the scheduler assigns each
JtEMPC := Gk (xk , uk , ξ ref
k )
unit of charging power to each unfinished EVs from the top
k=t
priority to the bottom (or until all the M R charging power has
− C(max(φt+W , ψ̆ ref
t+W )) − C(ψ ),
ref
(30)
been scheduled). If the number of unfinished EVs is greater
where parameters ψ ref and ψ̆ ref t+W denote the peak consumption
than M, then the scheduler would leave the rest to the next
over the entire (from 0 to T − 1) and the remaining horizon interval. Thus, LLF-LD could yield significant demand charge
(from t + W to T − 1) of the reference trajectory respectively, when the traffic is intensive.
and are computed as
( t+ℓ N ) A PPENDIX C
1 X X ref R EFERENCE T RAJECTORY OF EMPC [32]
ψ := max′
ref
u , (31)
t∈T ℓ τ =t i=1 i,τ The reference trajectory sets the states and decisions for the
EMPC to track. Here we consider to generate the reference
( k+ℓ−1 N )
1 X X ref
ψ̆ t := max
ref
u , t ∈T. (32) trajectory by solving the EV-DC problem (7) in one-shot, and
k∈T ′ ,k≥t ℓ τ =k i=1 i,τ
therefore the whole trajectory of all the EVs must be forecast,

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 YANG et al.: EV CHARGING SCHEDULING UNDER DEMAND CHARGE: A BLOCK MPC APPROACH 2137

i.e., from t = 0 to T − 1. We denote {ξ ref t }t∈T as the forecast [18] Q. Huang, Q. S. Jia, Z. Qiu, X. Guan, and G. Deconinck, “Matching EV
EV loads, then the reference trajectory can be computed by charging load with uncertain wind: A simulation-based policy improve-
ment approach,” IEEE Trans. Smart Grid, vol. 6, no. 3, pp. 1425–1433,
the following deterministic EV-DC problem: May 2015.
X [19] Q. Huang, Q.-S. Jia, and X. Guan, “A multi-timescale and bilevel
max Gt (xt , ut , ξ ref
t ) − C(ψ) coordination approach for matching uncertain wind supply with EV
{ut }t∈T charging demand,” IEEE Trans. Autom. Sci. Eng., vol. 14, no. 2,
t∈T
( pp. 694–704, Apr. 2016.
xi,t − (u i,t , 1) if τi,t > 1, (35) [20] Z. Jiang, Q. Jia, and X. Guan, “A computing budget allocation method
s.t. xi,t+1 =
ξi,t
ref
if τi,t ≤ 1, for minimizing EV charging cost using uncertain wind power,” IEEE
Trans. Autom. Sci. Eng., vol. 18, no. 2, pp. 681–692, Apr. 2021.
and (2)(3)(4), [21] Y. Xu, F. Pan, and L. Tong, “Dynamic scheduling for charging electric
vehicles: A priority rule,” IEEE Trans. Autom. Control, vol. 61, no. 12,
By denoting (xref , uref ) as the optimal solution of (35), the pp. 4094–4099, Dec. 2016.
[22] J. Jin and Y. Xu, “Joint scheduling of electric vehicle charging and
reference trajectory can be obtained as (xref , uref , ξ ref ). energy storage operation,” in Proc. IEEE Conf. Decis. Control (CDC),
Dec. 2018, pp. 4103–4109.
[23] Z. Yu, Y. Xu, and L. Tong, “Deadline scheduling as restless ban-
R EFERENCES dits,” IEEE Trans. Autom. Control, vol. 63, no. 8, pp. 2343–2358,
Aug. 2018.
[1] K. Clement-Nyns, E. Haesen, and J. Driesen, “The impact of charging
plug-in hybrid electric vehicles on a residential distribution grid,” IEEE [24] L. Hao and Y. Xu, “Index policies for stochastic deadline scheduling
Trans. Power Syst., vol. 25, no. 1, pp. 371–380, Feb. 2009. with time-varying processing rate limits,” in Proc. Amer. Control Conf.
(ACC), Jul. 2020, pp. 204–210.
[2] Z. Ma, D. S. Callaway, and I. A. Hiskens, “Decentralized charging
control of large populations of plug-in electric vehicles,” IEEE Trans. [25] J. Jin and Y. Xu, “Optimal policy characterization enhanced actor-critic
Control Syst. Technol., vol. 21, no. 1, pp. 67–78, Dec. 2012. approach for electric vehicle charging scheduling in a power distribution
network,” IEEE Trans. Smart Grid, vol. 12, no. 2, pp. 1416–1428,
[3] L. Gan, U. Topcu, and S. H. Low, “Optimal decentralized protocol
Mar. 2021.
for electric vehicle charging,” IEEE Trans. Power Syst., vol. 28, no. 2,
pp. 940–951, May 2013. [26] P. Whittle, “Restless bandits: Activity allocation in a changing world,”
J. Appl. Probab., vol. 25, pp. 287–298, Jan. 1988.
[4] J. Rivera, C. Goebel, and H.-A. Jacobsen, “Distributed convex optimiza-
tion for electric vehicle aggregators,” IEEE Trans. Smart Grid, vol. 8, [27] J. Jin, L. Hao, Y. Xu, J. Wu, and Q.-S. Jia, “Joint scheduling of deferrable
no. 4, pp. 1852–1863, Jul. 2017. demand and storage with random supply and processing rate limits,”
IEEE Trans. Autom. Control, vol. 66, no. 11, pp. 5506–5513, Nov. 2021.
[5] C. Le Floch, S. Bansal, C. J. Tomlin, S. J. Moura, and M. N. Zeilinger,
“Plug-and-play model predictive control for load shaping and voltage [28] M. Ellis, H. Durand, and P. D. Christofides, “A tutorial review of
control in smart grids,” IEEE Trans. Smart Grid, vol. 10, no. 3, economic model predictive control methods,” J. Process Control, vol. 24,
pp. 2334–2344, May 2019. no. 8, pp. 1156–1178, Aug. 2014.
[6] Y. Yang, Q.-S. Jia, X. Guan, X. Zhang, Z. Qiu, and G. Deconinck, [29] R. Halvgaard, N. K. Poulsen, H. Madsen, J. B. Jorgensen, F. Marra,
“Decentralized EV-based charging optimization with building inte- and D. E. M. Bondy, “Electric vehicle charge planning using economic
grated wind energy,” IEEE Trans. Autom. Sci. Eng., vol. 16, no. 3, model predictive control,” in Proc. IEEE Int. Electr. Vehicle Conf.,
pp. 1002–1017, Jul. 2019. Mar. 2012, pp. 1–6.
[7] K. Ojand and H. Dagdougui, “Q-learning-based model predictive control [30] J. Engel, T. Schmitt, T. Rodemann, and J. Adamy, “Hierarchical
for energy management in residential aggregator,” IEEE Trans. Autom. economic model predictive control approach for a building energy
Sci. Eng., vol. 19, no. 1, pp. 70–81, Jan. 2022. management system with scenario-driven EV charging,” IEEE Trans.
[8] Z. J. Lee et al., “Adaptive charging networks: A framework for smart Smart Grid, vol. 13, no. 4, pp. 3082–3093, Jul. 2022.
electric vehicle charging,” IEEE Trans. Smart Grid, vol. 12, no. 5, [31] R. Kumar, M. J. Wenzel, M. J. Ellis, M. N. ElBsat, K. H. Drees, and
pp. 4339–4350, Sep. 2021. V. M. Zavala, “A stochastic model predictive control framework for
[9] Z. J. Lee, S. Sharma, D. Johansson, and S. H. Low, “ACN-Sim: stationary battery systems,” IEEE Trans. Power Syst., vol. 33, no. 4,
An open-source simulator for data-driven electric vehicle charging pp. 4397–4406, Jul. 2018.
research,” IEEE Trans. Smart Grid, vol. 12, no. 6, pp. 5113–5123, [32] M. J. Risbeck and J. B. Rawlings, “Economic model predictive control
Nov. 2021. for time-varying cost and peak demand charge optimization,” IEEE
[10] C. L. Liu and J. W. Layland, “Scheduling algorithms for multipro- Trans. Autom. Control, vol. 65, no. 7, pp. 2957–2968, Jul. 2020.
gramming in a hard-real-time environment,” J. ACM, vol. 20, no. 1, [33] M. Dabbagh, B. Hamdaoui, A. Rayes, and M. Guizani, “Shaving data
pp. 46–61, 1973. center power demand peaks through energy storage and workload shift-
[11] A. K.-L. Mok, “Fundamental design problems of distributed systems for ing control,” IEEE Trans. Cloud Comput., vol. 7, no. 4, pp. 1095–1108,
the hard-real-time environment,” Ph.D. dissertation, Massachusetts Inst. Oct. 2019.
Technol., Cambridge, MA, USA, 1983. [34] J. Jin and Y. Xu, “Optimal storage operation under demand charge,”
[12] S. S. Panwar, D. Towsley, and J. K. Wolf, “Optimal scheduling policies IEEE Trans. Power Syst., vol. 32, no. 1, pp. 795–808, Jan. 2017.
for a class of queues with customer deadlines to the beginning of [35] S. Chen and L. Tong, “IEMS for large scale charging of electric
service,” J. ACM, vol. 35, no. 4, pp. 832–844, 1988. vehicles: Architecture and optimal online scheduling,” in Proc. IEEE
[13] S. Chen, L. Tong, and T. He, “Optimal deadline scheduling with 3rd Int. Conf. Smart Grid Commun. (SmartGridComm), Nov. 2012,
commitment,” in Proc. 49th Annu. Allerton Conf. Commun., Control, pp. 629–634.
Comput. (Allerton), Sep. 2011, pp. 111–118. [36] Z. Yu, S. Chen, and L. Tong, “An intelligent energy management system
[14] A. Subramanian, M. J. Garcia, D. S. Callaway, K. Poolla, and P. Varaiya, for large-scale charging of electric vehicles,” CSEE J. Power Energy
“Real-time scheduling of distributed resources,” IEEE Trans. Smart Syst., vol. 2, no. 1, pp. 47–53, Mar. 2016.
Grid, vol. 4, no. 4, pp. 2122–2130, Dec. 2013. [37] Y. Qin, Z. Bei, and M. Kezunovic, “Optimized operational cost reduction
[15] N. Chen, C. Kurniawan, Y. Nakahira, L. Chen, and S. H. Low, for an EV charging station integrated with battery energy storage and
“Smoothed least-laxity-first algorithm for electric vehicle charging: PV generation,” IEEE Trans. Smart Grid, vol. 10, no. 2, pp. 2096–2106,
Online decision and performance analysis with resource augmentation,” Mar. 2018.
IEEE Trans. Smart Grid, vol. 13, no. 3, pp. 2209–2217, May 2022. [38] M. S. Athulya, A. Visakh, and M. P. Selvan, “Electric vehicle recharge
[16] Y. Xu and F. Pan, “Scheduling for charging plug-in hybrid electric scheduling in a shopping mall charging station,” in Proc. 21st Nat. Power
vehicles,” in Proc. IEEE 51st IEEE Conf. Decis. Control (CDC), Syst. Conf. (NPSC), Dec. 2020, pp. 1–6.
Dec. 2012, pp. 2495–2501. [39] H. Zhang, Z. Hu, Z. Xu, and Y. Song, “Optimal planning of
[17] Z. Yu, Y. Xu, and L. Tong, “Large scale charging of electric vehicles: PEV charging station with single output multiple cables charging
A multi-armed bandit approach,” in Proc. 53rd Annu. Allerton Conf. spots,” IEEE Trans. Smart Grid, vol. 8, no. 5, pp. 2119–2128,
Commun., Control, Comput. (Allerton), Sep. 2015, pp. 389–395. Sep. 2017.

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.
 2138 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 21, NO. 2, APRIL 2024

[40] M. Yilmaz and P. T. Krein, “Review of battery charger topologies, Xinbo Geng (Member, IEEE) received the B.E.
charging power levels, and infrastructure for plug-in electric and hybrid degree in electrical engineering from Tsinghua Uni-
vehicles,” IEEE Trans. Power Electron., vol. 28, no. 5, pp. 2151–2169, versity, Beijing, China, in 2013, and the M.S. and
May 2013. Ph.D. degrees in electrical engineering from the
[41] L. Zhang, M. Yang, and Z. Zhao, “Game analysis of charging service Department of Electrical and Computer Engineering,
fee based on benefit of multi-party participants: A case study analysis Texas A&M University, College Station, TX, USA,
in China,” Sustain. Cities Soc., vol. 48, Jan. 2019, Art. no. 101528. in 2015 and 2019, respectively.
[42] E. Bitar and Y. Xu, “Deadline differentiated pricing of deferrable electric He was a Power System Design and Studies Intern
loads,” IEEE Trans. Smart Grid, vol. 8, no. 1, pp. 13–25, Jan. 2017. with the National Renewable Energy Laboratory in
[43] Y. Kim, J. Kwak, and S. Chong, “Dynamic pricing, scheduling, and 2016; a Data Analyst Intern with ISO New England
energy management for profit maximization in PHEV charging stations,” in 2017; a Visiting Student with the University
IEEE Trans. Veh. Technol., vol. 66, no. 2, pp. 1011–1026, Feb. 2017. of California at Berkeley and the Massachusetts Institute of Technology,
[44] H. Chen, Z. Hu, H. Luo, J. Qin, R. Rajagopal, and H. Zhang, “Design Cambridge, MA, USA, in 2018; a Visiting Student with the University
and planning of a multiple-charger multiple-port charging system for of Wisconsin–Madison, Madison, WI, USA, in 2019; and a Post-Doctoral
PEV charging station,” IEEE Trans. Smart Grid, vol. 10, no. 1, Researcher jointly affiliated with Cornell University, Ithaca, NY, USA, and
pp. 173–183, Jan. 2017. Texas A&M University from 2020 to 2021. His research interests include
[45] N. Machiels, N. Leemput, F. Geth, J. Van Roy, J. Büscher, and J. Driesen, decision-making under uncertainties.
“Design criteria for electric vehicle fast charge infrastructure based on
flemish mobility behavior,” IEEE Trans. Smart Grid, vol. 5, no. 1,
pp. 320–327, Jan. 2014.
[46] M. Rahmani-Andebili, “Scheduling deferrable appliances and energy
resources of a smart home applying multi-time scale stochastic model Xiaohong Guan (Life Fellow, IEEE) received the
predictive control,” Sustain. Cities Soc., vol. 32, pp. 338–347, Jul. 2017. B.S. and M.S. degrees in control engineering from
Tsinghua University, Beijing, China, in 1982 and
[47] Gurobi. Gurobi 9.5 Performance Benchmarks. Accessed: Nov. 1, 2022.
1985, respectively, and the Ph.D. degree in electrical
[Online]. Available: https://www.gurobi.com/features/gurobi-optimizer-
and systems engineering from the University of
delivers-unmatched-performance
Connecticut, Storrs, in 1993.
[48] CPLEX Optimization Studio 12.10 Performance Improvements. He was a Senior Consulting Engineer with Pacific
Accessed: Nov. 1, 2022. [Online]. Available: https://www.ibm.com/ Gas and Electric from 1993 to 1995. He visited the
products/ilog-cplex-optimization-studio/cplex-optimizer Division of Engineering and Applied Science, Har-
[49] W. Charytoniuk, M. S. Chen, P. Kotas, and P. V. Olinda, “Demand vard University, from 1999 to 2000. From 1985 to
forecasting in power distribution systems using nonparametric prob- 1988 and since 1995, he has been with Xi’an Jiao-
ability density estimation,” IEEE Trans. Power Syst., vol. 14, no. 4, tong University, Xi’an, China. He has been the Cheung Kong Professor
pp. 1200–1206, Nov. 1999. of Systems Engineering since 1999. He was the Director of the State Key
[50] A. K. Singh, S. Khatoon, M. Muazzam, and D. K. Chaturvedi, “Load Laboratory for Manufacturing Systems from 1999 to 2009 and the Dean of
forecasting techniques and methodologies: A review,” in Proc. 2nd Int. the School of Electronic and Information Engineering from 2008 to 2018.
Conf. Power, Control Embedded Syst., Dec. 2012, pp. 1–10. Since 2001, he has also been with the Center for Intelligent and Networked
[51] V. Jain, A. Sharma, and L. Subramanian, “Road traffic congestion in Systems, Tsinghua University, where was the Head of the Department of
the developing world,” in Proc. 2nd ACM Symp. Comput. Develop., Automation from 2003 to 2008. He has been the Dean of the Faculty of
Mar. 2012, pp. 1–10. Electronic and Information Engineering since 2019. His research interests
[52] Z. J. Lee, T. Li, and S. H. Low, “ACN-Data: Analysis and applications include economics and security of networked systems, including power and
of an open EV charging dataset,” in Proc. 10th ACM Int. Conf. Future energy systems, manufacturing systems, and cyber-physical systems. He is a
Energy Syst., Jun. 2019, pp. 139–149. member of the Chinese Academy of Sciences.

Lang Tong (Fellow, IEEE) received the B.E. degree

from Tsinghua University, Beijing, China, in 1985,
and the Ph.D. degree from the University of Notre
Dame, USA, in 1991.
He was a Post-Doctoral Research Affiliate with
Stanford University. He held visiting positions at
Lei Yang (Student Member, IEEE) received the B.E. Stanford University, the University of California at
degree from Zhejiang University, Hangzhou, China, Berkeley, the Delft University of Technology, and
in 2014, and the M.Sc. degree from the University the Chalmers University of Technology, Sweden.
of Bristol, Bristol, U.K., in 2016. He is currently He is currently the Irwin and Joan Jacobs Professor
pursuing the Ph.D. degree with the Department of Engineering and the Cornell Site Director of the
of Electronic and Information Engineering, Xi’an Power Systems Engineering Research Center (PSERC), Cornell University,
Jiaotong University, Xi’an, China. Ithaca, NY, USA. His research interests include energy and power systems,
He was a Visiting Student with Cornell University, electricity markets, smart grids, and the electrification of transportation
Ithaca, NY, USA, in 2020. His current research inter- systems. His expertise lies in data analytics, machine learning, optimization,
ests include energy and power systems, smart city, and market design. He received numerous publication awards from the IEEE
and the integration of transportation and information Signal Processing Society, the IEEE Communications Society, and the IEEE
systems. His expertise lies in Markov decision process, optimization, and Power and Energy Systems Society. He was the 2018 Fulbright Distinguished
model predictive control. Chair in Alternative Energy.

Authorized licensed use limited to: National Institute of Technology - Jamshedpur. Downloaded on March 08,2025 at 11:53:06 UTC from IEEE Xplore. Restrictions apply.