This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2017.2652485, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 1

Distribution Network Electric Vehicle Hosting

Capacity Maximization: A Chargeable Region
Optimization Model
Jian Zhao, Student Member, Jianhui Wang, Senior Member, Zhao Xu, Senior Member, Cheng Wang, Student
Member, Can Wan, Member, Chen Chen, Member, IEEE

r/R Index/set of reactive power support facilities, i.e., shunt

AbstractTo coordinate electric vehicle (EV) charging, the EV capacitors and automatic voltage regulators (AVRs).
aggregator (EVA) is usually assumed to obtain the privilege from
(j) Set of child nodes of DN node j.
EV owners (EVOs) to determine the EV charging profile, and
complex communication between EVA and EVOs is demanded, EV(j)/ Set of EV charging facilities/reactive power support
which poses difficulties for practical applications. In contrast, this RP(j) facilities under DN node j.
paper proposes the concept of an EV chargeable region to evaluate
the distribution network (DN) EV hosting capacity, i.e., how much B. Parameters
ch
EV charging demand can be accommodated in a DN, within which P k,t Charging rate per EV at k,t
the technical constraints of DN (e.g., voltage deviation) are tarr dep
EV arrival and departure time at k.
k , tk
guaranteed and EVOs charging requests are maximally ensured.
The optimization of the EV chargeable region is formulated as a DTk Total travel distance during the day at k.
two-stage robust optimization model with adjustable uncertainty ECS Per-mile energy consumption.
set. The EV chargeable region and DN decision variables are
ch EV charging efficiency.
optimized in the first stage and the feasibility in the real-time
CH,S
worst-case scenario is checked in the second stage, considering the P c,t,s Aggregated charging demand under at c,t in scenario
uncertainty of EV charging demand and DN active and reactive s
power. A modified column & constraint generation and outer PCH,AV Average aggregated EV charging demand at c,t.
c,t
approximation method is adopted to address the proposed EV,max
problem. Simulations on an IEEE 123-node DN demonstrate the P c,t Maximum EV charging demand at c,t.
effectiveness of the proposed model. pj,t/qj,t Active/reactive power load at j,t.
Index TermsElectric vehicle, hosting capacity, chargeable
rij/xij Resistance/reactance of DN line ij.
region, charging strategy, distribution network, two-stage
optimization, robust optimization, adjustable uncertainty set. V0 Voltage reference value.
VST Voltage at DN substation.
NOMENCLATURE
r Step value of shunt capacitor r.
A. Indices and Sets
QAVR,AV
r,t Average output of AVR at r,t.
t/T Index/set of time slots.
LCij Capacity of DN line ij.
K/K Index/set of the trips whose destination is home
Voltage fluctuation limitation
Nc,t Set of EV charging demand in all scenarios
t Electricity price at t.
/ Index/set of piecewise linearization approximation
CH,LB
method, where DE, AP, RP represent set for PLA P c,t Lower bound of uncertain EV charging demand at c,t.
method of EV delayed charging energy, quadratic term of /
CH
t
CH
c Uncertainty budget of EV charging demand at t/at c.
active and reactive power respectively
pLB UB
j,t /pj,t Lower/upper bound of uncertain active power at j,t.
j/J Index/set of DN nodes.
AP AP
t /c Uncertainty budget of active power at t/at c.
c/C Index/set of EV charging facilities.
qLB UB
j,t /qj,t Lower/upper bound of uncertain reactive power at j,t.

This work was partially supported by Hong Kong RGC Theme Based
RP
t /c
RP
Uncertainty budget of reactive power at t/at c.
AVR,LB
Research Scheme Grants No. T23-407/13N and T23-701/14N. J. Wangs work Qr Lower/upper bound of AVR output of r.
is supported by the Natural Science Foundation (Award number: 1638348) and /QAVR,UB
r

the U.S. Department of Energy (DOE)'s Office of Electricity Delivery and C. Variables
Energy Reliability.
J. Zhao and Z. Xu are with the Department of Electrical Engineering, The DEc,t Delayed EV charging demand at c,t.
Hong Kong Polytechnic University, Hung Hom, Hong Kong. (e-mails: CCc,t Compensation cost for delayed EV charging demand at
zhaojianzju@gmail.com, eezhaoxu@polyu.edu.hk). c,t.
J. Wang and C. Chen are with Argonne National Laboratory, Argonne, IL
60439, USA (e-mails: jianhui.wang@anl.gov, morningchen@anl.gov). PCH
c,t Average EV charging demand within chargeable bound
C. Wang is with the State Key Laboratory of Power Systems, Department of c,t.
Electrical Engineering and Applied Electronic Technology, Tsinghua
University, 100084 Beijing, China. (e-mail: c-w12@mails.tsinghua.edu.cn).
PBound
c,t EV chargeable bound at c,t.
C. Wan is with the College of Electrical Engineering, Zhejiang University, c,t Actual EV charging demand at c,t.
Hangzhou 310027, China (e-mail: can.wan@ieee.org).

0885-8950 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2017.2652485, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Pj,t/Qj,t Active/reactive power flow at j,t.
PQU QU
Quadratic term of active/reactive power flow at j,t.
j,t /Qj,t
Vj,t Voltage at j,t.
SC
d r,t Integral variables for discrete shunt capacitors at r,t.
VPCH
c,t Real-time uncertain EV charging demand at c,t.
Vpj,t/Vqj,t Real-time uncertain active/reactive power at j,t.
QAVR
r,t AVR output at r,t.
PWj,t /QWj,t Real-time worst-case active/reactive power flow at j,t.
QU,W
P j,t / Quadratic term of real-time worst-case active/reactive
Q QU,W
power flow at j,t.
j,t
VWj,t Real-time worst-case voltage at j,t.
V PQ
S /S
j,t j,t Slack variables of real-time voltage deviation
requirement and DN line capacity limitation at j,t.
I. INTRODUCTION
R ECENT years have seen a constantly increasing
penetration of electric vehicle (EV) charging demand in
distribution networks (DNs) because of the growing concerns
on global warming and environmental pollution issues. A
review of [1-3] indicates that the large-scale random EV
charging demand will introduce significant impacts on the
secure and economical operation of DNs. These negative
influences mainly include increasing distribution energy losses,
voltage deviation, and overload of DN lines and substations.
Voltage deviation is one of the major issues caused by
large-scale EV charging demand. For example, case study in [1]
demonstrates that the voltage of residential distribution grid will
drop very likely below 0.95 p.u. when the uncontrolled EV
charging demand increase.
Coordinated/smart charging control of EVs is the widely
accepted method to improve voltage profile in distribution
network [4-11]. Many researches in open literatures
demonstrate that significant voltage drop in DN can be
prevented using optimal EV flexible charging strategy by
properly rescheduling the EV charging demand [4-6]. Various
implementation algorithms and models of coordinated/smart
charging strategies are proposed to deal with voltage issues. A
market mechanism is proposed in [7] to optimally allocate
available charging capacity ensuring both DN voltage security
and EVOs preferences on charging rates. A market based
multi-agent control mechanism is proposed in [8] to use
remaining capacity of each EV charger for reactive voltage
control where the iterative exchange of messages can be omitted.
In [9], a real-time smart load management control strategy is
proposed and developed for the coordination of EV charging
based on real-time (e.g., every 5 min) minimization of total
operation cost while complying with network operation criteria.
In [10], a rolling multi-period optimization based EV charging
control method is proposed to deter excessive voltage
deviations and overloading of equipment with the input data
updated at each time step. A local control technique for EV
coordinated charging strategy in low voltage DN is proposed in
[11] and the advantage and disadvantage of local and
centralized EV charging strategy are discussed.
0885-8950 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2
Moreover, as the large-scale EVs cannot be scheduled and
managed directly by the power system operator, the concept of
EV aggregator (EVA) is proposed to manage a huge number of
EVs and act as an intermediary to communicate with power
system operators to ensure DN operation security and to
improve social welfare [12-14]. For example, a hierarchical
decomposition model is proposed in [13] to minimize the total
cost of dispatching generators and EVAs in the upper-level
model and to design detailed charging and discharging
strategies in the lower-level model. However, in these methods,
the EVA is usually assumed to obtain the privilege from EVOs
to determine the charging schedule of each EV as long as the
customers travel demand is met. This implies that EVOs should
report their daily departure time and required energy in the
battery to the EVA, which will create a huge inconvenience for
end-user customers. Another issue is that needs for unexpected
and urgent usage of EVs may not be satisfied, as this
information cannot be captured in advance by the EVA. Besides,
it still remains uncertain whether complex communication
facilities will be built up among the DNO, EVAs, and EVOs in
the DNs.
To address these issues, we first assume that EVs are charged
randomly according to users requests. Then we try to maximize
EV hosting capacity, by determining the largest admissible
charging demand at each DN node that can be accommodated.
Within the hosting capacity, the DN voltage and line capacity
will not vary beyond the requirement, and outside of the hosting
capacity, some EV charging demand should be postponed. The
main contributions of this paper are threefold:
1) The concept of EV chargeable region is innovatively
proposed to evaluate the largest amount of EV charging
demand under a DN node that will not lead to network
constraint violations. The EV chargeable region
optimization problem is formulated as a two-stage model,
where the EV chargeable region and DN decision variables
are optimized in the first stage and the feasibility of the DN
worst-case scenario is checked in the second stage. This
model not only guarantees the secure operation of DN, but
also maximizes the EV hosting capacity for DN. Compared
with the previous EVA-oriented coordinated charging
strategy method, the proposed model is user-friendly and
enforceable for practical application.
2) To model the EV charging demand uncertainties, the
Monte Carlo simulation method was usually used in the
previous work to generate user travel behaviour data from
either the travel statistical results in travel survey reports or a
self-defined distribution function of travel parameters [13,
15-19]. In comparison, we directly use sampling with the
replication method to model EV uncertainties from raw
vehicle travel data, which can better resemble the reality.
Then the piecewise linearization approximation (PLA)
method is used to estimate out-of-region delayed EV
charging energy on the basis of sampled EV uncertainties.
3) The framework is mathematically formulated as a robust
optimization problem with an adjustable uncertainty set. In
contrast with the typical robust optimization problem, the
uncertainty set in the framework is affected by the first-stage
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2017.2652485, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 3

variables. Thus, a modified column & constraint generation ,t , s

PcCH ,S

k N c ,s
Pkch,t , c C , t T , s S (2a)
(C&CG) and outer approximation (OA) method is proposed
to solve the optimization problem. 1
The remainder of this paper is organized as follows: Section
PcCH
,t
, AV

N c' ,t
P CH , S
c ,t , s , c C,t T (2b)
sN c' ,t
II describes the modeling of EV charging demand. Section III
where PCH,S
c,t,s denotes the aggregated charging demand under a
describes the formulation of the EV chargeable region
DN node c in scenario s and PCH,AV
c,t denotes its average value,
optimization model. The compact formulation, dual problem,
where || denotes the Cardinality of the set. When the sample
and algorithm are given in Section IV. Case studies to verify the
size S is large enough, the uncertainty of charging demand can
proposed model are conducted, analyzed, and discussed in
be properly modeled and the estimated cumulative distribution
Section V. Finally, Section VI draws some useful conclusions.
function (CDF) can be obtained from the statistical results.
II. MODELING ELECTRIC VEHICLE CHARGING DEMAND
B. Modeling Delayed Electric Vehicle Charging Energy
A. Modeling Electric Vehicle Charging Demand In this subsection, we illustrate the relationship between the
Uncertainty compensation cost of delayed charging energy and the EV
It is widely accepted that time-varying and location-varying chargeable region, based on the statistical results of EV
EV charging demand can be modeled according to EVOs travel charging demand uncertainties. As discussed, when the
behavior. Monte Carlo simulation method was mostly used to charging demand exceeds the upper bound of the chargeable
region, the excessive portion of the energy should be delayed, as
acquire EVOs travel behavior data on the basis of travel survey
follows:
reports, from which the parking duration and the charging PcEV ,max

DEc ,t ( c ,t PcBound DEc ,t 1 )g ( c ,t )d c ,t , c C , t T (3a)

,t

demand of each vehicle can be estimated [13, 15, 16]. To model PcBound
,t DEc ,t 1 ,t

more detailed spatial-temporal dynamics of EVs charging

g ( c ,t ) KPDF
, c , t c ,t B , c ,t ,
PDF DE
, c C , t T (3b)
demand, some other researchers randomly generated a daily trip
chain based on self-defined hypothetical distribution functions where (3a) describes the relationship between delayed charging
[17-19]. energy DEc,t and EV chargeable bound PBound c,t . c,t denotes the
In contrast, the present paper models the uncertainty of actual EV charging demand, and g(c,t) denotes the probability
charging behavior by sampling directly from the 2011 Raw Data distribution function (PDF) ofc,t. As formulated in (3b), the
of Travel Behavior released by the Atlanta Regional PDF is obtained using the PLA method from statistical results,
Commission (ARC), which can be found in the ARC rather than the exact distribution function, where KDE DE
,c,t and B,c,t

Metropolitan Travel Survey Archive [20]. This survey dataset are the auxiliary coefficients of linearization approximation.
contains 119,480 trips in total, collected within 3 months. For The PDF is actually the linearly estimated gradient of its CDF,
acquired from the sampling results in subsection IIA. It should
each trip, the information provided by the data includes:
be mentioned that the delayed energy is affected not only by the
departure time, departure location, arrival time, arrival location,
chargeable bound but also by the charging demand increase due
trip distance, transit access mode, etc. The large number of trips
to the delayed energy of the last time interval, as shown in (3a).
and the detailed and comprehensive record of each trip provide Thus, the delayed EV charging energy is calculated as the
valuable assistance in modeling the uncertainty of EV charging cumulative energy beyond the bound and within the maximum
demand. In this paper, all the EVs are first assumed to be EV charging demand PEV,max . However, (3) is still difficult to
c,t
charged upon arriving home, where the charging load for each solve because of the nonlinearity and the integral term ofc,t.
EV can be formulated as To cope with this problem, the PLA method is used again to
DTk ECS obtain the approximate linear term of (3). Thus the EV charging
Pkch,t Pch , if t tkarr & t tkarr & t tkdep , k K ,t T (1)
ch Pch demand can be modeled as:
where and are the index and set of the trips whose CCc ,t KDE , c , t ,
DEc ,t 1 ) BDE , c C , t T
Bound
, c , t ( Pc , t
DE
(4a)
destination is home, which will trigger the charging reaction. Pch , t Pc , t
PcCH CH , AV
DEc ,t 1 DEc ,t , c C , t T (4b)
denotes the fixed charging rate, which can be randomly selected
DEc ,t R , CCc ,t R , P CH
c ,t R ,P Bound
c ,t R , c C , t T (4c)
from the predefined charging power dataset with corresponding
probability, and is the index of variously rated charging where (4a) are the linearization approximate versions of
constraints (3). They are auxiliary constraints induced by the
power. DT is the total travel distance during the day, ECS is the
PLA, where K DE ,c,t and B DE
,c,t are the constant coefficients of
per-mile energy consumption, ch is the charging efficiency, linearization approximation. Also, (4b) describes the actual
and tarr
k and tdep
k are the EV arrival and departure time. charging demand considering the delayed charging energy. The
To model the aggregated charging demand under a DN node boundaries of the variables are quantified in (4c).
with Nk total EVs, sampling with the replication method is Some remarks on the EV chargeable region follow:
adopted to randomly select Nk samples from the EV charging 1) Although an EVA is not considered, an EV controller
demand database K. The summarized value can be regarded as should be installed at the DN node to interrupt the excessive
one scenario of charging demand at each DN node. By repeating charging demand beyond the chargeable region.
this procedure continually, various scenarios can be obtained, 2) The EV charging delay priority may depend on the EV
given as charging urgent level, e.g., the remained EV charging

0885-8950 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2017.2652485, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 4

demand. The EVs with larger SOC of the battery are terms of transmitting voltage is the same as in the previous work,
interrupted with higher priority, which guarantees the urgent while the voltage in the denominator of the voltage drop terms is
usage of EVs to the maximum extent. replaced with its reference value, which should be acceptable
3) Although immediate charge is modeled, EVOs can and more accurate than neglecting the whole term in linear
schedule their charging profiles on the basis of their DistFlow [21-26]. The voltage of the substation is given as VST,
intentions. The case of charging at the lowest electricity defined by (6d). The quadratic terms PQU QU
j,t Qj,t can be estimated
price is considered and discussed in the case study. using the PLA method with the following auxiliary constraints:
, t K , j , t Pj , t B , j , t , j J ,
PjQU , t T
AP AP AP
(7a)
III. FORMULATION OF ELECTRIC VEHICLE CHARGEABLE
j , t K , j , t Q j , t B , j , t , j J , , t T
QQU RP RP RP
REGION OPTIMIZATION MODEL (7b)
AP RP AP RP
where K K and B B are the constant coefficients of
,c,t ,c,t ,c,t ,c,t
A. Distribution Network Operation Constraints linearization approximation. The operation limitation
The complex DN power flow at each node j can be described constraints are given as
using DistFlow equations from [21, 22]:
, t Q j , t LCij , j J , i ( j ), t T
PjQU QU 2
(8a)
Pi ,2t Qi2,t
Pj ,t

i ( j )
Pi ,t

r
i ( j )
ij
Vi ,2t
p j ,t , j J , t T (5a) 1 V j ,t 1 , j J , t T (8b)
where (8a) describes the DN branch active and reactive power
Pi ,2t Qi2,t
Q j ,t Q
i ( j )
i ,t
i ( j )
xij
Vi ,2t
q j ,t , j J , t T (5b) limitation and (8b) describes the voltage fluctuation limitation.
B. Two-stage Electric Vehicle Chargeable Region
Pi ,2t Qi2,t Optimization Framework
V j2,t Vi ,2t 2(rij Pi ,t xij Qi ,t ) (rij2 xij2 ) , j J , i ( j ), t T (5c)
Vi ,2t
In this subsection, a two-stage EV chargeable region
where (5a) describes the active power flow, (5b) describes the optimization framework is proposed. To keep the DN voltage
reactive power flow, and (5c) describes the voltage transmit deviation within the requirement and to minimize the DN line
along the branch. To address the problem of nonlinearity, the loss, the coordinated operation of shunt capacitors and AVRs is
linear version of the DistFlow equations is proposed and widely used. In practice, shunt capacitors cannot react
justified in [21, 23] by Baran and Wu and later adopted by continually and immediately according to the rapid fluctuation
various researchers [24-26]. The approximation is based on two of demand and voltage in the DN, while the AVRs can respond
assumptions. First, the nonlinear terms representing the loss quickly to the real-time DN status. Also, the EV chargeable
should be much smaller than the branch power Pj, Qj and region is assumed to be optimized on a day-ahead basis before
voltage terms V2j , so that they can be neglected in the calculation the uncertainty is revealed, so that real-time interaction and
of power flow. Second, the approximation of (Vj V0)2 = 0 is complex communication between the EVA and EVOs are
adopted and it will be valid as long as the voltage violation is avoided.
always within the requirement so that the quadratic terms of In this regard, a two-stage robust optimization framework is
voltage can be replaced. proposed to determine the optimal EV chargeable region and
DN decision variables, hedging against any possible realization
In contrast with the previous work involving a linear version
of uncertainty. The uncertainty of the DN consists of EV
of Distflow, we propose the piecewise linearized Distflow,
charging demand and other active/reactive power loads.
achieved by linearizing the quadratic terms of active power and
Therein, shunt capacitors, the chargeable region, and expected
reactive power so that the loss terms can be maintained. The set DN operation variables act as the first-stage decision variables,
of power flow, taking into consideration linearization, EV served as here-and-now, which cannot be adjusted after the
charging demand, and reactive power facilities, can be uncertainty is revealed, while the AVR and real-time power
characterized by the following constraints: flow act as the second-stage decision variables, served as
Pi ,QU QiQU wait-and-see, which will respond to the uncertainties. It
Pj ,t

P
i ( j )
i ,t

r
i ( j )
ij
t

V02
,t
p j ,t
c EV ( j )
,t , j J , t T
PcCH (6a)
should be noticed that both distributed generators and other
schedulable loads in distribution network can be modeled as
, t Qi , t
Pi QU QU

Q j ,t

Q
i ( j )
i ,t

i ( j )
xij
V02
q j ,t second-stage wait-and-see variables and thus affect EV
chargeable region. However, they are not taken into

r RP ( j )
(d rSC,t rV j ,t QrAVR
,t
, AV
), j J , t T (6b) consideration, as this paper focuses on the EV charging demand
itself. Thus, the average operation cost is factored into the
rij Pi ,t xij Qi ,t Pi ,QU QiQU
V j ,t Vi ,t (rij2 xij2 ) t ,t
, j J , i ( j ), t T (6c) first-stage decision-making and the feasibility in the real-time
V0 2V03
worst-case scenario should be checked during the second-stage
V j ,t VST , j 1, t T (6d) decision-making, with the following formulation:
,t Q j ,t
PjQU QU

min (t rij CCc ,t )

where constraints (6a)(6c) are adjusted from constraints (9a)
(5a)(5c), where P QU QU
j,t Q j,t are used to estimate the quadratic t T ijL V02 cC

terms of Pj,t Qj,t and the EV charging demand and the reactive s.t. (4a)(4c), (6a)(6d), (7a)(7b), (8a)(8b)
power output of shunt capacitors and AVRs are taken into ( max
M1 US ( PcBound
min
) M 2 WS ( d rSC
S
,t ) t T jJ
V
j ,t S PQ
j ,t ) 0 (9b)
account in the power flow. The disposition of the quadratic ,t

0885-8950 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2017.2652485, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 5

The objective function (9a) is to minimize total DN operation Pi ,QU ,W

QiQU ,W

cost, in which the first term denotes the total DN line loss and
PjW,t

i ( j )
PiW,t

r
i ( j )
ij
t

V02
,t
Vp j ,t

the second term denotes the compensation cost paid to the

EVOs because of the charging delay. Constraints of power flow,

c EV ( j )
, t , j J , t T
VPcCH (11a)

active/reactive power quadratic term linearization, and the Pi ,QU ,W

QiQU ,W

operation limitation are defined in (4), (6), and (7), respectively, QWj ,t

Q
i ( j )
W
i ,t

i ( j )
xij t

V02
,t
Vq j ,t
in subsection IIIA above. The relationship among delayed EV
charging demand, EV chargeable bound, and the compensation

r RP ( j )
(d rSC,t rV jW,t QrAVR
, t ), j J , t T (11b)
cost is defined in (8). Constraint (9b) is used to check the rij PiW,t xij QiW,t Pi QU ,W
QiQU ,W

feasibility of the real-time worst-case scenario, given the EV V jW,t ViW,t (rij2 xij2 ) ,t ,t
,
V0 2V03
chargeable region and the operation status of shunt capacitors,
j J , i ( j ), t T (11c)
where SVj,t and SPQ
j,t denote the slack variables of voltage deviation

constraints and DN line capacity violation constraints. The DN V W

j ,t VST , j 1, t T (11d)
operation remains secure only if all the slack variables are zero P QU ,W
j ,t KAP, j ,t PjW,t BAP, j ,t , j J , AP , t T (11e)
in any scenario. The uncertainty set US() of constraint (9b) is Q QU ,W
K RP
Q BW RP
, j J , , t T
RP
(11f)
j ,t , j ,t j ,t , j ,t
given as follows:
P QU ,W
Q QU ,W
S PQ
LC , j J , i ( j ), t T
2
(11g)
US ( PcBound
,t ) { M1 : M1 [VPcCH
, t ,Vp j , t ,Vq j , t ]
j ,t j ,t j ,t ij

1 V S , j J , t T
W V
(11h)
, t Pc , t
VPcCH , c C , t T
Bound
(10a) j ,t j ,t

Q AVR , LB
Q AVR
Q AVR ,UB
, r R, t T (11i)
, t Pc , t
VPcCH , c C , t T
CH , LB r r ,t r
(10b)

P QU ,W
,Q QU ,W
,S ,S V PQ
R , j J , t T (11j)}
VP
cC
CH
c ,t tCH , t T (10c) j ,t j ,t j ,t j ,t

where power flow in the real-time worst-case scenario is

VPcCH,t CH
t T
c , c C (10d) illustrated by (11a)(11d). Compared with the first-stage power
flow in (6), EV charging demand, active power, and reactive
Vp j ,t p LB
j , t , j J , t T (10e) power become variables which are quantified in uncertainty set
Vp j ,t p , j J , t T
UB
j ,t (10f) (10) and the AVR becomes the real-time controllable variables
with the boundary given in (11i). The first-stage variable dSC
Vp
j J
j ,t , t T
AP
t (10g) r,t is

fixed in the real-time worst-case feasibility check. Quadratic

Vp
t T
j ,t AP
j , jJ (10h) terms of active and reactive power flow in the real-time
worst-case scenario are linearized in (11e)(11f). The
Vq j ,t q LB
j , t , j J , t T (10i) constraints of voltage deviation limitation and DN power flow
Vq j ,t q , j J , t T
UB
(10j) capacity limitation in the worst-case scenario are relaxed with
j ,t
variables S Vj,t and S PQ
j,t , as explained by (11g)(11h). The
Vq
j J
j ,t , t T
RP
t (10k)
boundaries of some variables are quantified by (11j). It should
be noticed that the feasible region defined by constraints (9)
Vq
t T
j ,t RP
j , jJ (10l)}
(10) guarantees the secure and reliable DN operation.
where the uncertainty set US() consists of three kinds of IV. SOLUTION METHOD
variables, i.e., EV charging demand, active power, and reactive
power. As neither the EV charging demand nor the active power A. Compact Formulation and Duality
and reactive power at each DN node can be precisely forecast on For simplicity, the compact formulation is written in this
a day-ahead basis, they are regarded as variables in the subsection as follows:
second-stage optimization so as to find the real-time worst-case min a T x1 + b T x 3 (12a)
x1 ,x 2 ,x 3
scenario. Constraints (10a)(10d) denote the uncertainty set of
s.t. Ax1 + Bx 2 + Cx 3 c (12b)
EV charging demand, (10e)(10h) denote the uncertainty set of
active power and constraints, and (10i)(10l) denote the ( max min d s) 0
T
(12c)
z1 , z 2 US ( x2 ) y,z1 ,z 2 ,sWS ( x1 )

uncertainty set of reactive power. In each set, the boundary of US (x 2 ) { z1 , z 2 : Dz1 + Ex 2 e, Fz 2 f } (12d)
the variables and the uncertainty budget at each time interval or WS (x1 ) { y, z1 , z 2 , s : Gx1 + Hy + Iz 1 + Jz 2 + Ks g } (12e)
at each DN node are given respectively. It should be noticed that
By the characteristics, the first-stage variables are divided
the uncertainty setUS() is affected by the first-stage variable,
into three groups; i.e., x1 [d rSC,t ] represents the first-stage DN
EV chargeable bound PBound c,t . The worst-case feasibility set
decision variable that affects second-stage DN operation status,
WS() of constraint (9b) is given as follows:
x 2 [ PcBound ] represents the one that affects the uncertainty set,
WS (d ) { M 2 :
SC
r ,t
,t

and x3 [ Pj ,t , Q j ,t , PjQU QU CH
, t , Q j , t ,V j , t , DEc , t , CCc , t , Pc , t ] represents other
M 2 [ PjW,t , QWj ,t , PjQU
,t
,W
, QQU
j ,t
,W
,V jW,t , QrAVR CH V PQ
, t ,VPc , t ,Vp j ,t ,Vq j ,t , S j ,t , S j ,t ]
first-stage variables. On the basis of whether or not they are
affected by first-stage variables, the variables in the uncertainty

0885-8950 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2017.2652485, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 6

, t ] and z 2 [Vp j , t ,Vq j , t ] , where the

set are divided into z1 [VPcCH step 1.
boundary of z1 is quantified by x 2 . The second-stage slack Initialization

variables are represented as s [S , S

V
j ,t
PQ
j ,t ] and the
Solve Master
second-stage operation variables are aggregated as Problem
y [ Pj ,t , Q j ,t , Pj ,t , Q j ,t ,V j ,t , QrAVR
W W QU ,W QU ,W W
,t ] . Subproblem
The optimization problem (12) can be divided into two stages, Solve OA subproblem
where the first-stage master problem can be described as Generate C&CG
(12a)(12b) and the second-stage subproblem can be described Solve OA master constraints
problem
as follows:
max min d Ts (13a)
z1 , z 2 US ( x 2 ) y,z1 ,z 2 ,sWS ( x1 ) UBIN LBIN OA
s.t. (12d)(12e) (13b)
Thus, (13) is formulated as a bi-level linear program. In
UBOUT LBOUT CCG
accordance with the strong duality of the linear program, the
inner optimization problem is replaced by its dual problem so
End
that (13a) can be reformulated as a single-level bilinear program,
Fig. 1. The framework of the algorithm.
as follows:
max T (g - Gx1 ) T (-Iz1 - Jz 2 ) (14a) The inner-level OA algorithm is given as,
z1 ,z 2 ,

DL { : HT 0, K T d, R + } , z1 , z 2 US (x 2 ) , (14b) Step 0: Initialization. Fix the first-stage decision variables

where is the dual variable vector of the inner problem of (13). x1* x*2 . Set the inner-level iteration index j 1 , lower bound
It should be noticed that the second term of (14a) is the bilinear LBIN , upper bound UBIN . Find an initial
term, as both and z1 z2 are variables. z1,* j , z *2, j .
The single-level bilinear program (14) can be solved by
either the OA method [27] or the big-M linearization method. Step 1: Solve the OA subproblem,
Although the big-M method can find the exact optimal value, S (x1* , z1,* j , z*2, j ) max T (g - Gx1* - Iz1,* j - Jz*2, j ) (16a)

the computational burden is closely dependent on the number of s.t. DL (16b)
bilinear terms. As the uncertainty set in the proposed model is
very large, the big-M method becomes low-efficient in Let be the optimal solution. Set LBIN = S (x , z , z ) .
* *
1
*
1, j
*
2, j

computation, or even intractable for addressing some cases. Step 2: Linearize the bilinear term Iz Jz at
Thus, the OA method is adopted to cope with the bilinear
(z1,* j , z*2, j , *j ) , as follows:
problem.
L j z1 z 2 *T -Iz1* - Jz*2 - Iz Jz
B. Algorithm
*T I z z J z - z (17)
To address the two-stage optimization problem with an
adjustable uncertainty set, a two-level algorithm is proposed in Step 3: Solve the OA master problem. Solve the linearized
this subsection. The outer level employs the C&CG method [28] version of the second-stage problem, defined as follows:
to obtain x1, x2, x3 with the results of inner-level optimization, U z1,* j , z*2, j , *j ) max T (g - Gx1 ) + (18a)
z1 ,z 2 , ,
and the inner level uses an OA algorithm [27] to cope with the s.t. L i (z1 , z 2 , ), i 1,..., j (18b)
bilinear problem.
The outer-level C&CG algorithm is given as, z1 , z 2 US (x ) DL
*
2 (18c)
Step 0: Initialization. Set outer-level iteration index k 1 , Let (z *
1, j 1 ,z *
2, j 1 , *
j 1 , j 1 ) be the optimal solution. Set UBIN
lower bound LBOUT , upper bound UBOUT . Find a = U(z *
,z *
, *
).
1, j 1 2, j 1 j 1
feasible solution (z1,* k 1 , z*2, k 1 , x k ) . Step 4: Check inner-level convergence. If
Step 1: Solve the master problem (12a)(12b), taking into UBIN LBIN OA , output the current solution. Otherwise,
consideration the following constraints, where wm is the set j = j+1, go to Step 1.
newly created variable vector at each iteration: It should be mentioned that the C&CG constraints
CCG (z*1, m , z*2, m , x*2, m ) { w m , x1 , x2 : additionally added at each iteration are originally formulated as
Gx1 + Hw m + Iz1*, m + Jz*2, m g , m 1,..., k (19)
Gx1 + Hw m + (Iz1*, m / x*2, m )x2 + Jz*2, m g }, m 1,..., k (15)
* * *
However, the first-stage variable x1 is not included in the
Let (x , x1, k 2, k , x ) be the optimal solution. Set LBOUT =
3, k constraints (19) and thus cannot be adjusted at each iteration.
aT x1,* k + b T x*3, k UBOUT = aT x1,* k 1 + b T x*3, k 1 . This is because the x1, representing the EV chargeable bound,
affects the boundary of the uncertainty set rather than the
Step 2: Solve the bilinear subproblem (14). Let (z*1,k , z*2,k )
worst-case scenario decision variables, which is different from
be the optimal solution. Check the outer-level convergence. If the typical robust optimization. To cope with the problem, the
UBOUT LBOUT <CCG, stop. Otherwise, let k = k+1, go to

0885-8950 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2017.2652485, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 7

auxiliary term x 2 / x*2,m is multiplied with the third term, and Case 2: Charging response to price; EV penetration is
16.82%.
finally reformulated as (15). The general framework of the
Case 3: Charging response to price; EV penetration is
algorithm is described as Fig. 1.
20.18%.
1

0.8

Probability
0.6

0.4
CDF at t = 20
0.2 CDF at t = 17

0
20 40 60 80 100 120 140 160
EV charging demand [kW]

Fig. 3. CDF of EV charging demand at two typical periods.

0.035
PDF at t = 20
0.03 PDF at t = 17

Fig. 2. Modified IEEE 123-node distribution network. 0.025

Probability
0.02
V. CASE STUDY 0.015

In this section, the experiments are presented to illustrate the 0.01

proposed model using the modified IEEE 123 node DN in this 0.005
section. All the algorithms are implemented on MATLAB,
0
using Cplex as the MILP solver. The optimality gap is set as 20 40 60 80 100 120 140 160
EV charging demand [kW]
10-3.
Fig. 4. PDF of EV charging demand at two typical periods.
A. Modified IEEE 123-node Distribution Network
B. Electric Vehicle Charging Demand Uncertainty
In the modified 123-node DN system, 12 DN nodes are
The sample size of nodal EV charging demand is set as 4000
connected with EV charging facilities and 6 DN nodes are
to effectively model the uncertainty of the total EV charging
connected with voltage regulator facilities, as shown in Fig. 2.
demand under each DN node at each time interval. Then the
The average DN peak load without consideration of the EV
estimated CDF of aggregated EV charging demand is obtained
charging demand is set as 2.84 MW/1.53 MVAR. The average
accordingly. Fig. 3 depicts the CDFs at two typical time
DN load at each of the 24 one-hour time intervals, compared
intervals. Results show that EV charging demand at t = 20 is
with the peak load, is given as 0.64 0.645 0.635 0.64 0.705 0.81
larger than that at t = 17. Then the PDF is obtained from the
0.86 0.83 0.85 0.87 0.905 0.83 0.81 0.85 0.89 0.96 1 0.99 0.94
gradient of CDF data, as depicted in Fig. 4. It should be
0.91 0.87 0.8 0.72 0.66. The upper and lower bounds of each
mentioned that the PDF and the CDF are described with discrete
DN active and reactive power are set as 1.2 and 0.8 times the
data rather than the exact distribution function, such that a
average value, respectively. Without loss of generalities, the
piecewise linearized function can be used to model the EV
uncertainty budgets AP t j t j
AP RP RP
are set as 1.05 times the
charging demand uncertainty.
average value. The voltage deviation limitation is set as [0.95
1.05]. The voltage at the substation node is set as 1.0125. The C. Optimal Electric Vehicle Chargeable Region
DN line loss fee is set as 0.1$/kWh. This subsection gives the simulation results of case 1. Two
At each EV node, 50 EVs are assumed to arrive during one typical charging bounds are depicted together with the various
day. The EV charging power rate is set as 4 kW, 6 kW and 8 kW uncertainty intervals of EV charging demand in Fig. 5. The EV
with a probability of 25%, 50% and 25%, respectively. The chargeable region is defined as the area lower than the
compensation cost of the EV charging delay is set as chargeable bound. It can be found that the EV charging demand
0.1$/(kWh*h). The EV penetration level is defined as the ratio is totally met during most times of the day, when both the DN
of total EV charging demand to total DN demand during the day, load and the EV demand are low. However, the EV charging
with the formulation demand may be larger than the chargeable region and thus
P CH , AV
c ,t should be delayed during the period hour 16 to hour 22. By
EVP t T cC
(20) comparing the difference between the chargeable bounds of
p P
t T jJ
j ,t
t T cC
CH , AV
c ,t
node 6 and node 114, it can be concluded that the location of EV
To effectively illustrate the proposed model, we propose charging demand in the DN area has a considerable impact on
three cases with different charging modes and different EV how much EV demand can be charged. Generally, voltage drop
penetration levels, as follows: issues become more serious at the end of the DN, where the EV
Case 1: Immediate charge; EV penetration is 16.82%. chargeable region is significantly affected and narrowed.

0885-8950 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2017.2652485, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 8

200 99.9% 98% 96% 94%

prevent the DN violation, especially if the EV charging demand

EV chargeable bound [kW]

90% 80% 70% 60%
is large. The proposed EV chargeable region method becomes
more advantageous with the increase of EV penetration level in
Node 6 the DN.
Node 114 TABLE I
SIMULATION RESULTS OF DIFFERENT CASES
EV compensation
Case Total cost ($) Chargeability
cost ($)
1 515.51 30.28 0.884
2 448.35 0.573 0.982
0
T[h] 3 492.60 4.568 0.707
6 12 18 24
Fig. 5. Optimal EV chargeable region in case 1. TABLE II
230 SIMULATION RESULTS AT EACH ITERATION OF CASE 1
99.9% 98% 96% 94%
Total cost Loss cost compensation Objective of
EV chargeable bound [kW]

90% 80% 70% 60% Iteration

($) ($) cost ($) SP ($)
Node 6
1 458.93 458.93 0 3.81*105
Node 114 2 513.31 482.57 30.74 5.99*103
3 515.51 485.23 30.28 0.556
TABLE III
COMPUTATION EFFICIENCY OF CASE 1
Total (s) MP (s) SP (s) Iteration
0 141 40 101 3
6 T[h] 12 18 24
Fig. 6. Optimal EV chargeable region in case 2. E. Optimization Procedure and Computation Efficiency
In this subsection, the optimization procedure and
D. Impact of EVOs Charging Response to Electricity Price
computational efficiency are discussed to verify the proposed
It is well known that EVOs can rearrange their charging solution algorithm. The simulation results at each iteration of
profile by responding to electricity price. To discuss the case 1 are listed in Table II, where the final solution is given
effectiveness of the EV chargeable region in this demand with fold type. It is observed that the objective value of
response scenario, two assumptions are first proposed: 1) half of subproblem is significantly reduced in each iteration and the
the EVOs respond to the electricity price by charging their EVs algorithm terminates in the third iteration when the objective
at the lowest electricity price; 2) the electricity price has a value of subproblem meets the convergence criterion. The
negative correlation with total DN demand. In this regard, EV computation efficiency is listed in Table III. It can be observed
charging demand is remodeled by sampling from the travel that the subproblem contributes more to the total computation
dataset with half of the charging demand shifting to the valley time. The reason is that it takes several iterations to achieve the
period of DN demand. convergence of outer approximation method for the bilinear
Fig. 6 depicts the EV charging demand uncertainty intervals subproblem.
and the chargeable region of case 2 at two typical DN node. The
results demonstrate that more EV charging demand can be F. Sensitivity Analysis of Piecewise Linearization
covered by the EV chargeable region compared with case 1 in Approximation Method
Fig. 5. It is well known that the demand response of EV In this subsection, sensitivity analysis of the PLA method is
charging demand helps to shift part of the DN load to the valley conducted to check the accuracy of the linearization
period, which may help to maintain the voltage profile and approximation. The simulation results with different segment
relieve DN congestion during peak period. Nevertheless, Fig. 6 numbers of piecewise linearized delayed EV charging demand
demonstrates that the network constraint violation limitation are listed in Table IV, where the simulation results of case 1 are
cannot necessarily be guaranteed without the proposed EV given with fold type. It can be observed that the EV
chargeable region, as EV charging demand is still random and compensation cost and EV chargeability are affected by the
uncontrolled. Especially if the EV penetration level in DN is number of segment of PLA model. It can be also noticed that the
increased, the DN will become more heavily loaded and the EV differences between simulation results of k=13 and k=10 are
chargeable region will take on a more important role in very small and the differences among simulation results of k=10,
maintaining DN secure operation. To illustrate this point, Table k=7 and k=4 are relatively larger, indicating the parameters in
I gives the simulation results for the three cases. The concept of the PLA method is proper for this case. Similarly, the simulation
EV chargeability in the DN is proposed to evaluate the general results with different segment numbers of piecewise linearized
level of the EV chargeable region in various DN nodes, with its Distflow are listed in Table V. It can be observed that the
definition given as differences among various costs become smaller when the
CB min( PcBound
,t / PcCH
,t
, AV
) (21) number of segment k increases. Thus, the PLA accuracy in
c ,t
modelling Distflow can be regarded within the acceptable
By comparing the simulation results of cases 2 and 3, it can be
range.
concluded that EVOs response to electricity price cannot TABLE IV

0885-8950 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2017.2652485, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) < 9

SIMULATION RESULTS WITH DIFFERENT SEGMENT NUMBER OF Mathematically, the proposed framework is formulated as a
PIECEWISE LINEARIZED DELAYED EV CHARGING DEMAND two-stage robust optimization problem with an adjustable
* Total cost Loss cost compensation uncertainty set. A modified C&CG/OA method is employed to
n k* Chargeability
($) ($) cost ($) solve the two-level problem. Case studies demonstrate the
13 515.88 486.48 29.40 0.8841 effectiveness of the proposed model in both immediate charging
10 10 515.51 485.23 30.28 0.8848
mode and charging demand response mode, considering various
7 526.71 487.14 39.57 0.8863
4 533.98 487.85 46.13 0.8894 EV penetration levels.
* k denotes segment number of the piecewise linearized delayed EV With the aid of the proposed model, not only operating
charging demand. constraint violation of DN is prevented, but also EVOs
* n denotes segment number of the piecewise linearized Distflow charging requests, i.e., immediate charging or price-response
TABLE V charging, are guaranteed to the largest extent. EVOs daily
SIMULATION RESULTS WITH DIFFERENT SEGMENT NUMBER OF report of charging demand to the EVA may be waived, as the
PIECEWISE LINEARIZED DISTFLOW EV charging profile can be well managed directly by the EVOs
Total cost Loss cost compensation themselves. Thus the urgent usage of EVs can be maximally
k n Chargeability
($) ($) cost ($)
guaranteed. Besides, communication mechanism of the
16 495.31 467.54 27.77 0.8848
13 503.68 474.99 28.69 0.8848
proposed framework is simple and the only communication is to
10 10 515.51 485.23 30.28 0.8848 pass message of EV chargeable region from distribution
7 552.58 519.65 32.93 0.8848 network operator to controller at distribution network node,
4 --* -- -- -- which is unidirectional and happens once a day. Overall, the
* -- denotes infeasibility of optimization model. proposed framework demonstrates a high potential for practical
300 applications.
1.1

250
REFERENCES
EV compensation cost [$]

1.0
[1] K. Clement-Nyns, E. Haesen, and J. Driesen, "The Impact of Charging
0.9 200 Plug-In Hybrid Electric Vehicles on a Residential Distribution Grid," IEEE
Chargeability

Transactions on Power Systems, vol. 25, pp. 371-380, 2010.

0.8 EV chargeability
EV compensation cost
0.7
0.6
0.5
0.4
0.12 0.14 0.16 0.18 0.20
EV penetration level
Fig. 7. Impact of EV penetration level on EV chargeability and compensation
cost.
G. Impact of EV Penetration Level
In order to address the prospective uncertainty of the EV
population in the DN, a sensitivity analysis of the EV
penetration level with EV immediate charging mode is
performed. Fig. 7 illustrates the impact of EV penetration level
on the EV compensation cost and the chargeability. The results
show that with increased EV penetration level, the EV
chargeability shows a slow decrease at first and then a rapid
decline as more EVs interact with the DN. This is because larger
EV charging demand may lead to DN line overload. Under such
circumstance, EV charging demand delay is the only solution to
avoid congestion. DN line capacity expansion planning can be
suggested if the chargeability is too small to be accepted by
EVOs.
VI. CONCLUSIONS
This paper innovatively proposes the concept of EV
chargeable region to evaluate the maximum DN EV hosting
capacity for each node. The uncertainty of EV charging demand
is modeled by sampling from raw vehicle travel data. The EV
chargeable region optimization problem is formulated as a
two-stage model where the chargeable region and DN decision
variables are optimized in the first stage and the feasibility of the
DN worst-case scenario is checked in the second stage.
150 [2] L. P. Fernandez, T. G. S. Roman, R. Cossent, C. M. Domingo, and P. Frias,
"Assessment of the Impact of Plug-in Electric Vehicles on Distribution
100 Networks," IEEE Transactions on Power Systems, vol. 26, pp. 206-213,
2011.
50 [3] S. Shafiee, M. Fotuhi-Firuzabad, and M. Rastegar, "Investigating the
Impacts of Plug-in Hybrid Electric Vehicles on Power Distribution
0
Systems," IEEE Transactions on Smart Grid, vol. 4, pp. 1351-1360, 2013.
[4] P. Richardson, D. Flynn, and A. Keane, "Optimal Charging of Electric
Vehicles in Low-Voltage Distribution Systems," IEEE Transactions on
Power Systems, vol. 27, pp. 268-279, 2012.
[5] O. Sundstrom and C. Binding, "Flexible Charging Optimization for Electric
Vehicles Considering Distribution Grid Constraints," IEEE Transactions
on Smart Grid, vol. 3, pp. 26-37, 2012.
[6] J. d. Hoog, T. Alpcan, M. Brazil, D. A. Thomas, and I. Mareels, "Optimal
Charging of Electric Vehicles Taking Distribution Network Constraints
Into Account," IEEE Transactions on Power Systems, vol. 30, pp. 365-375,
2015.
[7] J. d. Hoog, T. Alpcan, M. Brazil, D. A. Thomas, and I. Mareels, "A Market
Mechanism for Electric Vehicle Charging Under Network Constraints,"
IEEE Transactions on Smart Grid, vol. 7, pp. 827-836, 2016.
[8] S. Weckx, D. R, x, Hulst, B. Claessens, and J. Driesensam, "Multiagent
Charging of Electric Vehicles Respecting Distribution Transformer
Loading and Voltage Limits," IEEE Transactions on Smart Grid, vol. 5, pp.
2857-2867, 2014.
[9] S. Deilami, A. S. Masoum, P. S. Moses, and M. A. S. Masoum, "Real-Time
Coordination of Plug-In Electric Vehicle Charging in Smart Grids to
Minimize Power Losses and Improve Voltage Profile," IEEE Transactions
on Smart Grid, vol. 2, pp. 456-467, 2011.
[10] A. O. Connell, D. Flynn, and A. Keane, "Rolling Multi-Period Optimization
to Control Electric Vehicle Charging in Distribution Networks," IEEE
Transactions on Power Systems, vol. 29, pp. 340-348, 2014.
[11] P. Richardson, D. Flynn, and A. Keane, "Local Versus Centralized
Charging Strategies for Electric Vehicles in Low Voltage Distribution
Systems," IEEE Transactions on Smart Grid, vol. 3, pp. 1020-1028, 2012.
[12] D. Wu, D. C. Aliprantis, and L. Ying, "Load Scheduling and Dispatch for
Aggregators of Plug-In Electric Vehicles," IEEE Transactions on Smart
Grid, vol. 3, pp. 368-376, 2012.
[13] W. Yao, J. Zhao, F. Wen, Y. Xue, and G. Ledwich, "A Hierarchical
Decomposition Approach for Coordinated Dispatch of Plug-in Electric
Vehicles," IEEE Transactions on Power Systems, vol. 28, pp. 2768-2778,
2013.

0885-8950 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TPWRS.2017.2652485, IEEE
Transactions on Power Systems
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
[14] M. Liu, Y. Shi, and H. Gao, "Aggregation and Charging Control of PHEVs
in Smart Grid: A Cyber Physical Perspective," Proceedings of the IEEE,
vol. 104, pp. 1071-1085, 2016.
[15] D. Wu, D. C. Aliprantis, and K. Gkritza, "Electric Energy and Power
Consumption by Light-Duty Plug-In Electric Vehicles," IEEE Transactions
on Power Systems, vol. 26, pp. 738-746, 2011.
[16] A. Lojowska, D. Kurowicka, G. Papaefthymiou, and L. v. d. Sluis,
"Stochastic Modeling of Power Demand Due to EVs Using Copula," IEEE
Transactions on Power Systems, vol. 27, pp. 1960-1968, 2012.
[17] S. Bae and A. Kwasinski, "Spatial and Temporal Model of Electric Vehicle
Charging Demand," IEEE Transactions on Smart Grid, vol. 3, pp. 394-403,
2012.
[18] G. Wang, J. Zhao, F. Wen, Y. Xue, and G. Ledwich, "Dispatch Strategy of
PHEVs to Mitigate Selected Patterns of Seasonally Varying Outputs From
Renewable Generation," IEEE Transactions on Smart Grid, vol. 6, pp.
627-639, 2015.
[19] D. Tang and P. Wang, "Probabilistic Modeling of Nodal Charging Demand
Based on Spatial-Temporal Dynamics of Moving Electric Vehicles," IEEE
Transactions on Smart Grid, vol. 7, pp. 627-636, 2016.
[20] Metropolitan Travel Survey Archive [Online]. Available:
http://www.surveyarchive.org/
[21] M. Baran and F. F. Wu, "Optimal sizing of capacitors placed on a radial
distribution system," IEEE Transactions on Power Delivery, vol. 4, pp.
735-743, 1989.
[22] M. E. Baran and F. F. Wu, "Optimal capacitor placement on radial
distribution systems," IEEE Transactions on Power Delivery, vol. 4, pp.
725-734, 1989.
[23] M. E. Baran and F. F. Wu, "Network reconfiguration in distribution systems
for loss reduction and load balancing," IEEE Transactions on Power
Delivery, vol. 4, pp. 1401-1407, 1989.
[24] K. Turitsyn, P. Sulc, S. Backhaus, and M. Chertkov, "Options for Control of
Reactive Power by Distributed Photovoltaic Generators," Proceedings of
the IEEE, vol. 99, pp. 1063-1073, 2011.
[25] H. G. Yeh, D. F. Gayme, and S. H. Low, "Adaptive VAR Control for
Distribution Circuits With Photovoltaic Generators," IEEE Transactions on
Power Systems, vol. 27, pp. 1656-1663, 2012.
[26] C. Chen, J. Wang, F. Qiu, and D. Zhao, "Resilient Distribution System by
Microgrids Formation After Natural Disasters," IEEE Transactions on
Smart Grid, vol. 7, pp. 958-966, 2016.
[27] M. A. Duran and I. E. Grossmann, "An outer-approximation algorithm for a
class of mixed-integer nonlinear programs," Mathematical programming,
vol. 36, pp. 307-339, 1986.
[28] B. Zeng and L. Zhao, "Solving two-stage robust optimization problems
using a column-and-constraint generation method," Operations Research
Letters, vol. 41, pp. 457-461, 2013.
Jian Zhao (S15) received the B.Eng. degree from
Zhejiang University, Hangzhou, China, in 2013. He is
currently working toward the Ph.D. degree in the
Department of Electrical Engineering, The Hong Kong
Polytechnic University, Hong Kong. He was a visiting
scholar at Argonne National Laboratory, Argonne, IL,
USA. He was a recipient of the 2016 IEEE PES Best
Paper Reward.
His research interests include distribution network
operation and planning, grid integration of electric vehicles and renewable
energies.
Jianhui Wang (M07-SM12) received the Ph.D.
degree in electrical engineering from Illinois Institute of
Technology, Chicago, IL, USA, in 2007.
Presently, he is the Section Lead for Advanced Power
Grid Modeling at the Energy Systems Division at
Argonne National Laboratory, Argonne, IL, USA. He is
also a Fellow of the Computation Institute at The
University of Chicago (Jianhui@uchicago.edu).
Dr. Wang is the secretary of the IEEE Power &
Energy Society (PES) Power System Operations, Planning & Economics
Committee. He is an associate editor of Journal of Energy Engineering and an
editorial board member of Applied Energy. He is also an affiliate professor at
Auburn University and an adjunct professor at University of Notre Dame. He
has held visiting positions in Europe, Australia and Hong Kong including a
VELUX Visiting Professorship at the Technical University of Denmark (DTU).
0885-8950 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
10
Dr. Wang is the Editor-in-Chief of the IEEE Transactions on Smart Grid and an
IEEE PES Distinguished Lecturer. He is also the recipient of the IEEE PES
Power System Operation Committee Prize Paper Award in 2015.
Zhao Xu (M06-SM12) received his B.Eng., M.Eng,
and Ph.D. degrees from Zhejiang University, China, in
1996, National University of Singapore, Singapore, in
2002, and The University of Queensland, Australia, in
2006, respectively.
He is now with The Hong Kong Polytechnic
University. He was previously with Centre for Electric
Power and Energy, Technical University of Denmark.
His research interest includes demand side, grid
integration of renewable energies and EVs, electricity market planning and
management, and AI applications in power engineering. He is an Editor of
IEEE TRANSACTIONS ON SMART GRID, IEEE POWER ENGINEERING LETTERS,
and Electric Power Components and Systems journal.
Cheng Wang (S16) received the B.Sc. degree in
electrical engineering from Tsinghua University,
Beijing, China, in 2012, where he is pursuing the Ph.D.
degree. He was a Visiting Ph.D. student at Argonne
National Laboratory, Argonne, IL, USA, from 2015 to
2016.
His research interests include operation and
economic analysis of integrated energy systems.
Can Wan (M15) received his B.Eng. and Ph.D.
degrees from Zhejiang University, China, in 2008, and
The Hong Kong Polytechnic University in 2015,
respectively.
He is a Hundred Talents Program Professor with
the College of Electrical Engineering, Zhejiang
University, Hangzhou, China. He was a Postdoc Fellow
at Department of Electrical Engineering, Tsinghua
University, Beijing, China, and a Research Associate at
Department of Electrical Engineering, The Hong Kong Polytechnic University.
He was a visiting scholar at the Center for Electric Power and Energy,
Technical University of Denmark, and Argonne National Laboratory, IL, USA.
His research interests include forecasting, power system uncertainty analysis
and operation, renewable energy, active distribution network, demand side, and
machine learning.
Chen Chen (M13) received the B.S. and M.S. degrees
in electrical engineering from Xian Jiaotong University,
Xian, China, in 2006 and 2009, respectively, and the
Ph.D. degree in electrical engineering from Lehigh
University, Bethlehem, PA, USA, in 2013. During
2013-2015, he worked as a Postdoctoral Researcher at
the Energy Systems Division, Argonne National
Laboratory, Argonne, IL, USA. Dr. Chen is currently a
Computational Engineer with the Energy Systems
Division at Argonne National Laboratory. His primary
research is in optimization, communications and signal processing for smart
electric power systems, cyber-physical system modeling for smart grids, and
power system resilience. He is an editor of IEEE Transactions on Smart Grid
and IEEE Power Engineering Letters.