Computers & Industrial Engineering 85 (2015) 306–315

Contents lists available at ScienceDirect

Computers & Industrial Engineering

journal homepage: www.elsevier.com/locate/caie

Electric vehicle charging under power and balance constraints as

dynamic scheduling q,qq
Alejandro Hernández-Arauzo a, Jorge Puente a, Ramiro Varela a,⇑, Javier Sedano b
a
University of Oviedo, Department of Computing, West Building Department, 1, Campus of Gijón, 33204 Gijón, Spain
b
Instituto Tecnológico de Castilla y León (ITCL), Calle López Bravo 70, Polg. Ind., Villalonquejar, 09001 Burgos, Spain

a r t i c l e i n f o a b s t r a c t

Article history: Scheduling appropriately the charging times for a set of electric vehicles may lead to energy savings but
Received 5 March 2014 at the same time it may be a hard problem. In this paper, we consider a problem of this family, which is
Received in revised form 18 November 2014 motivated by a private community park where each parking space belongs to a particular user and has a
Accepted 2 April 2015
charging point connected to one line of a three-phase electric feeder. The number of vehicles in each line
Available online 13 April 2015
and the difference in the vehicles in every pair of lines charging at the same time are limited. We model
this problem in the framework of Dynamic Constraint Satisfaction Problems (DCSP) with optimization,
Keywords:
and so it is defined by a sequence of CSPs over time. We propose a solution method that decomposes each
Scheduling electric vehicles charging
Dynamic Constraint Satisfaction Problems
CSP into three instances of a one machine sequencing problem with variable capacity. This method was
Power and balance constraints evaluated by simulation on a set of instances defined from different scenarios of vehicle arrivals, depar-
One machine sequencing tures and energy requirements. The results of the experimental study show clearly that the proposed
Tardiness minimization algorithm is efficient and that it outperforms a classic dispatching rule.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction In this paper we consider a real life scheduling problem moti-

It is well known that the use of Electric Vehicles (EVs) may have
a positive impact on the economy and the environment due to pro-
moting the use of alternative sources of energy and relieving the
dependency of petrol. Furthermore, the energy stored in EVs may
be utilized as an ancillary service resource (Kang, Duncan, &
Mavris, 2013) for regulating frequency and voltage profiles as well
as to compensate fluctuations in renewable energy generation
(Dallinger, 2014).
At the same time, the emerging fleet of EVs introduces some
inconveniences such as the additional load on the power system
or the time required to charge the batteries. As pointed in (EDSO,
2012), one of the challenges in EVs’ technology is developing smart
systems for charging control to avoid increasing peak demand. In
fact, a number of charging control systems have already been
proposed (Gan, Topcu, & Low, 2007; Wu, Aliprantis, & Ying,
2012; Ma, Callaway, & Hiskens, 2013) that, in some cases, try to
fill the overnight valley in order to reduce daily cycling and
operational cost of power plants.
q
Some preliminary results of this research work has been discussed in
(Hernández-Arauzo, Puente, González, Varela, & Sedano, 2013).
qq
This manuscript was processed by Area Editor T.C. Edwin Cheng.
⇑ Corresponding author.
E-mail addresses: alex@uniovi.es (A. Hernández-Arauzo), puente@uniovi.es
(J. Puente), ramiro@uniovi.es (R. Varela), javier.sedano@itcl.es (J. Sedano).
vated by the charging station described in (Sedano et al., 2013).
This system was designed to be installed in a community park
under simplicity and economy criteria. Each stall belongs to a par-
ticular user and has a charging point connected to one of the lines
of a three-phase electric feeder. A centralized control establishes
the available power in each charging point over time. There are
power constraints limiting the number of vehicles that can be
charging at the same time on the same line and balance constraints
that limit the difference in the number of vehicles charging in
every pair of lines. Due to these technological constraints and to
the dynamic nature of the problem derived from the fact that the
arrival of the vehicles is not known in advance, it is not easy to
organize the charging periods of the vehicles over time in such a
way that all customers are satisfied and the system makes the
best use of the contracted power at the same time. Therefore, a
scheduling algorithm is a key element of the control system for
the charging station performance.
The main goal of this paper is the design and analysis of an effi-
cient and effective algorithm for the electric vehicle scheduling
problem in a charging station with the particular characteristics
mentioned above. To this end, we model the problem in the frame-
work of Dynamic Constraint Satisfaction Problems (DCSP) with
optimization. Therefore, it requires solving a number of static
CSPs over time. We propose a solution method that decomposes
each CSP into a number of instances of a one machine scheduling

http://dx.doi.org/10.1016/j.cie.2015.04.002
0360-8352/Ó 2015 Elsevier Ltd. All rights reserved.
 A. Hernández-Arauzo et al. / Computers & Industrial Engineering 85 (2015) 306–315
problem with variable machine capacity, which as far as we known
is the first time it appears in the literature. The scheduling algo-
rithm is evaluated by simulation and compared with a classic
Latest Starting Time (LST) rule.
The rest of the paper is organized as follows. In the next
section we summarize the characteristics of the charging station
that are relevant from the point of view of the scheduling
algorithm. Then, in Section 3, we review some of the recent
literature related to EVs charging scheduling. After that, in
Section 4, we define the problem as a DCSP. Section 5 follows
with a detailed description of the proposed solving procedure.
In Section 6, we report the results of the experimental study
and finally, in Section 7, we give some conclusions and ideas
for future research.
2. Description of the structure and the operation of the charging
station
The main characteristics of the electrical structure and the
operation mode of the charging station are detailed in (Sedano
et al., 2013). Fig. 1 shows the schema of the distribution grid, which
is designed to be installed in a private community park where each
user has his own space. The design criteria of the control system
were based on simplicity, economy and maintenance easiness.
The grid is fed by a three-phase source of electric power with a
voltage between phases of 400 V. Each charging point is connected
to one single-phase and supplies energy at 230 V and 2.3 kW
(Sedano et al., 2013). So, for a given contracted power, there can
be a maximum number of vehicles charging in a line at the same
time. Also, the consumption in the three lines should not be too
different at any time. Otherwise, the net is imbalanced and there
is current in the neutral point. This causes higher losses than those
of a balanced system and lowers the energy transmission effi-
ciency. Moreover, the Spanish regulations (BOE, 2013) do not allow
the installation of devices that produce large imbalances without
the consentient of the supplier company, which can penalize the
customer for it.
The station is controlled by a distributed system composed of a
master and a number of slaves. Each slave controls two consecu-
tive charging points in the same line. The master accesses the data-
base where the vehicles’ data and the charging schedule are stored.
Fig. 1. General structure of the distribution net of charging stations. It is formed by different parts such as: (1) power source, (2) three-phase electric power, (3) charging
points, (4) masters, (5) server with database, (6) communication RS 485, (7) communication TCP/IP, and (8) slaves.
307
It gathers information about the demanding vehicles from the
slaves, and sends connection/disconnection orders to them. So
the slaves are responsible for activating and deactivating charging
points as well as recording asynchronous events such as a new
vehicle arriving to the system. When entering in the station, the
user parks the vehicle in his own space (as he cannot use the space
of another user) and connects the vehicle to the charging point.
Then, he has to provide the charging time and the time he will take
the vehicle away (due date). From these data, the control system
schedules the charging times of the vehicles. The objective is in
principle that all users can be served by their due dates.
However this is not always possible and in that case what we try
to do is minimize the overall time beyond the due dates for all
users; i.e., the total tardiness. Of course, other objective functions
could be considered instead as, for example, the maximum tardi-
ness in order to not penalize a particular user in excess with
respect to the remaining ones.
In this paper we consider a simplified model of the operating
mode of the charging station that makes the following assump-
tions: the user never takes the vehicle away before the due date
and the battery does not get completely charged before the
charging time provided by the user. These are in fact unrealistic
assumptions, however the model may be adapted by introducing
new asynchronous events if they do not hold.
Each time a new vehicle requires charging, a new schedule
should be build. In principle, the system could try to accommodate
the new vehicle without changing the schedule for the remaining
vehicles in the system. However, this may not be the best option.
So, in order to obtain a new feasible schedule as good as possible,
all vehicles in the system that have not yet started to charge may
be rescheduled, and a new schedule is build with these vehicles
together with the new ones. In this process, the available resources
determined by the contracted power, the imbalance constraints
and the vehicles that are currently charging, must be taken into
account to build the new schedule.
Moreover, in order to avoid the system to collapse if, for exam-
ple, many vehicles arrive in a very short period of time, new sched-
ules are built at regular time intervals. In order to do that, after
every time interval DT (typically 120 s) or larger, a supervisor pro-
gram running on the server checks for the events produced in the
last interval. If at least one event was produced then the scheduler
308 A. Hernández-Arauzo et al. / Computers & Industrial Engineering 85 (2015) 306–315
is executed to obtain a new schedule. This schedule is active at
least during the next DT time units and so the vehicles scheduled
in this interval will start charging at their allocated times.
However, the vehicles scheduled beyond this interval may be
allocated different starting time, earlier or later, in the next
schedule. Notice that, due to the imbalance constraints, the vehi-
cles in a line may be delayed due to low demand of charge in the
other lines. So, the arrival of new vehicles to these lines may
contribute to schedule earlier a waiting vehicle in the first line.
Therefore, the operation of the charging station requires solving
a sequence of scheduling problems over time. Each one of these
problems requires in its turn scheduling the vehicles in each line
so as the three lines are balanced and the system makes the
maximum use of the contracted power. So, it seems natural to
model the overall electric vehicle scheduling problem as a DCSP.
The formal definition of this problem is given in Section 4, before
this we make a review of some related scheduling models for
electric vehicle charging.
3. Literature review on scheduling of electric vehicle charging
In this section we review some recent works dealing with
scheduling charging times of EVs. Our purpose is to demonstrate
that the problem may be considered from different points of view
and also to establish the differences of our problem with others.
A scenario where a number of EVs have to be scheduled over T
time slots satisfying due date constraints is considered in (Gan
et al., 2007; Gan, Topcu, & Low, 2011). Each EV may be charged
at different rates in different time slots and the objectives are
minimizing the peak demands and filling the valleys of power
availability. The reduction of peak power consumption is also
considered in (Kim, Lee, & Park, 2012), where feasible schedules
are built based on the activation time, the deadline, and the power
load profile of each charging task. In (Kang et al., 2013), a real-time
scheduling EV charging system is proposed that tries to optimize
total load variance and owners’ preferences as well. In (Ma et al.,
2013), the authors consider the overnight valley and some EVs
requirements. Their model establishes a feasible charging schedule
in accordance with the power constraints and sends activation/
deactivation signals to the charging stations. As in this last model,
we propose to use a centralized scheduler to control the charging
station. However, we do not consider a valley-filling strategy and
assume a constant available power instead, but our model could
be easily adapted to variations of power availability along the
day or from one day to the next, as done, for example in
(Wu et al., 2012).
A vehicle charging facility with N charging docks is described in
(Tran, Dogru, Ozen, & Beck, 2013). Based on information about arri-
vals, charging times and deadlines, a scheduling algorithm decides
which dock the vehicle is assigned, what makes a difference w.r.t.
our model where a vehicle in only allowed to park in its own space.
The vehicles in a dock are then sequentially scheduled in a First
Come First Served basis.
In (Lopes, Soares, Almeida, & Moreira da Silva, 2009), some
charging strategies are evaluated using a three-phase power flow.
The load imbalance between phases is an objective to be
minimized. This is different to considering the imbalance as a hard
constraint, as we do in this paper, which makes it more difficult to
find feasible schedules.
In (Clement-Nyns, Haesen, & Driesen, 2010), the impact of
Plug-in Hybrid EVs (PHEVs) on the distribution grid is analyzed.
The authors try to reduce power losses or voltage deviation by
means of coordinate charging when these vehicles are charged at
home. The relationship between feeder losses, load factor and load
variance are also analyzed in (Sortomme, Hindi, MacPherson, &
Venkata, 2011), where three charging algorithms are developed
with the objective of minimizing the impact of charging EVs on
the distribution system. We try to attain similar goals by means
of the balancing constraints.
These are some of the models recently proposed to dealing with
scheduling of EV charging which may be considered representative
of the state-of-the-art. From them, we can establish the differences
of the scheduling model we propose here to control the charging
station described in Section 2. The most relevant difference of
our model w.r.t. others is the fact that the imbalance among lines
is considered as a hard constraint, at difference of most of the
models which do not consider the imbalance at all or consider it
just as an objective to be optimized. Furthermore, other character-
istics of the charging station, such as for example that a stall is only
connected to one of the three lines and that each customer can
only use his own stall, make it difficult to obtain feasible and good
schedules.
4. Definition of the electric vehicle charging scheduling
problem
This problem will be termed PI herein, and for the sake of
clarity, we first consider a static version and then give a more
realistic dynamic definition.
4.1. The PI problem as a static CSP
In an instance P of the static PI problem there are 3 charging
lines Li ; 1 6 i 6 3, each one with ni charging points. N > 0 is the
maximum number of charging points that can be active at the
same time in a line. The line Li receives a number of M i vehicles
fv i1 ; . . . ; v iMi g from a time 0 up to a planning horizon. Each v ij
has an arrival time t ij P 0, a charging time pij > 0 and a due date,
dij P tij þ pij .
The goal is to obtain a feasible schedule for P, i.e., to assign
starting times to the decision variables st ij for each
v ij ; 1 6 i 6 3; 1 6 j 6 Mi , so that the following constraints are
satisfied:
i. For all vehicle v ij ; st ij P tij .
ii. No preemption is allowed, so a vehicle v ij cannot be discon-
nected before its charging time C ij ¼ st ij þ pij .
iii. The number of active charging points in a line cannot exceed
N, i.e.,
max Ni ðtÞ 6 N ð1Þ
ðtP0;i¼1;2;3Þ
where N i ðtÞ denotes the number of charging points of line Li
that are active during the time interval ½t; t þ 1Þ.
iv. The imbalance between lines must fulfill
max ðjNi ðtÞ  Nj ðtÞj=NÞ 6 D ð2Þ
ðtP0;16i;j63Þ
where D 2 ½0; 1 is a parameter.
The evaluation function is the total tardiness defined as
X
maxð0; C ij  dij Þ ð3Þ
i¼1;2;3;j¼1;...;M i
which must be minimized.
4.2. The PI problem as a dynamic CSP
Due to the fact that the arrival of vehicles is not known in
advance, it is more natural to define the PI problem as a dynamic
CSP. So, an instance P of the dynamic PI problem is given by a
 A. Hernández-Arauzo et al. / Computers & Industrial Engineering 85 (2015) 306–315
sequence of instances, P1 ; P2 ; . . . of a static CSP that we call PII
herein. An instance of the PII problem is defined by a set of vehicles
that demand charging at a given time, but they have not yet started
to charge. More formally, each Pk is defined as follows.1
In an instance Pk of the PII problem, we are given a set of vehi-
cles fv i1 ; . . . ; v ili ; . . . ; v imi g at time T k in each line Li ; 1 6 i 6 3, with
k ¼ 0; 1; . . . and T kþ1  T k ¼ DT. Each v ij has a charging time pij
and a due date dij . The vehicles v i1 ; . . . ; v ili are already active, i.e.,
they started to charge at times st ij < T k ; j ¼ 1; . . . ; li and have
not yet finished, i.e., C ij ¼ st ij þ pij > T k . While the vehicles
v ili þ1 ; . . . ; v imi 2 have not yet started to charge. So, in the iteration
k, the capacity of the line Li to charge new vehicles, denoted
is given by
X
M ki ðtÞ ¼ N  X ij ðtÞ; t P T k
16j6li
where
 Generate a new instance Pk of the problem PII with all
1; t < C ij
X ij ðtÞ ¼
0; t P C ij
The goal is to obtain a feasible schedule such that all vehicles
can be charged, even if no more vehicles arrive after T k . This
requires assigning starting times st ij to all vehicles unscheduled
at time T k so that all the constraints naturally derived from the
static PI problem are satisfied. Here it is worth to remark that a
given vehicle may get different values for st ij in subsequent
iterations while it does not start charging. However, once it starts,
it cannot longer be rescheduled.
The evaluation function is minimizing the total tardiness.
However, looking a the whole scheduling problem, we may con-
sider the following heuristic strategy: as the solution to Pk is
expected to be useful over a relatively short time (probably during
a small number of iterations until the arrival of new vehicles), we
will also try to maximize the number of charging vehicles at the
beginning. So, we consider a time horizon t h and try to maximize
the charge along the interval ½T k ; T k þ t h . This new objective may
be expressed as maximizing
Z T k þt h
ðN1 ðtÞ þ N2 ðtÞ þ N3 ðtÞÞdt
Tk
5. Solving the dynamic PI problem
To solve this problem, we adopted a similar strategy to that
used in (Rangsaritratsameea, Ferrell, & Kurzb, 2004) for the
dynamic job shop scheduling problem where the jobs are unknown
until they arrive. However, we consider rescheduling each time the
Supervisor is activated instead of each time a vehicle arrives.
Algorithm 1 shows a simulation of the algorithm we propose to
solve the dynamic PI problem. In the simulation, the problem data
and the sequence of times for the Supervisor to be executed are
given in advance. At each time T k , a new instance Pk of PII is
created if some vehicle arrived in ½T k1 ; T k Þ. P k includes all the
vehicles in the system that arrived before T k and have not yet
1
We use here a formulation that differs from the conventional definition given in
(Dechter & Dechter, 1988) where a DCSP is a sequence P 0 ; P1 ; . . . of static CSPs in
which P i is a restriction or a relaxation of P i1 . As we will see, in our case, each
instance may be obtained from the previous one by adding new constraints and doing
some relaxations.
2
Some of these vehicles has arrived by T k and so they were scheduled at T k but
were assigned a starting time greater or equal than T k and so they have not started to
charge by T k , while others have arrived in the interval ðT k1 ; T k  and so they are going
to be scheduled for the first time. From the point of view of the scheduling algorithm
there are not differences among these vehicles.
309
started to charge. The new solution S establishes starting times st ij
for all the vehicles and remains active at least along ½T k ; T kþ1 Þ. This
means that for all st ij 2 ½T k ; T kþ1 Þ; st ij is fixed to st ij in the solution
to the P instance and so the vehicle is disconnected at the comple-
tion time C ij ¼ st ij þ pij . If no vehicle arrives in ½T k1 ; T k Þ then the
current solution remains active in ½T k ; T kþ1 Þ.
Algorithm 1. Solving the PI problem.
Require: The data of a P instance of the PI problem: tij ; pij and
dij for all vehicles v ij ; and the sequence of times T 1 ; T 2 ; . . .
M ki ðtÞ,
at which the Supervisor is activated.
Ensure: A schedule S for P, i.e., starting times st ij for all
ð4Þ vehicles v ij .
S ¼ £;
for all k ¼ 1; 2; . . . do
if some vehicle arrived in ½T k1 ; T k Þ then
ð5Þ vehicles v ij s.t. tij < T k and that have not yet started to
charge, i.e., st ij has no assigned value;
Calculate a solution S to Pk ; {S establishes st ij P T k for
all vehicles v ij in Pk }
end if
For each st ij in S such that T k 6 st ij < T kþ1 , set st ij ¼ st ij in
S;
end for
return the schedule S and its total tardiness;
5.1. Solving the PII problem
The PII problem is really hard to solve because of the constraint
derived from constraint (iv.), which for this problem may be
expressed as
max ðjNi ðtÞ  Nj ðtÞj=NÞ 6 D ð7Þ
ðtPT k ;16i;j63Þ
ð6Þ It is not easy to build a schedule satisfying this constraint and at
the same time maximizing expression (6). To solve this problem,
we propose to use a dispatching rule and a more sophisticated
algorithm based on problem decomposition.
5.1.1. Solving PII with dispatching rule
We propose to use the following dispatching rule, termed LST
(Latest Starting Time), to solve a Pk instance. The unscheduled
vehicles at time T k are sorted in accordance with their latest start-
ing times given by dij  pij from low to high values; then these
vehicles are scheduled in this order and each one is given the
earliest starting time such that the scheduled vehicles satisfy all
the constraints.
This rule can be easily implemented, but enforcing all the
constraints after scheduling each vehicle may clearly prevent the
algorithm from reaching near optimal solutions.
5.1.2. Solving PII by problem decomposition
We propose here a more sophisticated method which does not
require keeping the balance constraints after each operation is
scheduled. The method relies on the following ideas. Firstly, we
establish profiles of maximum charge for each line,
N max
i ðtÞ; i ¼ 1; 2; 3. Then, to solve Pk , we try to obtain a schedule
for each line Li , so that N i ðtÞ is as close as possible to N max
i ðtÞ for
t P T k . If the combination of the solutions to the three lines does
not give rise to a feasible solution to Pk , the profiles N max
i ðtÞ must
310 A. Hernández-Arauzo et al. / Computers & Industrial Engineering 85 (2015) 306–315
be adjusted and new schedules have to be computed for one or
more lines.
The problem of calculating a schedule for a line subject to a
maximum load is denoted PIII herein. In an instance Pki of the
PIII problem we are given the set of vehicles fv i1 ; . . . ; v ili ; . . . ; v imi g
at time T k in the line Li , and a maximum charge profile for the line,
N max
i ðtÞ; t P T k . The objective is to obtain a schedule for the
vehicles, i.e., starting times st ij P T k for the inactive vehicles
v il þ1 ; . . . ; v imi , such that N i ðtÞ 6 N max ðtÞ, for all t P T k .
i i
The evaluation function is the total tardiness, defined as
X profile of the line Lj undergoes two adjustments before reaching
maxð0; C ij  dij Þ ð8Þ
j¼li þ1;...;mi
which must be minimized.
So, to solve Pk , the proposed method starts from some initial
profiles and then they are updated as long as the solutions
obtained to the Pki instances, 1 6 i 6 3, do not make up a solution
to the Pk instance.
Algorithm 2. Solving the PII problem.
Require: The data of an instance Pk of the PII problem: pij and
dij for all unscheduled vehicles v ij that arrived by T k ; and
the values st ij and pij for all vehicles scheduled before T k
such that st ij þ pij P T k .
Ensure: A schedule S for Pk .
Set the initial maximum profiles to
N max
i ðtÞ ¼ N; t P T k ; 1 6 i 6 3;
while Pk remains unsolved do
Solve the instances Pki under the current profiles N max ðtÞ;
i machines with variable availability denoted ðP; NC inc k T i Þ follow-
if these solutions do not make up a solution to Pk then
Let t0 P T k be the earliest time such that
N i ðt 0 Þ  N j ðt 0 Þ > D  N; 1 6 i; j 6 3;
Adjust the maximum profile of Li such that
N max
i ðtÞ 6 N j ðt0 Þ þ D  N; t P t 0 ;
end if
end while
return the schedule S for Pk and its total tardiness;
Algorithm 2 describes the calculation of a solution to a Pk
instance. The algorithm starts from trivial profiles N max i ðtÞ and then
iterates until a solution is reached. In each iteration, it solves the
Pki instances subject to the profiles N maxi ðtÞ. If these solutions make
up a solution for Pk , the algorithm finishes; otherwise, some profile
is adjusted from the earliest time t 0 at which an imbalance is
detected onwards.
We propose to use the following model for the profiles of max-
imum load. A profile N max i ðtÞ is given by a stepwise non increasing
function of the form:
 Pj ¼
dj sj 6 t < sjþ1 ; 1 6 j < k
N max
i ðtÞ ¼ ð9Þ
dk sk 6 t
where N P d1 >    > dk and s1 <    < sk ; k P 1. We represent this
profile as a sequence of tuples as: hðd1 ; s1 Þ; ðd2 ; s2 Þ; . . . ; ðdk ; sk Þi.
In Algorithm 2, the initial profiles are N max i ðtÞ ¼ hðN; 0Þi for all
three lines. When an imbalance of the form N i ðt0 Þ  N j ðt0 Þ > D  N
is detected after solving the three Pki instances, the profile
N max
i ðtÞ is modified in such a way that a new element
ðd; sÞ ¼ ðD  N þ N j ðt0 Þ; t 0 Þ is inserted and all tuples ðdj ; sj Þ with
dj > d and sj > s are removed from N max i ðtÞ.
This is a very simple model which helps to keep the load in the
three lines as large as possible at the beginning, hopefully along the
interval ½T; T þ th  for a sufficiently large t h , so maximizing the
expression (6). However, it may have some inconvenience as well.
For example, a new imbalance may be produced at a time just after
t0 . To avoid this drawback, we may adjust the new tuple as
ðd  dH ; s  sH Þ, where dH P 0 and sH P 0 are parameters to be
established empirically. Also, the next imbalance may be at a time
lower than t0 due to the non-preemption constraint. In any case,
the value of N max i ðtÞ, for each time t, is non increasing on
subsequent adjustments. This guarantees that Algorithm 2 stops
after a finite number of steps. Fig. 2 shows a situation where the
a compatible solution with that of line Li .
5.2. Solving the PIII problem
A Pki instance may be viewed as that of scheduling a number of
mi  li non-preemptive jobs, all of them available at time T k , on a
machine whose capacity varies over time and the objective is
minimizing the total tardiness. The processing time of the jobs
are the charging times of the vehicles v ili þ1 ; . . . ; v imi , and the
capacity of the machine at time t for the line Li , denoted Capki ðtÞ,
is calculated as
Capki ðtÞ ¼ minðMki ðtÞ; Nmax ðtÞÞ; t P T k ð10Þ
i
Fig. 3 shows the capacity of the machine for an instance of this
P
problem. This problem may be denoted ð1; CapðtÞk T i Þ following
the conventional notation ða j b j cÞ proposed in (Graham, Lawler,
Lenstra, & Kan, 1979).
In the simple case where the capacity CapðtÞ is non decreasing,
the problem is equivalent to the problem of identical parallel
P
ing the notation used in (Schmidt, 2000), where P is the number of
parallel machines and N inc denotes that the availability of machine
is non decreasing over time. Scheduling problems with machine
availability appear in many situations, for example when mainte-
nance periods are considered, with different profiles of machine
availability. This kind of problems is surveyed in (Ma, Chu, &
Zuo, 2010).
P
In (Koulamas, 1994), the ðPk T i Þ problem, in which the
machine is continuously available, is proved to be at least binary
NP-hard. An efficient simulated annealing algorithm for this prob-
lem is proposed in (Sang-Oh Shim & Kim, 2007). In this algorithm,
the starting solution is obtained by means of the apparently
tardiness rule. This rule was adapted for similar problems in
(Kaplan & Rabadi, 2012), to deal with ready times and due date
constraints. In this paper, we propose to adapt this rule to solve
P
the ð1; CapðtÞk T i Þ problem as follows: let CðaÞ be the earliest
starting time for an unscheduled job in the partial schedule a built
so far. Then for all unscheduled jobs that can start at CðaÞ a
selection probability is calculated as
 
1 maxð0; dj  CðaÞ  pj Þ
exp ð11Þ
pj gp
where p is the average processing time of the jobs and g is a look-
ahead parameter to be fixed empirically. The job with the largest
probability is scheduled at time CðaÞ.
5.3. Complexity considerations
As the time taken by the scheduling algorithm issued at each
time T k is critical, it is worth justify that Algorithm 2 terminates
in polynomial time. First of all, the algorithm that solves an
instance of the PIII problem with n vehicles requires calculating
CðaÞ and p and then expression (11) for a subset of the vehicles.
So, the time taken is of order OðnÞ. After solving the three PIII
 A. Hernández-Arauzo et al. / Computers & Industrial Engineering 85 (2015) 306–315 311

Fig. 2. Illustration of adjustments of the profile of maximum load of the line Lj to get a compatible schedule with that of line Li . D = 0.2 and N = 5, so the maximum difference
in the number of active EVs in the two lines is 2.

P bution of vehicles on the lines. In the first one (type 1) the vehicles
Fig. 3. Capacity of the machine for an instance of the ð1; CapðtÞk T i Þ obtained at a
time T k where three vehicles are charging in line j and the profile of maximum load
has been adjusted from s1 to N  1 and from s2 to N  2.
instances, checking for the imbalance constraints takes a time
lower than the total units of duration of the charging periods, T.
If there is some imbalance, new instances of the PIII have to be
solved, but the maximum number of times is finite as indicated
in Section 5.2 and also limited by T. So, in the worst case,
Algorithm 2 runs in a time of order Oðn  T 2 Þ.
6. Experimental study
arriving on the whole day through, but with arrival peaks at some
times. It was generated from the probability distributions given in
the first two columns of Table 1 (scenario 1). Also, we considered
another situation where most of the vehicles arrive almost at the
same time, with two different due date profiles. Table 1 (scenarios
2 and 3) shows the probability distributions used to generate the
arrival times of these instances.
The initial charge of the batteries was distributed as shown in
Table 2. The charging times are then calculated assuming that all
the vehicles require charging at 100% of their capacity, 23 kW h,
and that the charging rate is 2.3 kW. Two different due date pro-
files were considered as it is also shown in Table 2.
In the three cases, we considered two different profiles of distri-
are uniformly distributed, while in the second (type 2) the distribu-
tion is 10%, 30% and 60% on lines 1, 2 and 3 respectively. From each
scenario and type 30 instances were generated. Also, we consid-
ered four values for D, 0.2, 0.4, 0.6 and 0.8; and three values for
N, 20, 30 and 40. So, combining all these options we have 2160
instances in all, which are organized into 72 classes each one
Table 1
Summary of percentages and probability distributions used to generate the arrival
times in the three scenarios considered. Uðx; yÞ denotes uniform distribution in the
interval ½x; y and Nðx; yÞ denotes normal distribution with mean x and standard
deviation y. x and y are given in minutes.

Scenario 1 Scenarios 2 and 3

As it was done with the models summarized in Section 3, we
%Vehicles Arrival time %Vehicles Arrival time
evaluated the scheduling algorithm by simulation. To do that, we
10 Uð0; 1440Þ 20 Uð0; 1440Þ
have firstly defined a set of instances of the PI problem and then
20 Nð510; 15Þ 80 Nð870; 30Þ
we implemented a simulator to run the Algorithm 1, termed EVS 10 Nð720; 15Þ
(Electric Vehicles Scheduling) herein.3 In this study, we evaluated 50 Nð1170; 15Þ
the proposed EVS algorithm and compared it with the algorithm 10 Nð1350; 15Þ
derived from the LST rule.

6.1. Definition of the benchmark set Table 2

Summary of percentages and probability distributions used to generate the initial
charges and due dates for all the instances. Times are expressed in minutes.
We consider a charging station installed in a car park with 180
spaces, a time horizon of one day and different profiles of arriving %Vehicles Initial charge (in Due date scenarios 1 Due date
times, battery loads, due dates and distribution on the lines. The percent) and 2 scenarios 3
first scenario may represent a normal weekly day with vehicles 10 Nð80; 10Þ Nð240; 120Þ Nð120; 60Þ
30 Nð50; 15Þ Nð360; 120Þ Nð300; 60Þ
3
30 Nð35; 7:5Þ Nð480; 120Þ Nð360; 60Þ
These instances and more details of the experimental study are available at 30 Nð12; 6Þ Nð660; 120Þ Nð540; 60Þ
http://www.di.uniovi.es/iscop (Repository).
312 A. Hernández-Arauzo et al. / Computers & Industrial Engineering 85 (2015) 306–315

characterized by one tuple (type, scenario, D; N). For all the requires to solve the instances in the subsets derived from scenario
instances, the generated data are consistent, i.e., t ij þ pij 6 dij . To 3. As we can observe, this number decreases in inverse ratio to D
ensure this, the charging time was set as pij ¼ dij  tij if the and N, strongly in the first case and moderately in the second. So,
randomly generated value were larger than this. these results show that larger values of these parameters make
the instances easier to solve. In any case, the average number of
adjustments is really low and so it can be expected not to have a
6.2. Evaluation of the proposed algorithm
great influence on the time taken by the scheduler.
To evaluate EVS, we started considering different values for the
lookahead parameter g of expression (11). Table 3 summarizes the 6.3. Comparison of EVS with LST
values of the total tardiness obtained for all instances with four
values of g, 0.25, 0.50, 0.75 and 1.00. In all cases, the parameters All the instances were also solved with the algorithm LST and it
dH and sH were set to 0. As it was expected, the instances of the sce- is important to start remarking that for most of the instances EVS
nario 3 produce the largest tardiness as they represent situations reached a solution better than that obtained by LST. Only for 4
with the tightest due dates and most of the vehicles arrive around instances were LST better than EVS. Fig. 5 shows the comparison
the same time. Also, the lowest tardiness are obtained for the of the results obtained with EVS ðg ¼ 0:25Þ with those from the
instances of scenario 1 where the vehicles arrive more uniformly LST rule, averaged for each one of the 72 subsets. As we can
along the day and the due dates are not so tight. observe, the tardiness obtained with LST is in some cases more
Regarding the parameter g, there is significant difference in than 50% larger than that obtained by EVS.
favor of 0.25. Table 4 shows the average ranking computed by In Fig. 5, we can also appreciate the differences in the tardiness
the Friedman test for all the instances. of the instances in each one of the 72 subsets. In general, instances
It is also worth analyzing the number of adjustments required of type 2 show more tardiness than instances of type 1, the
to reach a solution to a PII instance, as it may have an important tardiness growing up quickly from scenario 1 to scenario 2, and
impact on the time required by EVS to reach a solution. Fig. 4 moderately from scenario 2 to scenario 3, as it was also shown in
shows the mean number of adjustments that EVS, with g ¼ 0:25, Table 3. This is reasonable as in scenario 1 the arrivals are more
uniformly distributed along the day, while in scenarios 2 and 3
almost all the vehicles arrive around the same time. Also in sce-
Table 3
Summary of results from EVS with four values of parameter g, the values are averaged nario 3 the due dates are more tight than they are in scenario 2.
for each scenario. The reported values have to be multiplied by E+06 to represent At the same time, the tardiness decreases when N and D increase
tardiness in minutes. for all subsets. However, for instances of type 2, the D has stronger
Value of parameter g influence on the tardiness than it has for type 1. This is quite
natural due to the fact that instances of type 2 have a non uniform
Scenario 0.25 0.50 0.75 1.00
distribution of vehicles on the three lines and so a large D allows
1 2.024 2.033 2.038 2.045
for large imbalance among the lines and then more vehicles can
2 3.081 3.096 3.099 3.104
3 3.256 3.258 3.257 3.258 be charging at the same time in the most loaded lines.
Special mention deserves the high similarity of graphics on
Average 2.787 2.795 2.798 2.803
Figs. 5(d) and 5(f). Remember that they represent results from
instances generated with the same parameters, with the exception
of those that control the due dates. So the charging times and the
arrival times of the vehicles are similar, but there are differences of
Table 4 about 90 min in the average due dates. So, in principle it is
Average rankings of the EVS algorithm with four values of parameter g given by a
reasonable to expect that the average tardiness for the instances
Friedman test with p-value of 8.97E11.
[scenario, type] = [3, 2] is at most 90  60  180 = 9.72E+05
Value of parameter g Ranking seconds larger than they are for instances [2, 2]. This difference
0.25 1.28 being negligible for tardiness of around 1.00E+07 seconds that
0.50 2.19 produce the hardest instances of type 2. So, this fact together with
0.75 2.76
the reasonable differences in the charging and arriving times make
1.00 3.76
the results almost indistinguishably.

Fig. 4. Number of adjustments required by EVS to reach feasible solutions to the instances of the PII problem along the time horizon. The y-axis represents the number of
adjustments (1000) and the x-axis represents the subsets T i N D  100; i denotes the type (1, 2). The results are averaged for the 30 instances in each subset. Only instances
of scenario 3 were considered.
 A. Hernández-Arauzo et al. / Computers & Industrial Engineering 85 (2015) 306–315 313

4,00E+06 2,50E+07

3,50E+06
2,00E+07
3,00E+06

2,50E+06 1,50E+07

2,00E+06

1,00E+07
1,50E+06

1,00E+06
5,00E+06
5,00E+05

0,00E+00 0,00E+00
20_20 20_40 20_60 20_80 30_20 30_40 30_60 30_80 40_20 40_40 40_60 40_80 20_20 20_40 20_60 20_80 30_20 30_40 30_60 30_80 40_20 40_40 40_60 40_80

(a) [1,1] (b) [1,2]

4,00E+06 2,50E+07

3,50E+06
2,00E+07
3,00E+06

2,50E+06
1,50E+07

2,00E+06

1,00E+07
1,50E+06

1,00E+06
5,00E+06
5,00E+05

0,00E+00 0,00E+00
20_20 20_40 20_60 20_80 30_20 30_40 30_60 30_80 40_20 40_40 40_60 40_80 20_20 20_40 20_60 20_80 30_20 30_40 30_60 30_80 40_20 40_40 40_60 40_80

(c) [2,1] (d) [2,2]

4,00E+06 2,50E+07

3,50E+06
2,00E+07
3,00E+06

2,50E+06
1,50E+07
,

2,00E+06

1,00E+07
1,50E+06

1,00E+06
5,00E+06
5,00E+05

0,00E+00 0,00E+00
20_20 20_40 20_60 20_80 30_20 30_40 30_60 30_80 40_20 40_40 40_60 40_80 20_20 20_40 20_60 20_80 30_20 30_40 30_60 30_80 40_20 40_40 40_60 40_80

(e) [3,1] (f) [3,2]

Fig. 5. Results from LST and EVS ðg ¼ 0:25Þ averaged for the six subsets defined by [scenario, type], scenario = 1, 2, 3; type = 1, 2. y-axis represent total tardiness in seconds,
with different scales for each type, and x-axis represent combinations of N D  100. Grey bars show LST values and black bars EVS values.

In order to enhance the comparison among EVS and LST, we
considered the performance gap in percent of LST and EVS against
the best known solution for each instance calculated as:
ðTd  BKTdÞ=BKTd  100
where Td refers to the tardiness obtained by one algorithm to an
instance and BKTd is the tardiness of the best known solution for
this instance, which in our case is taken as the best obtained by
LST and EVS. So, large values of the performance gap mean poor per-
formance of a method. As in some cases BKTd is 0, following Bülbül
(2011) we opted to show the average gap for the instances in each
subset with BKTd different from 0 and additionally the total gap
defined for each subset of 30 instances as
, exception of 0.8; while for type 2 instances, the difference is in
X X
ðTdi  BKTdi Þ BKTdi  100
i¼1...30 i¼1...30
These values are summarized in Table 5. As we can observe,
only in 4 out of the 72 subsets the average or the total gap of
EVS is not 0 due to the 4 instances where LST reached better tardi-
ness than EVS. In all subsets, the gaps are clearly much larger for
ð12Þ
LST than they are for EVS (they are indeed zero in this case for
almost all instances) in accordance with the results given in
Fig. 5. Furthermore, Table 5 allows us to highlight the differences
between LST and EVS depending on the four parameters: type, sce-
nario, D and N. Firstly, it is clear that the superiority in perfor-
mance of EVS w.r.t. LST grows up from scenario 3 to scenario 1.
Also, EVS gets better and better than LST with increasing values
of N. However, the influence of D is different for instances of types
1 and 2. For type 1, the difference is in direct ratio to D, with the
ð13Þ inverse ratio. Overall, the total performance gap in percent of
EVS is almost 0 for all subsets and it varies from 6.2% to 517.6%
314 A. Hernández-Arauzo et al. / Computers & Industrial Engineering 85 (2015) 306–315

Table 5
Summary of average and total gaps in percent for all 72 subsets of 30 instances each. The columns labelled with #T = 0 show the number of instances with best known tardiness 0.
The symbol ‘–’ means 0.

Scenario 1 Scenario 2 Scenario 3

LST EVS LST EVS LST EVS
t ND #T = 0 Av. Tot. Av. Tot. #T = 0 Av. Tot. Av. Tot. #T = 0 Av. Tot. Av. Tot.
1_20_20 – 140.0 117.3 – – – 52.6 51.1 – – – 42.6 41.3 – –
1_20_40 – 78.6 75.2 – – – 35.2 35.2 – – – 29.2 29.2 – –
1_20_60 – 33.9 33.9 – – – 27.7 27.6 – – – 21.6 21.6 – –
1_20_80 – 25.1 24.3 – – – 21.5 21.3 – – – 16.7 16.6 – –
1_30_20 – 506.1 251.3 – – – 85.8 80.6 – – – 53.7 52.3 – –
1_30_40 – 150.5 126.2 – – – 43.7 43.2 – – – 26.0 26.5 – –
1_30_60 – 63.3 54.5 – – – 22.0 21.1 – – – 11.6 11.7 – –
1_30_80 – 59.0 48.6 – – – 18.6 17.8 – – – 9.9 9.7 – –
1_40_20 – 3439.5 277.0 – – – 187.7 160.7 – – – 107.9 93.5 – –
1_40_40 20 1928.2 517.6 – 1.2 – 58.2 50.1 – – – 14.4 15.9 0.2 0.1
1_40_60 25 217.3 37.7 – – – 29.2 25.9 – – – 6.1 6.3 0.2 0.1
1_40_80 25 217.3 37.7 – – – 29.1 25.7 – – – 6.0 6.2 0.2 0.1
2_20_20 – 32.3 31.9 – – – 26.1 25.9 – – – 24.3 24.0 – –
2_20_40 – 37.4 35.9 – – – 28.8 28.5 – – – 26.7 26.3 – –
2_20_60 – 46.0 43.3 – – – 32.6 32.2 – – – 28.7 28.2 – –
2_20_80 – 39.4 40.1 – – – 29.8 29.5 – – – 25.5 25.4 – –
2_30_20 – 33.4 32.7 – – – 26.3 26.0 – – – 24.3 23.9 – –
2_30_40 – 35.9 34.1 – – – 29.8 29.3 – – – 27.7 26.7 – –
2_30_60 – 53.8 46.2 – – – 39.0 37.9 – – – 34.0 33.2 – –
2_30_80 – 42.4 44.0 – – – 32.6 32.2 – – – 25.5 25.4 – –
2_40_20 – 34.9 33.9 – – – 27.0 26.8 – – – 24.5 24.0 – –
2_40_40 – 37.5 34.3 – – – 30.3 29.8 – – – 25.9 24.8 – –
2_40_60 – 57.7 47.6 – – – 40.0 38.5 – – – 36.9 35.2 – –
2_40_80 – 59.2 49.0 – – – 35.4 34.6 – – – 25.2 25.1 – –
Av. 86.4 38.8 27.2

for LST, being 50.8% in average. Of course, it is important to take
into account that this gap is calculated w.r.t. the best known solu-
tion, which in almost all cases is the one obtained by EVS, and not
w.r.t. an optimal solution or a lower bound, in which case the gap
would be less favorable to EVS.
Regarding the time taken by the algorithms, For EVS it was
about 15% larger than it was for LST in average, and in no case this
time was greater than 12 ms, what is negligible w.r.t. the time
between two rescheduling points (120 s).
contracted power, what would allow the operator to adjust the
amount of contracted power to the consumption valleys of the
grid, for example. Furthermore, the model may be adapted to situ-
ations where the vehicles can be charged at variable rate. As it is
pointed in (Sedano et al., 2013), this is technically possible under
some restrictions and offers much more flexibility to organize
the charging of vehicles over time. However, this adaptation is
not trivial and will require reformulating some aspects of the prob-
lem and adapting the scheduling algorithms accordingly.

6.4. Discussion of the results 7. Conclusions

The results of this study show that the proposed EVS algo-
rithm performs better than a dispatching rule such as LST. In
our opinion, the superiority of EVS relies on how it deals with
the balance constraints. Instead of keeping this constraint after
each operation is scheduled, as it is done by the algorithm
derived from the LST rule, EVS establishes profiles of maximum
load in each of the three lines and then adapts the schedules to
these profiles. Even though these profiles may require a number
of adjustments, the time taken by the algorithm is quite reason-
able as it is negligible w.r.t. the time between two consecutive
executions of the scheduler.
As it was pointed in Section 2, we have considered here a sim-
plified version of the problem. However, the proposed model and
solving methods can be extended to deal with a number of charac-
teristics of the real situations that have been skipped here. For
example, the users may take the vehicle away before the due date
dij , or the battery may get fully charged before the expected charg-
ing time pij . In both cases an imbalance may be produced in the
system. To deal with these situations, some new asynchronous
events have to be added to the model. Also, the charging time of
the vehicles may be reduced in situations of saturation in order
to reduce the tardiness; or even some vehicles may not be
attended. The proposed model can also be adapted to variable
We have seen that scheduling the charging of electric vehicles
may be formulated as a Dynamic Constraint Satisfaction Problem
(DCSP) with optimization. In this paper, we have considered a
problem of this family motivated by a real charging station fed
by three-phase power and designed to be installed in a commu-
nity park, where each vehicle has its own space. We have seen
that this problem is hard to solve mainly due to the maximum
imbalance in the lines of the three-phase electric feeder. We have
proposed an effective algorithm that reduces the calculation of a
solution to solving a large number of instances of the one
machine sequencing problem with variable capacity, denoted
P
ð1; CapðtÞk T i Þ, which is unedited in the literature and was here
solved by a dispatching rule. A formal study of the
P
ð1; CapðtÞk T i Þ problem may lead to devise new solution meth-
ods, which may contribute to improve the performance of the
whole scheduling algorithm. As future work, we will analyze
the use of some exact, greedy and evolutionary algorithms to
solve this problem as it was recently done, for example, in
(Kodaganallur, Sen, & Mitra, 2014) for a version of the one
machine sequencing problem with setup times and quadratic
penalty completion times, or in (Ji, Ge, Chen, & Cheng, 2013) for
a more complex version with common due-window and deterio-
rating rate of the machine and job durations.
 A. Hernández-Arauzo et al. / Computers & Industrial Engineering 85 (2015) 306–315
Acknowledgments
This research has been supported by the Spanish Government
under research Projects TIN2010-20976-C02-02 and TIN2013-
46511-C2-2-P, by the ITCL under Contract FUO-EM-075-13 an by
the Principality of Asturias under Contract FC-13-COF13-035.
References
BOE (22 September 2013). Low Voltage Electrotechnical Regulation (TBR). Royal
decree 842/2002, of 2 August 2002. Official Gazette of Spain (BOE). <http://
www.boe.es/>.
Bülbül, K. (2011). A hybrid shifting bottleneck-tabu search heuristic for the job shop
total weighted tardiness problem. Computers & Operations Research, 38,
967–983.
Clement-Nyns, K., Haesen, E., & Driesen, J. (2010). The impact of charging plug-in
hybrid electric vehicles on a residential distribution grid. IEEE Transactions on
Power Systems, 25, 371–380.
Dallinger, D. (2014). Plug-in electric vehicles integrating fluctuating renewable
electricity. Kassel University Press GmbH.
Dechter, R., & Dechter, A. (1988). Belief maintenance in dynamic constraint
networks. In Proceedings of the seventh annual conference of the American
association of artificial intelligence (pp. 37–42).
EDSO (10 April 2012). Position paper on electric vehicles charging infrastructure.
European Distribution System Operators for Smart Grids (EDSO). <http://www.
edsoforsmartgrids.eu>.
Gan, L., Topcu, U., & Low, S. (2011). Optimal decentralized protocol for electric
vehicle charging. In IEEE conference on decision and control and European control
conference (pp. 5798–5804). CDC-ECE.
Gan, L., Topcu, U., & Low, S. (2007). Optimal decentralized protocols for electric
vehicle charging. IEEE Transaction on Power Systems, 6, 1–10.
Graham, R., Lawler, E., Lenstra, J., & Kan, A. (1979). Optimization and approximation
in deterministic sequencing and scheduling: A survey. Annals of Discrete
Mathematics, 5, 287–326.
Hernández-Arauzo, A., Puente, J., González, M. A., Varela, R., & Sedano, J. (2013).
Dynamic scheduling of electric vehicle charging under limited power and phase
balance constraints. In Proceedings of SPARK’13. ICAPS’13 scheduling and planning
applications workshop (pp. 1–8). ICAPS Council.
Ji, M., Ge, J., Chen, K., & Cheng, T. (2013). Single-machine due-window assignment
and scheduling with resource allocation, aging effect, and a deteriorating rate-
modifying activity. Computers & Industrial Engineering, 66, 952–961.
315
Kang, J., Duncan, S. J., & Mavris, D. N. (2013). Real-time scheduling techniques for
electric vehicle charging in support of frequency regulation. Procedia Computer
Science, 16, 767–775.
Kaplan, S., & Rabadi, G. (2012). Exact and heuristic algorithms for the aerial
refueling parallel machine scheduling problem with due date-to-deadline
window and ready times. Computers & Industrial Engineering, 62, 276–285.
Kim, H. J., Lee, J., & Park, G. L. (2012). Constraint-based charging scheduler design for
electric vehicles. Proceedings of the 4th Asian conference on intelligent information
and database systems (Vol. Part III, pp. 266–275). Berlin, Heidelberg: Springer-
Verlag.
Kodaganallur, V., Sen, A. K., & Mitra, S. (2014). Application of graph search and
genetic algorithms for the single machine scheduling problem with sequence-
dependent setup times and quadratic penalty function of completion times.
Computers & Industrial Engineering, 67, 10–19.
Koulamas, C. (1994). The total tardiness problem: Review and extensions.
Operations Research, 42, 1025–1041.
Lopes, J., Soares, F., Almeida, P. M., & Moreira da Silva, M. (2009). Smart charging
strategies for electric vehicles: Enhancing grid performance and maximizing the
use of variable renewable energy resources. In EVS24 international battery,
hybrid and fuel cell electric vehicle symposium, 2009 (pp. 392–396).
Ma, Z., Callaway, D., & Hiskens, I. (2013). Decentralized charging control of large
populations of plug-in electric vehicles. IEEE Transactions on Control Systems
Technology, 21, 67–78.
Ma, Y., Chu, C., & Zuo, C. (2010). A survey of scheduling with deterministic machine
availability constraints. Computers & Industrial Engineering, 58, 199–211.
Rangsaritratsameea, R., Ferrell, W. G., & Kurzb, M. B. (2004). Dynamic rescheduling
that simultaneously considers efficiency and stability. Computers & Industrial
Engineering, 46, 1–15.
Sang-Oh Shim, S. O., & Kim, Y. D. (2007). Scheduling on parallel identical machines
to minimize total tardiness. European Journal of Operational Research, 177,
135–146.
Schmidt, G. (2000). Scheduling with limited machine availability. European Journal
of Operational Research, 121, 1–15.
Sedano, J., Portal, M., Hernández-Arauzo, A., Villar, J. R., Puente, J., & Varela, R.
(2013). Intelligent system for electric vehicle charging: Design and operation.
DYNA, 88, 640–647.
Sortomme, E., Hindi, M., MacPherson, S., & Venkata, S. (2011). Coordinated charging
of plug-in hybrid electric vehicles to minimize distribution system losses. IEEE
Transactions on Smart Grid, 2, 198–205.
Tran, T.T., Dogru, M.K., Ozen, U., Beck, C., 2013. Scheduling a multi-cable electric
vehicle charging facility. In Proceedings of SPARK’13. ICAPS’13 scheduling and
planning applications woRKshop (pp. 20–26). ICAPS Council.
Wu, D., Aliprantis, D., & Ying, L. (2012). Load scheduling and dispatch for
aggregators of plug-in electric vehicles. IEEE Transactions on Smart Grid, 3,
368–376.