3100000016

Foundations and Trends
R
in Electric Energy Systems
Vol. 2, No. 1 (2017) 1–132

c 2017 F. Kong and X. Liu
DOI: 10.1561/3100000016
Sustainable Transportation with Electric Vehicles
Fanxin Kong Xue Liu

University of Pennsylvania McGill University
Contents
1 Introduction and Overview 3

1.1 Electric Vehicle Overview . . . . . . . . . . . . . . . . . . 5
1.2 Charging Infrastructure Overview . . . . . . . . . . . . . . 6
1.3 Electric Vehicle Charging Overview . . . . . . . . . . . . . 8
1.4 A Systematic Solution Framework . . . . . . . . . . . . . 11
1.5 Road Map . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2 Charging Rate Control 13

2.1 Recent Developments . . . . . . . . . . . . . . . . . . . . 14
2.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Optimal Charging Rate Control . . . . . . . . . . . . . . . 23
2.4 Performance Evaluation . . . . . . . . . . . . . . . . . . . 25
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.6 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3 Charging Demand Response 32

3.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3 The Stackelberg Game Between A Charging Station and EVs 40
3.4 The Stackelberg Equilibrium with Co-located Renewable . 48
ii
iii
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.7 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4 Charging Load Scheduling 64

4.2 Problem Description . . . . . . . . . . . . . . . . . . . . . 69
4.3 Algorithm Analysis Based on Competitive Analysis . . . . . 74
4.4 Algorithm Analysis Based on Resource Augmentation . . . 84
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5 Charging Demand Balancing 97

5.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . 100
5.3 Algorithm Design for Demand Balancing . . . . . . . . . . 102
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6 Conclusions 117
Acknowledgements 120
References 121
Abstract
Electric vehicles are gaining more and more popularity due to low
oil dependency and low emission. Their deep penetration will signif-
icantly benefit the environment, but meanwhile will cause two crucial
consequences. First, electric vehicles introduce heavy load impact into
the power grid by shifting energy demand from gasoline to electricity.
The surging load will compromise the grid’s reliability and jeopardize
its power supply quality. Second, charging stations become indispens-
able infrastructure to support large deployment of electric vehicles. The
availability in public destinations comes with electric vehicles compet-
ing for both power supply and service points of charging stations. The
competition degrades quality of service and thus can compromise the
original intent of advocating electric vehicles.
There are many research efforts addressing either of the two conse-
quences above. Different with them, we consider both and jointly study
quality of service for electric vehicle users and reliability of the power
grid. We review recent developments on this topic in this article. In
Chapter 1, we introduce the ecosystem of electric vehicles and discuss
motivations for managing charging load. This chapter further presents
a systematic solution framework for smart electric vehicle charging. The
following chapters then study each block of the framework. Specially, in
Chapter 2, we investigate charging rate control, which handles how to
allocate power supply to electric vehicles within a charging station. In
Chapter 3, we address electric vehicle demand response, which is how
to make electric vehicles follow the power supply of charging stations
and the power grid. In Chapter 4, we study electric vehicle scheduling,
which copes with how to schedule electric vehicles to multiple charg-
ing points within a charging station. In Chapter 5, we discuss charging
demand balancing, which deals with how to balance electric vehicles
among multiple charging stations.
In these chapters, we first present deployable algorithms and mech-
anisms that are designed for each framework blocks. Then, we evalu-
ate the proposed approaches by two complementary ways. One way is
leveraging theoretical analysis to demonstrate their performance guar-
antees, while the other is using extensive simulations based on realistic
2
data traces and simulation tools. We also review studies that align
with the corresponding framework blocks and consider additional di-
mensions and/or different optimization goals. Finally, in Chapter 6,
we conclude the article with summaries of main ideas discussed in the
previous chapters.
F. Kong and X. Liu. Sustainable Transportation with Electric Vehicles. Foundations

and Trends R
in Electric Energy Systems, vol. 2, no. 1, pp. 1–132, 2017.
DOI: 10.1561/3100000016.
1
Introduction and Overview
The transportation sector is by far the largest oil consumer and thus
a prime contributor to air pollution. For example, the sector accounts
for about 23% of the global GHG emissions in 2014 [1]. The growing
concerns over environmental impacts and oil scarcity have boosted the
need to electrify the transportation sector and spurred efforts to accel-
erate the adoption of electric vehicles (EVs). As shown by Fig. 1.1(a),
the yearly EV sales for U.S. have grown more than 6 times from 2011
to 2015, and as shown by Fig. 1.1(b), the worldwide yearly sales would
be over 6 million by 2020. Further, the number of EVs in the globe
would be over 35 million by 2022 [2].
The popularity of EVs will significantly benefit the environment
and alleviate energy crisis, but meanwhile will cause heavy load im-
pact to the power grid due to shifting energy demand from gasoline to
electricity. The potential impact includes compromising the grid’s reli-
ability and jeopardizing its power quality. For example, uncoordinated
EV charging can increase the peak load and energy losses and overload
distribution lines and transformers [3, 4, 5, 6]. Overloading can over-
heat the transformers, and accelerate their degradation, and eventually
cause premature failure to them. This impact would be even severer
3
4 Introduction and Overview
(a) U.S. EV sales 2010-2015 (b) World-wide EV sales 2013-2020
Figure 1.1: Market research about EV sales.
with some eco-friendly areas such as neighbourhoods with higher pen-

etration of EVs.
Recently, various charging facilities have been deployed such as
charging points in residential areas and working places [3, 7, 8, 9,
10, 11, 12]. Among them, charging stations have become indispens-
able infrastructure to support the deep integration of electric vehicles
[10, 11, 12, 13, 14]. Meanwhile, one crucial consequence following their
public availability is that EVs will compete for charging resources such
as power supply and service points of these charging stations. This
competition can much degrade the quality of service if there is no coor-
dination performed among EV users when making decisions on choosing
charging stations. For example, some charging stations may be over-
loaded with long waiting queues of charging demand (and thus long
waiting time for EVs) while others barely have EVs to serve. Some EVs
that have a tight schedule may be allocated with lower power supply
than EVs with adequate spare time for charging.
There are two different methods to accommodate the large-scale
EV charging load. The first method is to make the required invest-
ment to upgrade the power grid and build more charging facilities. The
second method is to exploit the elasticity of charging load and exist-
ing communication networks to coordinate and control EV charging,
i.e., to enable smart EV charging. This method postpones upgrading or
building new charging infrastructure and thus substantially reduces the
required reinforcement investment [15, 16]. This article focuses on the
1.1. Electric Vehicle Overview 5
latter method and studies smart EV charging. In the following, we first

present an overview for electric vehicles and charging infrastructure,
and then reveal emerging challenges and propose a systematic solution
framework for smart EV charging.
1.1 Electric Vehicle Overview
Although EVs were first introduced many decades ago, their resurgence
and actual popularity start recently due to technological developments
and the environmental impact of petroleum-based transportation [17].
Electric vehicles use electric motors for propulsion and can be powered
by electricity from on-board batteries. In contrast, conventional vehicles
use internal combustion engines for propulsion and usually depend on
non-renewable fossil fuels. These two kinds of vehicles are different in
that the electricity EVs consume can be generated from a wide range
of energy sources, including fossil fuels and renewable sources such as
solar power and wind power or the combinations of those sources. EVs,
if sourcing from renewable, thus come with lower carbon footprint and
other emissions than conventional vehicles.
This article confines to studying plug-in electric vehicles, which can
be recharged from any external source of electricity, such as the power
grid and local renewable generation. Plug-in EVs can be further clas-
sified into two different types: all-electric or battery electric vehicles
(BEVs) and plug-in hybrid electric vehicles (PHEVs). BEVs only use
electric motors for propulsion instead of internal combustion engines.
They have no fuel tank and derive all power from rechargeable batteries
on-board. Examples of BEVs include Nissan Leaf, Ford Focus Electric,
Tesla Model S, BMW i3, and BYD Qin EV300. Fig. 1.2(a) shows Nis-
san Leaf 2017. PHEVs share the characteristics of both conventional
vehicles and BEVs, and thus have both electric motors and internal
combustion engines. They can derive power both from on-board batter-
ies that can be recharged by plugging into external sources, and from
combusting fossil fuel from fuel tanks. Examples of PHEVs include
Chevrolet Volt, BYD F3DM, BMW i8, and Toyota Prius. Fig. 1.2(b)
shows Chevrolet Volt 2017.
(a) Nissan Leaf 2017 (b) Chevrolet Volt 2017
Figure 1.2: Examples of BEV and PHEV.
Most plug-in EVs of this generation use lithium-ion (li-ion) bat-

teries. Compared to most other rechargeable batteries, li-ion batteries
have advantages such as higher energy density, higher power density,
and long life span. For example, Nissan Leaf has 30 kW h battery pack
with 172 km on a full battery charge, while the latest Tesla Model S is
equipped with 100 kW h battery pack, which can allow a driving range
as long as 539 km. Battery life span is defined as the number of full
charge-discharge cycles before significant capacity loss. Li-ion batteries
degrade, in terms of capacity, energy efficiency, as the number of cy-
cles increases. To maintain a long life span, li-ion batteries should not
operate with frequent mode switching, i.e., switching between charg-
ing and discharging. One disadvantage is that li-ion batteries can pose
safety issues since they contain flammable electrolytes. For example, it
would be very risky to charge EV batteries with a power supply larger
than the allowed maximum charging rate. In order to operate safely
and efficiently, li-ion batteries should be used within safe temperature
and certain power ranges.
1.2 Charging Infrastructure Overview
Charging stations or charging points are a fundamental element in

charging infrastructure that supplies electricity for the recharging of
EVs. The charging infrastructure usually exists in three different con-
texts. The first is residential charging stations, where an EV owner
plugs in when he or she at home, and the EV usually charges overnight
1.2. Charging Infrastructure Overview 7
Table 1.1: Charging Standard Specification [23].
Charging Level Voltage [V] Max. Current [A] Max. Power [kW]
AC Level 1 120 12 1.44
AC Level 2 208-240 48 11.5
AC Level 3 208-240 400 96
DC Charging 208-600 400 240
[3, 18]. The second context is park-and-charge, i.e., charging while

parked, where a parking lot is equipped with charging function and
EVs can receive recharging while parked [19, 20]. This scenario encour-
ages EV users to take advantage of nearby facilities, such as parking
stations including parking at malls, working places and airports. The
third is publicly available fast charging stations, which usually can pro-
vide an individual charging rate larger than 40 kw and takes dozens of
minutes to deliver tens of miles [5]. This article mainly studies charg-
ing stations falling into the latter two contexts, i.e., publicly available
charging stations.
As EV ownership is expanding, there is a growing need for charg-
ing stations available in public destinations. As of the end of 2015, the
number of public charging points deployed in the globe reached around
190k, up from 110k in 2014 and about 50k in 2013 [21]. As of May 2017,
around 16k electric stations and about 43k charging points are available
to the public in North America [22]. These charging stations leverage
multiple sensors such as current sensors to provide better safety and
reliability than residential charging points. These sensors can moni-
tor the power consumed and EVs can be automatically connected or
disconnected according to the power demand status. With this capabil-
ity, public stations can protect the charger and EVs from overheating
and thus potential damages. Further, public charging stations are usu-
ally capable of charging EVs with higher charging rate than residential
charging points, because different charging standards and techniques
are applied to the two different contexts.
Charging standards used in North American mainly follow the SAE
International standard SAE J1772, where four standards are devel-
oped [23]. Table 1.1 shows the specification for the four charging stan-
dards. AC level 1 has the lowest power capacity and is mainly used
in residential charging points, or other application scenarios that EV
users have plenty of time for recharging their vehicles. By contrast,
the other three standards have much higher power capacity and are
used by most public charging stations, or other application scenarios
that there are high expectations on charging rate and charging time.
The fast charging capability of AC level 3 and DC charging can signif-
icantly shorten the charging time of EVs. For example, charging time
for 100 km is about 20 − 30 minutes with AC level 3 and it is about
10 minutes with DC charging.
Fast charging stations, especially those enabled with AC level 3 and
DC charging, introduce high power load on the power grid. There are
two ways that can mitigate this load impact. One way is to schedule and
coordinate EV charging according to the load status of the power grid,
e.g., charging EVs when the power grid has reduced load or reduced
electricity cost. One enabler to coordinating EV charging is that EVs,
charging stations and the power grid can communicate with each other.
The other way is to locally install power generation such as solar power
to loose the need for the power grid, which thus alleviates the load
impact caused by EV charging. Another advantage here is that EVs can
source from the power grid at opportune time to reduce their charging
cost, such as at off-peak time or when the electricity price is high. In
this article, we study both methods as well as the combination of these
two methods.
1.3 Electric Vehicle Charging Overview
The increasing integration of electric vehicles has motivated many re-

search efforts from both industry and academia. These efforts can be
partitioned into two important groups, which respectively address the
two crucial consequences mentioned above. The first group centers
around the power grid oriented consequence, which is the associated
heavy load impact that can compromise the stability and reliability of
the power grid. The second group focuses on the EV user oriented con-
1.3. Electric Vehicle Charging Overview 9
sequence, which is the quality of service degradation caused by EVs’

competition on charging resources including charging points and the
power supply.
Power Grid Oriented EV Charging. Studies falling into this
group address how to mitigate the potential grid impact associated
with large-scale EV charging. Unbalance between charging demand
and power supply will occur if the interaction between EVs and the
power grid is uncontrolled or uncoordinated. This fact is revealed in
[24], which shows that there is a positive dependence between charging
load impact and the penetration levels of EVs. Uncoordinated loads can
cause a series of issues such as power loss, low energy efficiency, fre-
quency deviation, and voltage deviation, which in turn jeopardize the
stability and reliability of the power grid. These issues are investigated
by two major research threads.
The first thread is load flattening and example works include
[4, 25, 26, 27, 28, 29, 30]. In general, these works alleviate the unbalance
between charging demand and power supply through shifting charging
demand to the power valley of the power grid, i.e., valley filling. Flat-
tening power load is further demonstrated to be effective to improve
energy efficiency and power loss of the power grid [6]. The second thread
is frequency regulation and example studies include [31, 32, 33, 34, 35].
A gap between power supply and demand on the power grid causes
the grid frequency to deviate from its nominal value, and the goal of
frequency regulation is to reduce this gap. EVs are capable of mak-
ing rapid response to frequency changes, and thus they are regarded
as suitable for providing frequency regulation service. However, con-
ventionally, power generators that are capable of providing MW-scale
power are utilized for frequency regulation. To meet this power capac-
ity threshold, these existing works investigate the aggregated power by
a large number of vehicles and regard these EVs as a whole to provide
frequency regulation.
The above works focus on addressing the power grid impact. This
confined focus can cause degradation to quality of service to EV users,
i.e., blindly following requirements from the power grid may not result
in beneficial results for EV users. For example, a number of EVs may
have to charge at rather low rates to reduce power load on the grid,
which will compromise user satisfaction. Different with these existing
works, this article considers both the power grid and EV users and
jointly studies the stability and reliability of the former and quality of
service for the latter.
EV User Oriented EV Charging. Works belonging to this group
study how to well handle the competition on charging resources in
order to improve EV user satisfaction or quality of service. EV user
satisfaction depends on key aspects such as charging cost, charging
time, charging rate, and travel distance/time to charging points. There
are two major research threads that present charging schemes to benefit
EV users in terms of the above aspects.
The first thread is opportune charging and/or discharging and ex-
ample works include [36, 37, 38, 39, 40, 41]. In general, these works
charge and/or discharge vehicles at some opportune time. For exam-
ple, charging EVs when the electricity price is low results in reduced
charging cost, and charging EVs when the charging rate is high will
reduce charging time. The second thread is EV routing and examples
studies include [42, 43, 44]. As to the routing problem for fuel-based
vehicles, it usually needs to consider only two factors: fuel price and
travel distance/time to gas stations. As to EV routing, besides charging
price and travel distance/time to charging stations, it needs to consider
another key factor, which is queueing time. The existence of queueing
for EV charging is mainly because of two reasons: EV charging takes
much longer time compared to gas fueling and there is a limited number
of charging points. Thus, there is a much higher opportunity for EVs to
queue and wait for the availability of charging points, and further their
waiting time is usually much longer. To reduce the queueing time and
improve quality of service, it should distribute EVs to charging stations
as evenly as possible. However, to follow routes with reduced queueing
time may conflict with the other goals such as reducing charging cost
and travel distance/time to stations.
These existing works only discuss one or two of the mentioned as-
pects related to user satisfaction. By contrast, we present a holistic
framework that studies multiple aspects including not only all the men-
1.4. A Systematic Solution Framework 11
͙ ^ŽůƵƚŝŽŶ&ƌĂŵĞǁŽƌŬ
dŚĞhƉƉĞƌ>ĞǀĞů
͙ ͙ ΀Śϱ΁ŚĂƌŐŝŶŐĞŵĂŶĚĂůĂŶĐŝŶŐ
dŚĞ>ŽǁĞƌ>ĞǀĞů
͙ ͙ ͙ ͙ ΀ŚϮ΁ŚĂƌŐŝŶŐZĂƚĞŽŶƚƌŽů
΀Śϯ΁ŚĂƌŐŝŶŐĞŵĂŶĚZĞƐƉŽŶƐĞ
΀Śϰ΁ŚĂƌŐŝŶŐ>ŽĂĚ^ĐŚĞĚƵůŝŶŐ
Figure 1.3: A figure illustrating the application scenario and a block diagram
showing the proposed systematic solution framework.
tioned aspects but also new aspects such as the level of users’ urgency
and user behavior’s uncertainty. Further, many works of the first thread
consider a direct interaction between the power grid and charging sta-
tions, and all works of the second thread only consider the interaction
between charging stations and EV users. By contrast, this article dis-
cusses a hierarchical interaction among the power grid, charging sta-
tions, and EV users.
1.4 A Systematic Solution Framework
This article considers an application scenario, as illustrated in Fig. 1.3,

where multiple charging stations are publicly available in an urban area
and many electric vehicles demands charging service from these charg-
ing stations. For this application scenario, we propose a systematic
solution framework that employs a hierarchical operational structure.
There are two levels in the framework.
The lower level determines how to allocate power supply to vehicles
plugged in within a charging station. For this level, we discuss three dif-
ferent cases according to the operation mode of the power grid and the
configuration of charging facilities. The first two cases both investigate
charging facilities that are able to change power rates continuously.
The difference between them is that one case, called charging rate con-
trol, focuses on the normal operation of the power grid, while the other
case, called charging demand response, confines to the demand response

period of the power grid. The two cases leverage time-driven methods
due to the continuity of charging rate regulation. By contrast, the third
case, called charging load scheduling, handles charging facilities with
fixed charging rates. This case adopts event-driven methods to schedule
vehicles to multiple charging points within a charging station.
The upper level decides how to balance vehicles that are demand-
ing charging service among multiple charging stations, which is called
charging demand balancing. The decision is made according to sev-
eral key factors including charging demand of each individual vehicle,
travel distances to charging stations, and detour distances to destina-
tions, charging prices of charging stations, and uncertainty of EV user
behavior. We propose several efficient and effective decision-making al-
gorithms for this level.
1.5 Road Map
Fig. 1.3 depicts the road map of this article. Chapter 2 to Chapter 4
focus on the lower level and study charging rate control, charging de-
mand response, and charging load scheduling respectively. Chapter 5
confines to the upper level and discusses charging demand balancing. In
these chapters, we first design deployable algorithms and mechanisms
and then evaluate the proposed approaches through two complemen-
tary methods. One method is to use theoretical analysis to demonstrate
their performance guarantees, while the other method is to leverage ex-
tensive simulations based on realistic data traces and simulation tools.
We also review existing studies that align with the problems studied in
each chapter and further consider additional dimensions and/or differ-
ent optimization goals. Finally, in Chapter 6, we conclude the article
with summaries of main ideas discussed in the previous chapters.
2
Charging Rate Control
This chapter investigates charging rate control, which addresses the

competition on power supply among EVs served in a charging station.
To guarantee the reliability and stability of the power grid, charging
stations may be sometimes unable to provide enough power supply to
enable maximum charging rates for their serving EVs. This case will
occur more and more frequently as the EV penetration keeps increasing
[4, 9, 45]. Equally allocating such limited power supply among EVs
may not make users equally satisfied because of their different state of
charges (SOC) and battery capacities. For instance, an EV user with a
drained battery (e.g., 20% SOC) may need higher power to be satisfied
than one with a quite full battery (e.g., 80% SOC) [39, 46]. We propose
a general charging rate control approach that agrees with both cases,
i.e., whether users with identical or distinct degrees of satisfaction.
The algorithm efficiently and optimally determines EV charging rates
for maximizing QoS while satisfying the grid stability. The core of the
algorithm is a binary search over a sorted list of marginal cost, which
is mainly inspired by the stationarity of Karush-Kuhn-Tucker (KKT)
conditions. We conduct extensive simulations using realistic data traces
and the result highlights the benefits and importance of our algorithm.
13
14 Charging Rate Control
The rest of this chapter is organized as follows. Section 2.1 presents

recent development related to charging rate control. Section 2.2 pro-
vides system model. Section 2.3 proposes an optimal charging rate
control algorithm. Section 2.4 evaluates the proposed algorithm. Sec-
tion 2.5 gives the summary of this chapter. 1
2.1 Recent Developments
There is a batch of research efforts that study charging rate control.

Most of them fall into the following two categories. The first category
centers around power grid oriented issues and its general goal is to guar-
antee the performance of the power grid. The second category focuses
on EV user oriented issues and its general goal is to benefit and satisfy
EV users. Further, these works have different optimization objectives
such as load flattening, cost minimization, fairness, and so on. In this
section, we investigate these recent developments. The discussion is
first organized according to the categorization and then according to
their optimization objectives.
2.1.1 Power Grid Oriented Rate Control

Direct Load Flattening. Existing works that directly flatten the
power grid’s load usually optimize either the aggregated load or the
variance of the aggregated load.
Aggregated Load Minimization. The authors of [25] aim at minimiz-
ing aggregated load and formulate the problem using a convex program.
In the formulation, the charging rate for each EV is continuously ad-
justable within a specified range, and this setting is similar to that in
this chapter. They solve the formulated problem via a decentralized
algorithm based on gradient projection. Although the algorithm can
alleviate the computation burdens for the charging station, it involves
much overhead about communicating intermediate results among dif-
ferent parties and needs computation support of each party. As to the
application scenario in this chapter, i.e., considering vehicles served
in a charging station, we argue that a centralized approach is more
1
A preliminary version of the discussion in this chapter appeared in [5].
2.1. Recent Developments 15
applicable. The reason is two-fold. First, the number of EVs in a charg-

ing station is not large (usually at most tens of vehicles) and thus the
computation load that puts on the charging station is also limited. Fur-
ther, EV users may be not willing to allow such computation support
or participate into such as a decentralized method.
The authors of [26] extend [25] and study the case charging facilities
that can only support a fixed power rate for each EV. Thus, the deci-
sion variable is interpreted as the start and resume time for charging
an EV with its fixed charging rate. This problem becomes more chal-
lenging since the decision set becomes discrete. They stipulate that
the charging procedure cannot be interrupted once it is initiated. With
this stipulation, it only needs to decide when to initiate charging. The
authors further relax the problem from discrete to continuous through
assuming a probability distribution for all potential feasible charging
profiles. One rationale of the above stipulation is that frequent mode
switching during EV charging can lead to battery degradation. In ad-
dition, another key factor affecting battery degradation is the depth
of charging, i.e., deep charging (e.g., charging to 100% SOC) usually
degrades battery more than shallow charging does (e.g., keeping SOC
between 50% and 80%). In Chapter 4, we will consider this discrete
case, but unlike [26], we leverage more applicable event-driven meth-
ods to schedule EV charging.
Load Variance Minimization. A more intuitive objective for load
flattening is to directly minimize the variance of the aggregated load.
The authors of [47] present comparison about the load variance min-
imization between global and local approaches. The former approach
is for a global aggregator to coordinate EV charging loads, while the
latter one is described as local facilities that are only responsible
for managing the loads within their own residential areas. The two
approaches use global- and local-scale of aggregated information
respectively. We also consider a local scenario in this chapter, where
charging rate control is carried out locally within a charging station.
The evaluation result in [47] demonstrates that both approaches can
effectively flatten the power loads and the global approach has better
performance.
The authors of [27] study local residential areas and deal with load
variance minimization in presence of the uncertainties of renewable
generation. They propose a real-time load control algorithm by shift-
ing the power consumption of deferrable loads (e.g., EV charging) to
time periods with high renewable generation. The algorithm minimizes
the expected load variance with updated predications of renewable at
each time step. The authors prove that the expected variance of the
aggregated load is linearly proportional to the square of the prediction
error. They further show that as time horizon expands, the predic-
tion can become accurate and thus the algorithm’s sub-optimality van-
ishes accordingly. Our charging rate control approach involves renew-
able generation, but does not consider its uncertainties in this chapter.
To demonstrate how the prediction error affects the performance of the
proposed rate control approach, one feasible way is to leverage a similar
analysis to that in Section 3.4.
Note that the above two optimization objectives, i.e., minimizing
the aggregated load and minimizing the variance of the aggregated
load, have been proved to be equivalent [6, 25].
Power Loss Minimization. Uncontrolled and random PEV
charging can cause increased overloads and power losses, especially
at the distribution level. The authors of [3] demonstrate power losses
caused by EV charging in two stages. They first assume perfect knowl-
edge of load profiles and compute optimal charging profile of vehicles
based on historical data. Then, as to the reality with inaccurate pre-
diction of load profiles, they introduce stochastic programming based
analysis that shows the positive relationship between prediction errors
and power losses. The authors of [29] identify the relationship between
power losses, load factor and load variance through formulating prob-
lems to optimize the three objectives respectively. It is demonstrated
that optimizing the three objectives is equivalent to one another if the
distribution system has a single feeder. For practical systems with mul-
tiple feeders, minimizing load variance just approximately minimizes
power losses. These results motivate several follow-up works that fur-
ther address the power loss minimization problem, such as [30, 48].
The authors of [30] propose a real-time smart load control approach,

which is to minimize power losses plus the power generation cost. The
approach defines three different charging time zones according to the
load status of the power grid, where on-peak zones only serve EVs with
high priorities and off-peak zones serve EVs without preference. The
authors of [48] address the same problem as in [30] and present another
solution using fuzzy coordination.
Frequency Regulation. The mismatch between power supply and
demand causes the power grid’s frequency to deviate from the normal
value, and the goal of frequency regulation is to alleviate this mis-
match. Existing works usually leverage financial incentives to motivate
EV users to participate into frequency regulation. The authors of [32]
make several critical observations about EVs providing regulation ser-
vice. First, providing regulation service is different from normal charg-
ing operation, because the payment for the former is based on the
capacity of available power, while the expense for the latter is based on
the amount of actual power dispatched. Then the authors further re-
strict frequency regulation from interrupting charging operations, since
simultaneous charging and regulating may lead to serious generation
oscillations. With these observations, the problem is formulated with
an objective to maximize the revenue obtained by EV users.
The authors of [49] further investigate frequency regulation jointly
provided by EVs and backup batteries. Frequent charging and discharg-
ing can lead to battery depreciation, the objective is to minimize the
usage of backup batteries (and to maximize the usage of vehicles) while
satisfying the required regulation capacity. These two objectives usually
are not compatible with each other and thus the authors leverage game
theory to solve the conflict between the two objectives. The authors of
[35] address frequency regulation in presence of unreliable information
acquisition. They propose a distributed acquisition approach based on
consensus filtering to enable all participated EVs to have consistent
and accurate frequency deviation information. To enable EVs for fre-
quency regulation needs a huge number of participated EVs. This in
turn will cause massive communication between the EVs and the grid
operator, and thus will increase the occurrence of unreliable informa-
tion acquisition. By contrast, the application scenario discussed in this

chapter focus on a charging station locally communicating with served
vehicles. It is likely to exist much less occurrence of unreliable informa-
tion acquisition. This is also one reason that we employ a centralized
approach for charging rate control.
2.1.2 EV User Oriented Rate Control
Charging Cost Minimization. A customer’s top interest may not

be to follow the control from the power grid and satisfy the needs of
the power grid or the aggregator [50]. Indeed, blindly following the
requirement from the grid or the aggregator may not lead to the most
beneficial result for an EV user. The authors of [36] first propose a
global scheduling optimization problem, which minimizes the total cost
of all vehicles that perform charging and discharging during the day.
The optimality is global but the problem is impractical, because it
requires all information from the future base loads, arrival times and
charging demands of all arriving vehicles. They then present a practical
scheduling optimization problem, which aims to minimize the total cost
of EVs in the local group. The advantage is that the local problem is
not only scalable to a large EV population but also resilient to dynamic
EV arrivals. Our charging rate control approach also focuses on a local
charging station and thus possesses the same advantages.
The authors of [51] focus on a goal of minimizing the combinational
cost consisting of both an EV customer’s charging cost and the penalty
of maintaining less sufficient energy to reach each intermediate stops.
Compared to [51], the authors of [40] employ a similar model of individ-
ual EV user’s cost, but study a different objective, which is to minimize
the charging cost for an EV during night time while guaranteeing the
EV fully charged at the end of the horizon. The authors assume the
vehicle-to-grid support and thus an EV owner can receive revenue by
providing regulation service for the smart grid through an aggrega-
tor. The authors of [37] also investigate the vehicle-to-grid support and
provide more detailed analysis about economics for EV charging. The
authors of [41] further expand [40] by integrating stochastic informa-
tion (hourly electricity prices, the hourly regulation service prices and
2.2. System Model 19
hourly regulation signals) and formulate the problem as a stochastic dy-

namic programming problem. The authors of [39] study an online EV
charging problem with a goal of maximizing the total value of served
vehicles minus the energy cost incurred.
These exiting works aim at minimizing the overall charging cost for
EV users, while the focus of this article is to address user satisfaction
or QoS degradation caused by competing for charging resource (here is
power supply of the charging station). The different focuses make our
work different from these existing ones.
Charging Fairness. Existing works that are most related to the
discussion in this chapter include [9, 52]. These two works represent EV
user satisfaction as the fairness about sharing the power supply and
design distributed algorithms to determine charging rates that achieve
proportional fairness among vehicles. There are several key differences
between them and our work. First, our work explicitly considers the
heterogeneity among different vehicles such as different charging limits
and SoCs, which makes the problem more realistic but also more dif-
ficult to solve. Second, as mentioned, centralized approaches are more
suitable to the application scenario of a local charging station, and thus
our work designs a centralized rate control algorithm. Third, our work
provides provides an optimal and efficient rate control algorithm, which
needs no iterations or approximation that exists in the distributed al-
gorithm presented in [9, 52].
2.2 System Model
As depicted in Fig. 2.1, charging station CS has a controller to ad-

just charging rates pk s of EVs within the charging station. We employ
a discrete-time model, where the control horizon is partitioned into a
sequence of time slots of equal length. The time slot length could be
several minutes in practice, e.g., 5 minutes. At each time slot, the con-
troller adjusts the charging rates of all EVs at the charging station. We
consider a single interval for the following algorithm design and anal-
ysis. The whole control process is decoupled into running the designed
algorithm at every time slot. The discrete-time model has been widely
‫ܵܥ‬
ܵ
‫݌‬ଵ ‫݌‬௄
‫݌‬௞
͙ ͙
‫ܸܧ‬ଵ ‫ܸܧ‬௞ ‫ܸܧ‬௄
Figure 2.1: A diagram for charging rate control. S: power supply of charging station
CS; pk : charging rate of EVk .
used in the EV charging study such as [9, 25, 39] and power market
literature such as [53, 54, 55].
Power supply model. Besides the power grid, charging sta-
tions today increasingly have some local renewable energy (e.g., solar
and wind power) available to meet the sustainable development needs
[56, 57]. We thus consider both of the two energy sources: the power grid
and co-located renewable generation. The power grid feeds a charging
station through a transformer. The capacity limit of the feeding trans-
former is denoted as L, which is the maximum power load that the
transformer can serve. Note that the following design and analysis can
be easily adapted to a transformer serving multiple charging stations.
The renewable power generation is denoted as R, which can be from
a mix of different sources such as solar power and wind power. The
charging station first uses the renewable and draws power from the
grid to compensate the insufficiency.
Power load model. The power load on the feeding transformer
can be differentiated into two parts: elastic EV load of the charging sta-
tion and inelastic base load of other consumers [9, 25, 39]. The elasticity
here is based on whether the load can be controlled by the charging
station.
The station is charging K EVs, each of which has a charging rate
pk . The EV load p equals the sum of charging rates of all K EVs, i.e.,
P
p = k pk . In this level, each EV is characterized by two parameters: a
charging rate limit and a utility function. First, EVs have charging rate
limits or maximum charging rates and different types of EVs support
‫ݑ‬
‫ݑ‬ଵ ‫ ݌‬ൌ ݈‫݃݋‬௔ଵ ሺ‫݌‬ሻ
‫ݑ‬ଶ ሺ‫݌‬ሻ ൌ ݈‫݃݋‬௔ଶ ሺ‫݌‬ሻ
‫ݑ‬ଷ ሺ‫݌‬ሻ ൌ ݈‫݃݋‬௔ଷ ሺ‫݌‬ሻ
Ͳ ‫݌‬
Figure 2.2: Illustration of utility functions of three different EVs.
different such limits. For example, Nissan Leaf is able to support 50kw
while Tesla Model S can support up to 135kw [58]. Thus, without loss
of generality, we assume that each EV has a different charging rate
limit, denoted by P k , i.e., pk ∈ [0, P k ]. Second, a utility function is
used to quantify the utility that EVk obtains when it has charging
rate pk . The utility can be seen as how much the EV user is satisfied
with charging rate pk . In addition, according to EV charging studies
such as [46, 39], EV users may have different degrees of satisfaction
even though their EVs have the same charging rates. Two possible
reasons for this are their different state of charge and different battery
capacity. To capture this heterogeneity, we associate each EV with a
different utility function, which is denoted as uk (pk ). We assume that
the utility function is continuously differentiable, non-decreasing and
strictly concave. Although this kind of modeling approach may seem to
be restrictive, it is standard and widely used within EV charging and
power market literature [9, 53, 25, 54, 59, 55, 60, 39].
We provide more discussion about the utility function as follow.
The non-decreasing property captures that the higher power allocated
to an EV, the higher (at least not lower) the user satisfaction will be.
The concave property reflects diminishing return, which is that the
increasing rate of user satisfaction decreases as the allocated power
rises. Fig. 2.2 gives illustration for utility functions of three differ-
ent EVs, which possess the above properties. The figures shows a
typical example for the utility function, i.e., the logarithm function,
u1 = loga1 (p1 ), u2 = loga2 (p2 ), u3 = loga3 (p3 ). For an objective of max-
P
imizing the total utility, u = max i ui , we discuss two different cases:
(i) ai = aj , ∀i, j, and (ii) ai 6= aj , ∃i, j.
Case (i): this case is called proportional fairness. To maximize the

total utility for this case, there must be p1 = p2 = p3 if no other
constraints. This means that users can be equally satisfied.
Case (ii): this case is when users have different degrees of satisfac-
tion. Suppose that a1 < a2 < a3 , which means that EV3 is harder to
satisfy than EV2 , and EV2 is harder than EV1 . To maximize the total
utility for this case, there must be p1 > p2 > p3 if no other constraints.
This means EV1 has more allocated power than EV2 , and EV2 has
more than EV3 .
The base load, denoted as D, is the total power consumed by other
consumers. Since it cannot be controlled by the charging station, the
maximum power supply S that the station can have is S = [L−D]+ +R,
where [a]+ = max{0, a}. The term [L − D]+ is the remaining load the
feeding transformer can serve, i.e., the maximum amount of power that
the station can draw from the power grid. When the base load is larger
than its capacity limit, i.e., L < D, the term would be zero. At this
time, the transformer is already heavily overloaded by the based load,
and thus the station would have no power from the grid.
2.2.1 Problem Formulation

We use a notion of social welfare to define the sum of EVs’ utility. The
goal of the charging rate control is to maximize the social welfare, and
the problem can be formulated as the following optimization problem.
X
max u(p) = uk (pk ), (2.1)
k
s.t. S = [L − D]+ + R, (2.2)
X
pk ≤ S, (2.3)
k
0 ≤ pk ≤ P k , ∀k, (2.4)
where p = (pk )∀k is the vector of charging rates. The objective

Eqn. (2.1) is for social welfare maximization. The constraint Eqn. (2.2)
is the maximum power supply that the charging station can provide to
EVs. The constraint Eqn. (2.3) restricts that the EV load is less than or
equal to the charging station’s maximum power supply. The constraint
Eqn. (2.4) limits the charging rate for each EV.
2.3. Optimal Charging Rate Control 23
2.3 Optimal Charging Rate Control
In this subsection, we design a charging rate control algorithm, which

not only has low algorithm complexity (polynomial time) but also op-
timally determines charging rates of EVs at a charging station. The
algorithm is shown by Algorithm 1. We now first provide brief descrip-
tion for the algorithm and then show its optimality.
Algorithm description. According to the relation between the
station’s maximum power supply and the maximum EV load, there
are three cases. Case (i): S = 0, i.e., when the power supply is zero;
this case happens when the upstream transformer is heavily overloaded
and the renewable generation is zero. Case (ii): S ≥ k P k , i.e., when
P
the power supply can support maximum charging rates for all the EVs;
this case occurs when the transformer is lightly loaded and/or sufficient
P
renewable generation. Case (iii): 0 < S < k P k , i.e., when the power
supply is non-zero and not enough to charge all the EVs at their max-
imum rates. Note that case (iii) is vital because this case would often
occur as the EV penetration keeps increasing [9, 45, 4]. A complete
algorithm should deal with all these cases.
We propose Algorithm 1, which determines the optimal charging
rates for all the three cases. Lines 1 and line 2 deal with the two trivial
cases (i) and (ii) respectively. Case (iii) is more difficult to solve than
the two cases, because the constraint Eqn. (2.3) is active with non-zero
S in case (iii). Inspired by the stationarity of Karush-Kuhn-Tucker
(KKT) conditions, we mainly use a binary search over a sorted list of
marginal utility to solve case (iii). Lines from 3 to 14 solve this case
and we describe them as follows.
Line 3 sorts the upper and lower bounds of each EVk ’s marginal
utility (i.e., the first derivative) in a decreasing order. The while loop
from line 5 to line 12 is the binary search. In each iteration, lines from 7
to 10 derive all pk values accordingly. Then, lines from 11 to 12 decide
to drop which half of the current interval [lf t, rgt]. If the sum of current
pk values is greater than S, i.e., some pk s cannot be assigned as P k but
smaller values to meet the power supply S, then drop the right half by
rgt ← mid. If the sum is less than S, i.e., some pk s cannot be assigned as
0 but larger values to further improve welfare, then drop the left half by
Algorithm 1: Optimal Charging Rate Control

Input : P k , ∀k, S;
Output: pk , ∀k;
1 if S == 0 then
∀k, pk ← 0;
k P k ≤ S then
P
2 if
∀k, pk ← P k ;
0 0
3 Sort all uk (0) and uk (P k ) in a decreasing order, and the sorted
list is denoted as {ϕ1 , ϕ2 , ..., ϕ2K };
4 lf t ← 1, rgt ← 2K;
5 while lf t < rgt − 1 do
j k
6 mid = lf t+rgt
2 ;
7 for k ← 1 to K do
8 if ϕmid ≥ u0k (0) then
pk ← 0;
9 else if ϕmid ≤ u0k (P k ) then
pk ← P k ;
10 else
pk ← x, where u0k (x) = ϕmid ;
P
11 if k pk > S then
rgt ← mid;
12 else
lf t ← mid;
13 Let Yrgt ← {k|u0k (0) < ϕrgt },
Ymid ← {k|uk 0 (0) ≥ ϕrgt ∧ uk 0 (P k ) ≤ ϕlf t },
Ylf t ← {k|uk 0 (P k ) > ϕlf t } ;
14 for k ∈ B do
pk = uk 0 −1 (λ), where
{λ|λ ≥ 0 ∧ k∈B uk 0 −1 (λ) = S − k∈Ymid P k } ;
P P
15 return pk , ∀k;
2.4. Performance Evaluation 25
lf t ← mid. The binary search finally finds an interval [lf t, rgt], which
divides the K EVs into three subsets as shown in line 13. Since the pk s
in subset Yrgt and Ylf t are already determined by the binary search,
line 14 decides the pk s in subset Ymid accordingly. Line 14 attains since
P
the welfare for case (iii) is maximized when k pk equals S.
Since the complexity of the binary search dominates in Algorithm 1,
the algorithm has a polynomial-time complexity of O(KlogK). This low
complexity brings a series of important benefits that advocate using
the algorithm. First, it would cause little overhead to the controller
and allow the algorithm to run at short time intervals (e.g., less than
several minutes). Second, the base load and renewable generation can
be predicted accurately in the short term, i.e., the hourly prediction can
be much less than 5% mean absolute error [61, 62]. This error would
much decrease for an interval of several minutes. Based on the accuracy
of the estimation, the result of the algorithm would deviate little from
reality. This further validates the applicability of the algorithm.
Optimality analysis. Besides the low algorithm complexity, opti-
mality is another good feature of Algorithm 1. The following theorem
states its optimality.
Theorem 2.1. Algorithm 1 optimally determines charging rates for the

K EVs in maximizing the social welfare.
Proof. The proof is in Appendix 2.6.
We see that Algorithm 1 efficiently and optimally solves the opti-

mization problem given by Eqn. (2.1) to Eqn. (2.4). It has lower com-
plexity and thus is better than traditional methods such as interior-
point methods, which have a complexity much higher than the cubic
order of the number of variables (i.e., O(K 3 ), here is the EV number
K) [63].
2.4 Performance Evaluation
To highlight the benefits of our design for charging rate control, we

carry out extensive simulations using realistic data traces. This subsec-
tion presents the experimental results and the analysis for them.
Experimental Setup. We consider a station with 20 fast charging

points, each of which supports a maximum rate of 50kW [58]. There are
20 EVs at this station and each EV plugs into a charging point. Each
EV’s maximum charging rate (P k ) is randomly picked from a range
of [30, 50]kW . The station is equipped with a PV solar panel array
that has a nameplate power of 1M W . We use the solar irradiation
trace from National Solar Radiation Data Base [64] and choose the
data observed at Los Angeles International Airport, CA, US, on Sep
08, 2010. Fig. 2.3(a) shows the hourly power output of the solar panel
array during the whole day. We further use a base load trace from
[65] and choose the data from the service area of Southern California
Edison, with ID of TOU-8-SEC, on Sep 08, 2014. Fig. 2.3(b) depicts
the base load trace. The upstream transformer’s load limit is set as
1.1M W . This experiment assumes an identical utility function for EVs
for ease of presentation, but our algorithm applies to heterogeneous
utility functions. Further, it employs proportional fairness among users
and thus logarithm (log) is used as the utility function.
Experimental Methodology. We compare three different meth-
ods. (i) Opt-Rate. Our optimal charging control algorithm, Algorithm 1.
(ii) Max-Rate. This method imposes no control on the charging rate and
always uses the maximum rate to charge EVs. The upstream trans-
former may be overloaded by this method. (iii) No-RES. This method
achieves optimal charging rates when sourcing only from the grid. We
run simulations independently for each of the 24 data points in Fig. 2.3
and plot the result for each run. For illustration, we use each hour here
(and the corresponding data point) to represent one independent time
slot. The following results explore the benefits of renewable integration
and the importance of charging control.
2.4.1 Benefits of Integrating Renewable.
Fig. 2.4(a) compares the total EV charging rate or user satisfaction

among the three methods. Comparing with No-RES, we observe that
Opt-Rate significantly increases EV charging rate while respecting the
transformer load limit (Fig. 2.4(f)). This is because of the co-located
renewable. To see how the charging station draws energy from different
1.2 0.6
Supply (MW)
Load (MW)
Solar 0.5
0.8
0.4 Base Load
0.4
0.3
0
1 6 12 18 24 1 6 12 18 24
Hour of the Day Hour of the Day
(a) Solar trace (b) Base load trace
Figure 2.3: The solar trace is from Los Angeles International Airport, CA, US,
Sep 08, 2010 [64]; the base load trace is from the service area of Southern California
Edison (ID: TOU-8-SEC), Sep 08, 2014 [65].
sources, we plot figures from Fig. 2.4(b) to Fig. 2.4(d) to demonstrate

the different sources’ power usage for the three methods. First, the
power drawn from the grid, i.e., the grid power, by Opt-Rate and No-
RES decreases (increases) as the base load increases (decreases); while
the grid power by Max-Rate is not affected by the base load. This
observation comes from that when solar supply is zero, i.e., hour 1 to
hour 5 and hour 19 to hour 24, Opt-Rate’s grid power varies with the
base load while Max-Rate does not. The reason is that all methods
except Max-Rate imposes charging control to respect the transformer
load limit. During the same time periods, the transformer is heavily
loaded and the total charging rate by Opt-Rate is less than that by
Max-Rate. Second, the grid power by Opt-Rate and Max-Rate goes
down (goes up) as the solar supply rises (declines) from hour 6 to hour
10 (hour 16 to hour 19). No-RES gives no such observation because
of employing no co-located renewable. From hour 10 to hour 15 when
the renewable generation is sufficient, Opt-Rate behaves the same as
Max-Rate in terms of that both charge EVs with their maximum rates.
2.4.2 Importance of Charging Control.
Fig. 2.4(e) shows the individual EV charging rate comparison between

Opt-Rate and Max-Rate for hour 6, when the transformer is heavily
loaded and there is little solar supply. First, each EV’s charging rate
by Opt-Rate is less than or equal to that by Max-Rate, because the
former method respects the transformer load limit. Further, with the
Sum Crg Rate (MW)

2
Sum Crg Rate (MW)
1
1.6 Grid power Solar supply
0.8
Solar usage Base load
0.6 Max-Rate 1.2
Opt-Rate
0.4 No-RES 0.8
0.2 0.4
0 0
1 6 12 18 24 1 6 12 18 24
(a) Sum rate comparison (b) Opt-Rate
Sum Crg Rate (MW)
Sum Crg Rate (MW)
2 2
1.6 Grid power Solar supply 1.6 Grid power Solar supply
Solar usage Base load Base load
1.2 1.2
0.8 0.8
0.4 0.4
0 0
1 6 12 18 24 1 6 12 18 24
(c) Max-Rate (d) No-RES

Charging Rate (kW)
50 2
Opt-Rate
Max-Rate
Max-Rate Opt-Rate 1.6
Load (MW)
45 No-RES Solar supply

1.2
40
0.8
35 0.4
30 0
1 5 8 10 15 20 1 6 12 18 24
EV number Hour of the Day
(e) Indiv. rate comp (6am) (f) Transformer load
Figure 2.4: Comparisons on sum charging rate (a), separate power usage (b)∼(d),
individual charging rate (e), and transformer load (f).
2.5. Summary 29
limit, EVs’ charging rates are fairly allocated by Opt-Rate, which partly
illustrates the importance of charging control. Another result indicating
the importance is by Fig. 2.4(f). First, although Max-Rate has the
maximum charging rate (Fig. 2.4(a)), it violates the transformer load
limit up to 31%. When the solar supply is inadequate, e.g., hour 1 to
hour 6 and hour 18 to hour 24, the total charging rate by Max-Rate
together with the base load keep overloading the transformer as long
as 13 hours. Second, No-RES always let the transformer work just at
its load limit; while Opt-Rate even makes it lightly loaded when solar
supply is sufficient. This observation further implies the benefits of
using co-located renewable.
From the above observations, we conclude that (i) Max-Rate has
the maximum EV charging rate and thus highest user satisfaction, but
it overloads the power grid a lot, (ii) Opt-Rate maximizes EV charging
rate and user satisfaction while respecting the capacity of the power
grid, and (iii) No-RES results in the least charging rate and thus lowest
user satisfaction among the three approaches.
2.5 Summary
This chapter addresses the competition on power supply among EVs

served in a charging station. We first propose a charging rate control al-
gorithm that optimally and efficiently performs power allocation among
vehicles. The algorithm maximizes EV user satisfaction while guaran-
teeing the reliability and stability of the power grid. We then conduct
extensive simulations using realistic data traces and the result high-
lights the benefits and importance of our algorithm. We finally present
recent developments about charging rate control, which also analyze
relationships and differences between our work and existing ones.
2.6 Appendix
Proof for Theorem 2.1.
Proof. The algorithm obviously finds the optimal charging rates for
cases (i) to (ii), respectively. Thus, we focus on discussing case (iii)
P
(0 < S < k P k ). To prove the optimality for this case, we need to
analyze the following two parts: Part (i) is the binary search by lines
from 5 to 12; Part (ii) is about lines from 13 to 14.
Part (i). The optimization problem given by Eqn. (2.1)-(2.4) is
concave and satisfies Slater’s condition (since all constraints are affine).
Thus, the Karush-Kuhn-Tucker (KKT) conditions are necessary and
sufficient for its optimality [66]. Let λ be the Lagrange multiplier for
Eqn. (2.3), and θk and θk be that for Eqn. (2.4). The Lagrangian is
then
X X
L= uk (pk ) − λ( pk − S)
k k
X X
− θk (pk − P k ) + θ p .
k k k k
The KKT conditions of stationarity, complementary slackness and fea-
sibility are then
uk 0 (pk ) − λ − θk + θk = 0, (2.5)
θk (pk − P k ) = 0, θk ≥ 0, (2.6)
θk (pk − 0) = 0, θk ≥ 0, (2.7)
X
pk = S, λ ≥ 0. (2.8)
k
After the binary search, the set {1, ..., K} is partitioned into three sub-
sets: Yrgt , Ymid , Ylf t , as shown by line 12. These subsets must have the
following properties:
•pk∈Yrgt = 0, where Yrgt = {k|uk 0 (0) < ϕrgt }, (2.9)
•0 < pk∈Ymid < P k , where
Ymid = {k|uk 0 (0) ≥ ϕrgt ∧ uk 0 (P k ) ≤ ϕlf t }, (2.10)
•pk∈Ylf t = P k , where
Ylf t = {k|uk 0 (P k ) > ϕlf t }. (2.11)
P
Note that the property Eqn. (2.10) must hold, otherwise k pk would
be greater or less than S. Based on the above properties, we have the
following deduction.
For all k ∈ Ymid , by Eqn. (2.6) and (2.7), we have θk = 0, θk = 0.
The utility function is concave, by Eqn. (2.5),
λ = u0k (pk ) > 0, ϕrgt ≤ λ ≤ ϕlf t . (2.12)
2.6. Appendix 31
Then, for all k ∈ Yrgt , by Eqn. (2.6) and (2.5), we have
θk = 0, θk = λ − u0k (0) > λ − ϕrgt ≥ 0.
Finally, for all k ∈ Ylf t , by Eqn. (2.7) and (2.5), we have
θk = 0, θk = u0k (P k ) − λ > ϕlf t − λ ≥ 0.
Therefore, all KKT conditions of Eqn. (2.5)-(2.8) are satisfied and the
charging rates for Yrgt and Ylf t are optimal.
Part (ii). This part is to prove that the charging rates for subset
Ymid given by line 13 to line 14 are optimal. Using Eqn. (2.8) and
Eqn. (2.9)-(2.11), we have
X X X
0+ pk + P k = S. (2.13)
k∈Yrgt k∈Ymid k∈Ylf t
Then, based on Eqn. (2.12) and Eqn. (2.13), we can determine the
charging rates for subset Ymid , which are right the results given by line
14.
Hence, according the two parts, the algorithm optimally solves the
optimization problem given by Eqn. (2.1)-(2.4), and thus optimally
determines charging rates of the K EVs.
3
Charging Demand Response
Demand response programs provide incentives that induce costumers

to adapt their electricity demand according to power supply conditions,
for example, to reduce their power loads in response to warning signals
of high price peak or requests from electric power companies. Demand
response thus can help the power grid transit from the paradigm of
supply-follow-demand to the one of demand-follow-supply. Demand re-
sponse is identified as one of the crucial areas of the future smart grid,
especially for incorporating distributed generation (e.g, co-located re-
newable) into the power grid and for serving deferable loads (e.g., elec-
tric vehicle charging) from customers [67, 68].
This chapter considers a notably promising resource for the adop-
tion of demand response: electric vehicles. First, EVs are proliferating
rapidly as mentioned above. Second, EVs cause large charging demand
to the power grid. The charging loads are highly controllable and thus
have a high degree of responsiveness, which also enables significant flex-
ibility for power demand management. Third, charging stations have
become indispensable infrastructure to support the deep penetration
of EVs. They are usually equipped with local power generation such as
renewable, which can help reduce the power need from the grid [56, 57].
32
33
Along with the wide recognition of the demand response potential

of EVs, researchers have begun to focus on new market designs towards
this end, e.g., [53, 69, 70, 71, 72]. However, a clear vision still remains
elusive. First, a key assumption in most existing studies is that EVs di-
rectly interact with the electric power company. Under this assumption,
they confine to designs of single-level markets between the two parties.
However, if this assumption is violated, inefficiency or even infeasibility
emerge in the markets. A canonical example is that EVs served at a
charging station communicate directly with the station instead of the
power company. Further, electric power companies usually prefer to
enable such direct interaction with medium and large consumers (e.g,
charging stations) for the market-based demand response [73, 74]. Fur-
thermore, even if permitting direct interaction with individual EVs, the
inefficiency from both computational and economic perspectives would
become more severe as the market share of EVs continues increasing.
Hence, the focus of this chapter is the hierarchical demand response
for EVs via charging stations.
Second, besides the economic aspects, a thoughtful market design
should also take account of the temporal constraints associated with
charging loads. Due to relatively long charging cycles, EV users usu-
ally leave their vehicles at a charging station and then return to pick
them up after some time [10, 12, 75]. The time is termed as deadline,
which can be explicitly specified by the users or implicitly regarded as
their daily routines such as the off-duty time [10, 12]. Rational users
wish that the charging station serves their charging demand as much as
possible before the deadline. However, solely responding to the warning
signals or requests from the electric power company overshadows the
importance of temporal aspects of EVs. Although this can cause signifi-
cant demand reduction, EVs would end up with little allocated power in
extreme cases. Third, charging stations are usually co-located with re-
newable generation such as wind and solar power [56, 57]. The effective
power output of renewable fluctuates and varies over time due to the
dependence on the external conditions (e.g., solar irradiation and wind
speed) [76, 77]. Both the economic and temporal performance would
be much compromised if overlooking the uncertainty of renewable.
34 Charging Demand Response
To achieve a market design respecting the operation mode of charg-

ing stations, we explore the feasibility and benefits of utilizing two-level
market models. In particular, we propose to design a two-level market
framework, where the first level is between a charging station and the
electric power company while the second level is between the station
and EVs. Finding such as a market mechanism that addresses all the
three key issues mentioned above, however, is a non-trivial task because
we need to address a series of challenges. First, both charging stations
and EVs usually act strategically to maximize their individual benefits.
A suitable market mechanism should be distributed, where both par-
ties are allowed to make decisions independently. Second, the market
framework should be compatible with both the economic and tempo-
ral aspects. Third, from an algorithmic perspective, a well-designed
mechanism should have robust guarantee of the impact of renewable’s
uncertainty on the market.
To address these challenges, we propose a distributed deadline-
aware two-level market mechanism for EV demand response. The mar-
ket mechanism has a hierarchical decision-making structure as follows.
The charging station leads the market by deciding the amount of grid
power that best responds to the demand reduction requests from the
electric power company. EVs follow the charging station’s actions by in-
dependently submitting to the station a unit price bid that best satisfies
its own economic and temporal requirements. The charging station’s
power supply (power drawn from the grid plus co-located renewable
generation) is allocated to EVs in proportion to their bids.
To demonstrate the feasibility and benefits of such a distributed
market mechanism, we study and analyze it using the framework of
Stackelberg games [78]. Through rigorous game-theoretic analysis, we
prove that 1) the market level among EVs converges to a unique Nash
equilibrium where no EV can improve its economic or temporal perfor-
mance by changing only its own bid; 2) the whole market converges to
a unique Stackelberg equilibrium where the charging station achieves
the best economic gain in presence of the Nash equilibrium. Another
key contribution is to bound the impact of renewable’s uncertainty on
the Stackelberg equilibrium. For EVs, we demonstrate that the uncer-
tainty does not impact the Nash equilibrium. For the charging station,
to obtain good economic performance in the average at the Stackelberg
equilibrium, we present a stochastic optimization based algorithm given
the estimation of the likelihood of renewable generation. The algorithm
has provable robust performance guarantee in terms of the variance of
the prediction errors, which is not dependent on any specific distribu-
tion of the prediction error. To be specific, our main contributions in
this chapter are summarized as follows.
• We study the novel scheme of hierarchical demand response for
EVs via charging stations. Specially, we propose a distributed deadline-
aware two-level market mechanism for this scheme.
• Through theoretical analysis, we demonstrate that 1) the pro-
posed market mechanism provably converges to the equilibrium solu-
tions; 2) the renewable’s uncertainty has provable bounded impact on
the equilibriums.
• Using detailed simulations, we show the efficacy for the market
mechanism as well as validate the theoretical results under various set-
tings with regard to both economic and temporal performance.
The rest of this chapter is organized as follows. Section 3.1 presents
recent developments. Section 3.2 provides system model. Section 3.3
proposes a two-level market. Section 3.4 analyzes the proposed market
in presence of renewable. Section 3.5 evaluates the propose approaches.
Section 3.6 gives the summary of this chapter. 1
As mentioned, there is a batch of research efforts on EV charging,

such as flattening grid load [25, 27], improving cost efficiency [30, 79],
charging load scheduling [20, 80], and so on. We mainly discuss existing
works studying market designs for EV demand response, which are
most related to the discussion in this chapter.
Along with the wide recognition of the demand response potential
of electric vehicles, researchers have begun to focus on new market de-
signs towards this end. The authors of [81] propose an auction-based
1
A preliminary version of the discussions in this chapter appeared in [59].
market mechanism between EVs and an aggregator, which supports

that the aggregator participates in the day-ahead and secondary re-
serve sessions. They further identify which variables are necessary to
be predicted and how to include these variables in the optimization
problem. In this chapter, we also carry out a similar job that identifies
the prediction of co-located renewable and integrates the prediction
into a stochastic programming program. The authors of [82] consider
EV demand response as a scalable resource allocation problem and thus
design a proportional allocation based market mechanism. They further
develop a decentralized algorithm based on auction theory to update
bids of vehicles in each iteration. The application scenario studied in
[82] is a parking lot enabled with charging functionality to serve a large
number of EVs. The authors of [19] also consider this park-and-charge
mode and study a similar problem, which is how to price charging
service. In the next chapter, we also study the park-and-charge mode
but investigate a total different problem, which is how to schedule EV
power loads in presence of differentiated charging requirements.
The authors of [83] consider virtual power plants that enable wind
power generators to participate in the electricity market through com-
bining them with EVs. The work uses EVs as a storage medium to
overcome the the intermittent nature of wind power. The proposed
program in [83] focuses on maximizing the profit of a virtual power
plant through optimizing the schedule of power supplied back to the
grid according to wind power production and the available storage.
EVs passively participate the program by following scheduling deci-
sions made by the aggregator. As the compensation, it entitles these
EVs with free charging. Different with [83], the discussion in this chap-
ter allows more freedom to EVs and considers EVs actively participate
into demand response programs. Another difference is that we employ
renewable prediction in the optimization problem and further study
how the uncertainty of renewable affects the demand response mecha-
nism. The authors of [71] consider the same scenario as in [83], where
both EVs and renewable generators partition into demand response
mechanisms. They propose a distributed algorithm based on alternat-
ing direction method of multipliers.
The authors of [84] study a forward market, where users can de-
fer service of pre-specified loads in exchange for a energy price reduc-
tion. They propose a deadline differentiated market mechanism, which
guarantees that the total quantity of energy can be delivered by a user-
specified deadline. The mechanism connects deadlines to energy prices,
i.e., the longer the deadlines are, the lower the energy prices are. One
feature of the proposed mechanism is that there is an efficient com-
petitive equilibrium between supply and demand. Similar to [84], the
authors of [53] also propose market mechanisms for demand repones
and characterize equilibria for the proposed mechanisms. Two key dif-
ferences between them are that the mechanisms in [53] leverage supply
function bidding instead of deadline differentiated pricing, and the au-
thors of [53] further present distributed demand response algorithms
that iteratively compute the equilibria. The authors of [69] extends to
consider power loads from different household appliances. The authors
in [70] focuses on the impact of charging EVs on distribution circuits of
residential areas and propose a demand response mechanism to help the
distribution circuit to accommodate EV penetration. The mechanism
only sends the demand limit to customers, which allows customers to
control their own power loads according to customers’s preference and
thus largely respects customers’ freedom.
These existing works mainly focus on a single market level and only
consider the economic aspects of EV demand response. Our work is
different from them and the novelty mainly lies in the following two as-
pects. First, we novelly studies a hierarchical demand response scheme.
Besides the level between a charging station and EVs, we additionally
consider the market level and interaction between the charging station
and the electricity company. Second, we take account of two additional
new dimensions into the market design. One is integrating temporal
aspects, because they are critical for users to decide their bids and
for the charging station to decide its power supply. The other is con-
sidering the uncertainty of renewable, which considerably affects the
performance of market mechanisms.
PowerGrid Collocated
Renewable
b1 xN
x1 bi xi bN
... ...
(C1,D1) (Ci,Di) (CN,DN)
PowerLines CommunicationLinks
Figure 3.1: N EVs bid for the power supply of a charging station that sources
power from the power grid and co-located renewable.
3.2 System Model
The scenario we consider in this chapter is that an electric power com-

pany requests power load reduction from participants of the demand
response (DR) program during a time period of T (e.g., one hour).
We focus on a charging station, as a DR participant, who responses
to the request by strategically controlling charging power allocated to
EVs at the station. Fig. 3.1 illustrates the schematic of the charging
system. The charging station sources power both from the power gird
(offered by the electric power company) and the co-located renewable
generation (owned by the charging station).
3.2.1 EV Model with Deadline
The charging station are serving N EVs during the demand response
(DR) period [0, T ). At current time t = 0, each EV i has a remaining
charging demand Ci and an absolutely deadline Di . As mentioned, the
deadline Di can be user-specific time when the user will pick up her
vehicle. Each user requires her vehicle fueled with a specific amount of
energy after the deadline. The remaining demand Ci is the amount of
energy that is still needed at current time to fulfil the user’s demand.
We assume a linear relation between the charged energy and charging
power allocated to an EV [26, 85]. Thus, at time t = T , the remaining

demand will become to Ci −xi T , where xi ≤ P is the allocated charging
power during [0, T ) and P is the charging limit that the charging station
can support. We define the urgency si for each EV i as
Ci − xi T
si = − Di , (3.1)
P
where we assume that EVs are charged with power P after the DR
period. Note that if si > 0, the urgency denotes the lag time of EV i
at time t = T ; if si < 0, the absolute value |si | is the slack time at time
t = T . For ease of presentation, let Ωi = CPi − Di , which represents the
original urgency (if without DR) at the current time t = 0, and then
rewrite the urgency by Eqn. (3.1) as si = Ωi − PT xi .
During the DR period, since the charging station carries out power
load reduction, its power supply may be less than N P and thus not
enough to make all EVs charged with P . In other words, EVs compete
for the station’s power supply to have large power allocation xi and
thus reduce their urgencies si . To have a desirable power allocation, at
time t = 0, each EV i strategically submits a unit power price bid bi > 0
to the charging station. EV i thus needs to pay bi xi to the station. A
larger bid results in more power (thus lower urgency) allocated to an
EV but also causes a higher payment by the EV. Generally speaking,
EV users wish to have both more allocated power and lower payment.
We use a disutility function for each EV i to describe its dissatisfaction
with the urgency and payment. The disutility of EV i is defined as
ui (bi , b−i , y) = U (si + λbi xi ), (3.2)
where b−i is the vector of bids excluding bid bi ; λ ≥ 0 captures how to

weigh between urgency and payment. The function U (·) > 0 is assumed
to be continuously differentiable, strictly increasing, and convex. Note
that this kind of modeling may seem to be restrictive, but it is a stan-
dard form and widely used within the network economics [86, 87, 88]
and power markets literature [53, 89, 84, 54].
3.2.2 Charging Station Model with Renewable

Before the DR period begins, the charging station allocates power P
to each EV and thus has power supply N P . During the DR period, we
denote by y and χ the power drawn from the power grid and the re-
newable power generation respectively. Thus, with respect to the power
grid, the load reduction by the station is N P − y during the DR pe-
riod. The station entirely allocates the power supply to the EVs, i.e.,
P
y + χ = i xi . Following the principle of proportional sharing, the
power xi allocated to each EV i is proportional to its bid, that is,
bi
xi = (y + χ) P . (3.3)
i bi
There are usually lower and upper bounds on the unit power price.
First, we assume that there is operational cost for providing such charg-
ing service, such as purchasing grid power y and operating charging
facilities. We assume that its selling price to the EVs is not less than
the unit operational cost B, i.e., bi ≥ B. Second, there is an upper
bound B of the unit price, i.e., bi ≤ B which can be set as the highest
price that EV users can take. This biding range, denoted as Bi , i.e.,
Bi = [B, B], is called the set of strategies for EV i.
Our design of the utility for the charging station consists of two
terms. The first term represents the compensation offered by the elec-
tric power company for the load reduction. The second term is the
received revenue from charging EVs. This design aims to balance be-
tween the two terms. The utility of the charging station is defined as
X
g(b, y) = G(N P − y) + bi xi , (3.4)
i
where G(·) is continuously differentiable, non-decreasing and concave,
and y ∈ [0, N P ]. Similarly, this kind of utility modeling is standard
and widely used when designing schemes as to network economics [86,
90, 91] and power markets [53, 89, 84, 54, 69].
3.3 The Stackelberg Game Between A Charging Station and EVs
The studied demand response problem includes two hierarchies. The

first hierarchy is that the charging station determines its grid power y
3.3. The Stackelberg Game Between A Charging Station and EVs 41
to maximize the utility. The second hierarchy is that EVs decides their
bids bi to minimize the individual disutility. Both the station and EVs
make their own decisions separately, i.e., without any cooperation. We
model this hierarchial noncooperative problem as a two-level Stackel-
berg game [78]. In this section, we omit the renewable generation, i.e.
setting χ = 0. We will consider it and study the effect of its uncertainty
on the Stackelberg game in Section 3.4.
In general, each player in a Stackelberg game is rational and self-
ish, i.e., aims to maximize its own utility or minimize its own disutility.
Such a game has two types of players: leader(s) and followers. The lead-
ers first choose their own strategies and then the followers determine
their strategies. On one hand, the followers know that the strategies
adopted by the leader(s) and thus they can adapt their own strategies
accordingly. In particular, due to the rationality of players, each fol-
lower will choose its best strategy (i.e., the one with the maximized
utility or minimized disutility), given the strategies of leaders. The
best strategy is also known as the best response. On the other hand,
the leaders also know the fact that the followers will choose their best
responses. Thus, the leaders will determine their best strategies based
on the best responses of the followers. The Stackelberg equilibrium is
usually regarded as the solution to this leader-follower game.
For our problem, we consider the charging station as the only leader
and EVs at the station as followers. In the EV-level game, each EV in-
dependently chooses its strategy, i.e., each EV i submits its unit power
price bid bi to the station. Given the strategy of the station, i.e., the
grid power y, its aim is to minimize the disutility defined in Eqn (3.2).
Because the EVs do not know the strategies of each other, the Nash
equilibrium is the solution of the EV-level game. The set of strategies in
the Nash equilibrium has the property that none of the EVs can reduce
its own disutility by using a different strategy, given the strategies of
the other EVs and the strategy of the station. In the station level, the
station decides its strategy, i.e., the grid power y, to maximize its utility
defined in Eqn (3.4). To find the Stackelberg equilibrium in this two-
level game, we leverage the backward induction technique [92]. That
is, we first find the best response of each EV in the EV-level game, i.e.,
the Nash equilibrium, and then based on this, optimize the utility of
the charging station accordingly.
3.3.1 Nash Equilibrium among EVs

In this subsection, we first present several key definitions used when
to find the Nash equilibrium, then prove its existence and uniqueness,
and finally propose the Nash equilibrium.
Definition 3.1 (Best Response for EVs). The best response Ri (b−i , y)
of EV i is the best strategy for EV i, given the other EVs’ strategies
b−i and the station’s strategy y, i.e.,
Ri (b−i , y) := arg min ui (bi , b−i , y), ∀i. (3.5)
bi
Definition 3.2 (Nash Equilibrium). A Nash equilibrium for the nonco-

operative game among the EVs is a profile of strategies b∗ = (b∗i )∀i
such that, given the stations’s strategy y,
b∗i = Ri (b∗−i , y), ∀i, (3.6)
which is the best response of EV i to the best responses of all the other
EVs.
Definition 3.3 (Standard Function [93]). A function f (z) =
(f1 (z), ..., fN (z)), where z = (z1 , · · · , zN ), is a standard function, if
for all z ≥ 0 the following properties are satisfied.
• Positivity: f (z) > 0;
• Monotonicity: If z ≥ z0 , then f (z) ≥ f (z0 );
• Scalability: For all α > 1, αf (z) > f (αz).
Lemma 3.1. The best response given in Eqn. (3.5) of EV i is a standard
function of b−i .
Proof. All proofs of this chapter are in Appendix 3.7.
We now prove the existence and uniqueness of the Nash equilibrium.

Note that Lemma 3.1 is key to show the uniqueness.
Theorem 3.2. There is a unique Nash equilibrium for the noncooper-
ative game among EVs.
Based on the proof for Theorem 3.2, the Nash equilibrium b∗ =

(b∗i )∀i
is the solution to the equation b = R(b, y) in the feasible set of
bi ∈ Bi , ∀i. We give a close-form expression of the Nash equilibrium as
follows.
Theorem 3.3. The Nash equilibrium b∗ = (b∗i )∀i of the noncooperative

game among EVs is given by
" #B
T (N − 1)
b∗i = , ∀i, (3.7)
λP (2N − 1) B
where [x]aa21 = max(min(x, a2 ), a1 ).
From the Nash equilibrium, we have several observations as fol-

lows. Firstly, if not considering urgencies or deadline, i.e., λ = +∞,
EVs would bid as low as possible to reduce their payments. The asso-
ciated result is thus little power allocation and significant degradation
of temporal performance. This observation reflects the necessity and
importance of taking account of temporal aspects of EV demand re-
sponse. Moreover, we observe from Eqn. (3.7) that b∗i , ∀i increases as
N increases (within the set Bi ), i.e., EVs have to bid higher if there are
more EVs competing for the station’s power supply. When N = 2, EVs
h iB h iB
T T
has the smallest bid 3λP B
. The bid is limited by 2λP B
, which is at-
tained as N tends to the infinity. Another observation from Eqn. (3.7)
is that the Nash equilibrium does not rely on y, i.e., not depend on the
station’s decision on its power supply. This property much eases the
analysis on the effect of renewable generation in the Stackelberg game
(see details in Section 3.4). Lastly, all EVs bid the same by Eqn. (3.7),
which counters the intuition that EVs with different urgencies Ωi should
bid differently. This observation follows because of the selfish behavior
of players (EVs). We will come back to this point with more explanation
in the next subsection.
3.3.2 Efficiency of the Nash Equilibrium

The efficiency of a noncooperative game is defined as the performance
gap between its Nash equilibrium and the Pareto optimality [94]. Pareto
optimality in our context is a state of the solution in which it is impos-

sible to make any EV with decreased disutility without making at least
one EV with increased disutility [66]. The selfish behavior of players
(EVs) causes degradation in the efficiency of a Nash equilibrium. There
are two widely-used concepts to describe this efficiency. Price of anar-
chy (PoA) is defined as the ratio between the worst equilibrium and
the Pareto optimum. Price of stability is defined as the ratio between
the best equilibrium and the Pareto optimum. Since our game has a
unique Nash equilibrium, the two concepts are equivalent here.
The centralized problem that assumes EVs are cooperative to min-
imize their total disutility F , i.e., to achieve Pareto optimality, is for-
mulated as
X
min F = ui (b, y),
i
s.t. bi ∈ Bi , ∀i, (3.8)
where b = (bi )∀i and we rewrite ui (bi , b−i , y) as ui (b, y). The optimal
solution to Eqn. (3.8) is denoted by b̂ = (b̂i )∀i . Hence, given the charg-
ing station’s strategy y, the price of anarchy of the Nash equilibrium
is
ui (b∗ , y)
P
P oA = Pi . (3.9)
i ui (b̂, y)
The rest of this subsection presents the analysis of difference between

the Nash equilibrium and the Pareto optimum. We consider the follow-
ing two different cases for the disutility function U (·): 1) U (·) is linear,
i.e., U 00 (·) = 0; 2) U (·) is strictly convex, i.e., U 00 (·) > 0.
Linear U (·)
Based on Eqn. (3.26), we have
T y λy i b2i
P
X X
F = zi = Ωi − + P .
i bi
i i P
Note that if some EVs have negative disutility, we can add an offset
to the disutility of all N EVs and align them all to have non-negative
disutility. This alignment does not change the optimal solution. Since
∂F
∂bi > 0, F is an increasing function with respect to bi . Hence, when
bi = B, ∀i, the total disutility F is minimized, i.e., b̂ = (B)∀i . The
closed-form expression of the price of anarchy in this case is
H + λyb∗
P oA = , (3.10)
H + λyB
where b∗ = b∗1 = · · · = b∗N and the total urgency H = i Ωi − TPy . Note
P
that b∗ is bounded according to Eqn. (3.7), and thus the P oA is also

bounded in this case.
Due to the linearity of U (·), U 0 (z1 ) = · · · = U 0 (zN ), which means
all EV contribute equally to the minimization of the total disutility
in spite of their different urgencies. Thus, the optimal solution b̂ is
uniform, i.e., b̂1 = · · · = b̂N .
Strictly convex U (·)

Due to U 00 (·) > 0, EVs with different urgencies have different U 0 (zi ), i.e.,
they contribute differently to minimizing the total disutility. Thus, in
general, the optimal solution b̂ is nonuniform. This can be also observed
from the stationarity of the Karush-Kuhn-Tucker (KKT) conditions
in Eqn. (3.12). To obtain this stationarity condition, we rewrite the
problem in Eqn. (3.8) as
X X T y bi b2
min F = ui = U (Ωi − + λy i ),
i i
X P w w
s.t. bi ∈ Bi , ∀i, w= bi . (3.11)
i
We use an intermediate variable w to decouple the disutility of different
EVs. Define its Lagrangian as
X T y bi b2 X
L(bi , θ, δ) = + λy i ) + δ(w −
U (Ωi − bi ).
i P w w i
The stationarity condition with respect to bi is thus

∂ui T y bi b2 Ty bi
= U 0 (Ωi − + λy i )(− + 2λy ) = δ, ∀i. (3.12)
∂bi P w w Pw w
∂ 2 ui ∂ui
Since ∂b2i
> 0, ∂bi is strictly increasing with respect to bi . Hence, for
∂ui ∂uj
any {bi , bj |i 6= j, B < bi < B, B < bj < B}, to make ∂bi = ∂bj = δ, it
must hold bi 6= bj , which validates the above observation. This observa-

tion also follows our intuition that EVs with different urgencies should
have different unit power prices. By contrast, according to Theorem 3.3,
all EVs submit the same bid, which contradicts to the intuition. This
is because the selfish behavior of EVs in the noncooperative game, i.e.,
EVs have no cooperation and each one wants to minimizes its own
disutility.
We now analyze the PoA in this case. We use (ŵ, b̂) and F (ŵ, b̂)
to denote the optimum and the optimal value to the problem in
Eqn. (3.11) respectively. The following lemma gives the upper and lower
bound of F (ŵ, b̂).
Lemma 3.4. F (ŵ, b̂) is bounded by [L, Q]:


Ty e
b b2
λye T
i U (Ωi − P N B + N B ), if B ≤ P λ ,
 P

L=
Ty e
b b2
λye T
−
 P
i U (Ωi + ), if B > ,

P NB NB Pλ
B
Ty λyB T
X
Q= U (Ωi − + ), where eb = .
i NP N 2P λ B
Based on Theorem 3.3 and Lemma 3.4, it is easy to find the bound
for the PoA in this case. It is well know that the PoA of a noncooper-
ative game is always greater than or equal to one [94]. Thus, the PoA
in this case is bounded by
P ∗ P ∗ , y)
i ui (b , y) i ui (b

, . (3.13)
Q 1 L 1
As mentioned, Price of Anarchy (PoA) measures how the efficiency

of a game degrades due to selfish behavior of the players, i.e., the
performance gap between Nash equilibrium and Pareto optimality.
Eqn. (3.13) right bounds the performance gap, i.e., bounds the disutil-
ity difference between Nash equilibrium and Pareto optimality.
3.3.3 Stackelberg Equilibrium of the Two-level Game

This subsection presents the analysis of the two-level game between the
charging station and EVs. Given that the station knows the profile of
strategies b = (bi )∀i for EVs, to find the optimal strategy, the station
only needs to maximize its utility by Eqn. (3.4). Plugging Eqn. (3.3)
into the utility function, we have
P 2
b
g(b, y) = G(N P − y) + Pi i y. (3.14)
i bi
Given Eqn. (3.14), we have the following definitions.
Definition 3.4 (Best Response for the Station). The best response of
the station is the best strategy for the station given the EVs’ strategy
profile b, i.e,
Y (b) = arg max g(y). (3.15)

y
Definition 3.5 (Stackelberg Equilibrium). A Stackelberg equilibrium of

the two-level game between the station and EVs is a strategy y ∗ for the
station and a profile of strategies b∗ = (b∗ )∀i for EVs, such that
y ∗ = Y (b∗ ), and b∗i = Ri (b∗−i , y ∗ ), ∀i. (3.16)
Using the backward induction technique in [92], i.e., replacing bi in

Eqn. (3.14) with b∗i at the Nash equilibrium given by Eqn. (3.7), we
have
g(b∗ , y) = G(N P − y) + b∗ y. (3.17)
The second derivative of g(b∗ , y) with respect to y is G00 (·), which is

negative. Thus, setting the first derivative to zero for maximization, we
have
h iN P
y ∗ = N P − G0−1 (b∗ ) , (3.18)
0
where G0 (·) is the first derivative and G0−1 (·) is its inverse function.
Since the noncooperative game among EVs has a unique Nash equilib-
rium and y ∗ is also unique due to the concavity of g(b∗ , y), we have
the following theorem.
Theorem 3.5. The two-level Stackelberg game between the station and
EVs has a unique Stackelberg equilibrium (y ∗ , b∗ ).
3.3.4 Further Discussions

In practice, there is a constraint that the allocated power to each EV i
is less than the charging limit, i.e., xi ≤ P , ∀i. We omit this constraint
when finding the Stackelberg equilibrium and solving the Pareto opti-
mum. We now consider this constraint and discuss how to adapt the
above solutions.
For the Stackelberg equilibrium, the constraint is satisfied by na-
ture. At the Nash equilibrium, all EVs submit the same bid b∗ given by
Eqn. (3.7), and thus they equally share the power supply y. Since y is
bounded by N P , give the allocation rule in Eqn. (3.3), the constraint
xi ≤ P , ∀i is satisfied.
The Pareto optimum is achieved by solving the centralized problem
given by Eqn. (3.8). When U (·) is linear, the optimal solution is uni-
form, i.e., b̂1 = · · · = b̂N . EVs in this case also equally share the power
supply y, and thus the constraint xi ≤ P , ∀i is satisfied by nature. When
U (·) is strictly convex, the optimal solution is generally nonuniform and
thus we need to consider the constraint when solving this centralized
problem. One way to incorporate the constraint, i.e, xi ≤ P , ∀i, is to
add a barrier item [66], so the centralized problem becomes
X
min ui (b, y) + ln(P − xi + 1) , (3.19)
bi ∈Bi ,∀i i
where ln(·) is the natural logarithm and ‘+1’ allows xi = P . The term
ln(P − xi + 1) is the barrier item. A similar analysis method to that
of Lemma 3.4 can be then applied to derive the bound of the total
disutility by Eqn. (3.19).
3.4 The Stackelberg Equilibrium with Co-located Renewable
We assume that a charging station is co-located with renewable gen-

eration such as wind and solar power [56, 57]. This power generation
provides further flexibility for optimizing the station’s power supply
and thus its utility. Meanwhile, the uncertainty of renewable brings
difficulty in determining its optimal strategy. In this section, we ad-
dress this difficulty and study the Stackelberg game in presence of the
renewable generation.
3.4. The Stackelberg Equilibrium with Co-located Renewable 49
3.4.1 Stochastic Modeling of Renewable Generation

At the beginning of the DR period, the charging station predicts the
renewable generation, using which it determines its strategy on the grid
power y. We use a prediction error to capture the renewable’s uncer-
tainty, that is, χ = (1 + )χ,
e where χ e is the predicted power output
and χ is the actual output of the renewable generator. Following the
standard assumptions in statistics, we assume unbiased prediction, i.e.,
the mean E[] = 0, and denote the variance V[] = σ 2 [95, 27]. With-
out loss of generality, we assume that the capacity of the renewable
generator is less than N P , i.e., χ ≤ N P and χ e ≤ N P . In addition,
according to Theorem 3.3, at the Nash equilibrium, the bids b∗ of EVs
do not depend on the station’s strategy, and thus they are not affected
by the renewable’s uncertainty. By this property, we only need to an-
alyze the effect of the uncertainty on the charging station’s strategy.
The station’s utility function with respect to the predicted renewable
generation is
g(b∗ , y) = G(N P − y) + b∗ (y + χ).

e (3.20)
We use ye∗ denotes its optimal solution and y ∗ to represent the opti-
mal solution with respect to the actual power output (replacing χ e in
Eqn. (3.20) with χ). Then, it is easy to have
h iN P −χe
ye∗ = N P − G0−1 (b∗ ) , (3.21)
0
h iN P −χ
y ∗ = N P − G0−1 (b∗ ) . (3.22)
0
Hence, the resulting utilities for the two cases are
g(b∗ , ye∗ ) = G(N P − ye∗ ) + b∗ [ye∗ + χ]N P , (3.23)

∗ ∗ ∗ ∗ ∗
g(b , y ) = G(N P − y ) + b (y + χ). (3.24)
There are two key points needed to be noted on Eqn. (3.23). First, the
predicted renewable generation χ e is only used to determine the grid
power ye∗ . When counting the actual utility obtained by the station, the
actual generation χ is used, as shown in the last term of Eqn. (3.23).
Second, since ye∗ is determined by the predicted output χ,e ye∗ + χ may
be greater than N P . Hence, we trim it into the corresponding range

and the surplus generation will be discarded because of the prediction
error.
3.4.2 Competitive Analysis

It is clear that the performance of Eqn. (3.23) depends on the accu-
racy of the predicted power output, thus it is important to charac-
terize this dependence. We leverage competitive analysis to analyze
the performance of Eqn. (3.23). The competitive ratio η here is de-
fined as the ratio between the expectations of the above two utilities
(Eqn. (3.23)(3.24)):
E [g(b∗ , y ∗ )]
η= . (3.25)
E [g(b∗ , ye∗ )]
In the following, we first give two useful lemmas (Lemma 3.6 and
Lemma 3.7) and then bound the competitive ratio η (Theorem 3.8).
Lemma 3.6. Given that the expectation and variance of the prediction
error of renewable generation are E[] = 0 and V[] = σ 2 , we have
E[+ ] ≤ σ2 , E[− ] ≥ − σ2 , where + = max(0, ) and − = min(0, ).
Lemma 3.7. For ye∗ given by Eqn. (3.21) and y ∗ given by Eqn. (3.22),
we have y ∗ = ye∗ + ϕ, and ϕ = ϕ+ + ϕ− , where ϕ+ =
+ χe 0
e − G0−1 (b∗ ) 0 , ϕ− = G0−1 (b∗ ) − χ − χe.

χ
This lemma can be proved by comparing Eqn. (3.21) and

Eqn. (3.22) for each case (+ and − ). We omit the proof here due
to space limit. We now give the main theorem of this section as fol-
lows. The theorem bounds the competitive ratio and demonstrates that
the station’s utility is close to the optimal if the prediction of renewable
generation is accurate.
Theorem 3.8. Given that the variance of the prediction error of re-
newable generation is bounded by σ 2 , the utility given by Eqn. (3.23)
e(G0 (N P −e
χ y ∗ )+b∗ )
has a competitive ratio of 1 + Aσ, where A = 2g(b∗ ,e
y∗ )
. That
E[g(b∗ ,y ∗ )]
is, η = E[g(b∗ ,e
y ∗ )]
≤ 1 + Aσ.
It is worth noting that Theorem 3.8 does not rely on any assump-
tions on the distribution of the prediction error other than the bounded
variance. One key observation from the theorem is that the competi-
tive ratio is a linear function of the standard deviation of the prediction
error. Its robustness is also reflected by the observation that this com-
petitive ratio decreases to 1 when the prediction error decreases to 0.
Till this point we have discussed the Stackelberg game between the
charging station and EVs during a demand response period. We have
also provided analytic guarantees for the Nash equilibrium among EVs
and for the Stackelberg equilibrium in presence of the uncertainty of
renewable generation. In this section, we evaluate the solutions using
detailed numerical simulations to obtain a better picture of urgencies
and payments of EVs and load reduction of the station in the game.
3.5.1 Experimental Setup
We consider a charging station that responds to a demand response

request with the duration of T = 0.5 hour. There are N EVs being
charged, and for each EV i, the charing demand Ci is uniformly gener-
ated from a range of [10, 60] kwh. This range has the same scale with
the real EV’s battery capacity, e.g., 24kwh for Nissan Leaf [96] and
70-85kwh for Model S of Tesla Motors [97]. The deadline Di is chosen
from the uniform distribution in [0.5, 5] hour. We consider two cases for
the disutility function: 1) a linear function ui = si + λbi xi + 10 (adding
10 to align all disutility to be non-negative); and 2) a natural expo-
nential function ui = exp(si + λbi xi ). The charging station supports a
maximum charging power of P = 7.2kw, which is the Level 2 charging
and is widely deployed in public places [98]. The range of unit power
price is set as B = 0.05$/kw and B = 0.5$/kw, which has the same
scale with the electricity price in the power market. The utility function
of the station uses a natural logarithm g = ln(N P − y) + i bi xi .
P
1.8 1.8
PoA & Bid
PoA & Bid

1.2 1.2
PoA-exp-upper
PoA-exp-upper
PoA-exp
0.6 0.6 Bid PoA-exp
Bid PoA-exp-lower
PoA-exp-lower
PoA-linear
PoA-linear
0 0
2 4 6 8 10 12 14 16 18 20 0 0.2 0.4 0.6 0.8 1
EV number (N) O
(a) PoA varying with N when λ = 0.2. (b) PoA varying with λ when N = 20.
Figure 3.2: Comparison on PoA for the linear and exponential case with varying
the EV number N and λ. PoA-exp-upper and PoA-exp-lower: the upper and lower
bound given by Eqn. (3.13) with the natural exponential function.
3.5.2 Results on the Nash Equilibrium
Since the renewable generation does not affect the Nash equilibrium,
we do not consider it in this subsection. We first analyze how the num-
ber of EVs (N ) and weight λ affect the price of anarchy (PoA) and
Nash equilibrim (NE), and then validate the difference between NE
and Pareto optimum.
Price of anarchy. Fig. 3.2 illustrates PoA varying with the EV
number N and the weight λ. First, by Fig. 3.2(a), as the number N
increases, the PoA of both the linear and exponential cases rises, which
means the gap between the NE and Pareto optimum becomes larger.
This observation follows the fact that in a non-cooperative symmetric
game, the PoA increases, i.e., the efficiency of NE decreases, as the
number of players (EVs here) rises [99]. Different with the above ob-
servation, by Fig 3.2(b), we can see that the PoA first increases and
then decreases as the weight λ grows. The reason is as follows. When
λ is small, EV users weighs urgency more than payment. Thus, they
should bid higher (i.e., greater than B) to reduce their own urgency,
and the bid would result in larger PoA. However, the bid is outside the
feasible set and thus invalid. After projecting the original bid as B, the
PoA will decline instead. Moreover, the more difference is between the
original bid and B, the smaller PoA will be. Hence, the PoA first rises
as λ grows. After the original bid falls into the feasible set, EV users
put more and more weight on payment than urgency as λ grows. They
thus bid lower and lower, and the PoA declines accordingly.
Second, by Fig. 3.2(a), as the EV number grows, the PoA of the
two cases increases slower and slower. This is because that the bid at
NE increases slightly and the difference between it and the unit price
at Pareto optimum declines. Similarly, by Fig 3.2(b), when the original
bid lies in the feasible set (i.e., needs no projection), the PoA declines
slower and slower as λ increases. Third, as shown in both Fig. 3.2(a)
and Fig 3.2(b), the curve PoA-exp lies between the curves of PoA-exp-
upper and PoA-exp-lower, which validates the upper and lower bounds
given by Eqn. (3.13) for the PoA of the exponential case.
The bid and urgency at NE. To see how the bid and urgency at
NE varies with the EV number N , we first generate the parameters for
all 20 EVs and then run simulation for each of the first N ∈ {2, 4, ..., 20}
EVs. This guarantees that each simulation run includes the same set-
tings of all EVs in the pervious run. Note that the linear and exponen-
tial cases have the same results on the bid and urgency at NE. Fig. 3.3
shows the results. By Fig. 3.3(a), the urgency declines as the EV num-
ber grows. The reason is that to maximize the station’s utility, the grid
power increases as the bid rises. Thus, more power can be allocated to
EVs, which results in the reduced urgency. For example, when N = 10,
the allocated power for each EV is 56.8/10 = 5.68kw; when N = 20, it
becomes larger and equals 129.2/20 = 6.46kw. In addition, the urgency
slows down its growth as the bid curve flats gradually. In contrast, by
Fig. 3.3(b), the urgency increases as the weight λ rises. This is be-
cause putting more weight on the payment leads to a lower bid and
thus smaller grid power. Hence, less power can be allocated for charg-
ing EVs. For instance, when λ = 0.2, each EV has an allocated power
of 66.0/10 = 6.60kw; when λ = 0.5, it decreases to 56.8/10 = 5.68kw.
Moreover, as the bid goes to its boundary, the urgency stays unchanged
anymore, e.g., when the bid equals B = 0.5 or B = 0.05.
Nash equilibrium vs. Pareto optimum. Since for the linear
case, the unit prices at Pareto optimum are identical across all EVs (see
Section 3.3.2), we focus on the comparison between NE and Pareto op-
timum for the exponential case. Fig. 3.4 demonstrates the comparison
0.08 1.6
Urgency Bid Grid power
0.07 1.2
Urgency
Bid
0.06 0.8
114.8 129.2
71.3 85.8 100.3
0.05 56.8 0.4
27.4 42.2
0 12
0.04 0
2 4 6 8 10 12 14 16 18 20
EV number (N)
(a) The bid and urgency of EV 1 varying with N when λ = 0.5.
0.5 1.1
Urgency
0.3 1
Bid Grid power
Urgency
Bid
0.1 69.6 0.9

0 70 69.0 66.0 62.9 59.8 56.8
53.8 52 52 52 52
-0.1 0.8
-0.3 0.7
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
O
(b) The bid and urgency of EV 1 varying with λ when N = 10.
Figure 3.3: Comparison on the bid and urgency at NE with varying the number N
and the weight λ. Left y-axis is for bid, and right y-axis is for urgency. Grid power
is labeled by values.
Urgency-Pareto Bid-NE Unit price-Pareto

0.25 6 0.25 6 0.25 6
0.2 4 0.2 4 0.2 4

Urgency
Urgency
Urgency
0.15 2 0.15 2 0.15 2

Bid
Bid
Bid
0.1 0 0.1 0 0.1 0
0.05 -2 0.05 -2 0.05 -2
0 -4 0 -4 0 -4
1 2 1 2 3 4 5 6 1 2 3 4 5 6 7 8 9 10
EV id EV id EV id
Figure 3.4: Comparison between Nash equilibrium and Pareto optimum. The bid at
Nash equilibrium, unit price and urgency at Pareto optimum with different settings
of N = 2, 6, 10, U = exp(·) and λ = 0.2. Left (right) y-axis is for bid (urgency).
results for the EV number N ∈ {2, 4, 6}. The bid at NE is identical;

while the unit prices at the Pareto optimum may be different. More-
over, at the Pareto optimum, EVs with larger urgencies have unit prices
higher than or equal to that of EVs with smaller urgencies. This ob-
servation follows our intuition. For example, when N = 2, EV 1 has
higher unit price; when N = 10, EV 3, EV 8 and EV 10 have the three
largest urgencies and possess higher unit prices. In addition, as the EV
number rises, the difference of unit prices at the Pareto optimum de-
creases. We conjecture that if N approached to infinity, all EVs would
have the same unit price. This conjecture can be theoretically validated
using the stationarity condition given by Eqn. (3.12). We rewrite the
condition as follows:
T δw
bi = + b2
, ∀i.
2λP 2λyU 0 (Ωi − TPy bwi + λy wi )
As EV number N increases, w also increases. Thus, the last item of

the above equation tends to be identical across all EVs and thus so do
their unit prices.
3.5.3 Results on the Stackelberg Equilibrium

In this subsection, we evaluate how the co-located renewable generation
and its uncertainty impact the charging station’s utility at the Stack-
elberg Equilibrium. The number of EVs is set as N = 20. We assume
that the prediction errors follow a uniform distribution for this simu-
lation. It is worth to note that our theoretical analysis in Section 3.4
does not rely on any specific distribution. To make the prediction error
2
with zero mean
√ and√ variance σ , the range of the uniform distribution
is set as [− 3σ, 3σ].
The station’s utility. Fig. 3.5 demonstrates how the station’s util-
ity varies with the renewable generation and EVs’ bids. The results in
this figure are averaged over 10000 runs in order to show the probabilis-
tic characteristics of the Stackelberg equilibrium. First, by Fig. 3.5(a),
when the standard deviation σ is low, e.g., σ = 0 and σ = 0.1, the util-
ity increases as the actual renewable generation χ grows. Interestingly,
when the standard deviation σ is high, e.g., σ = 0.3 and σ = 0.4, the
33 80
V=0 V=0
32 V=0.1 V=0.1
60
V=0.2 V=0.2
Utility
Utility
31 V=0.3 V=0.3
V=0.4 40 V=0.4
30
20
29
5 10 20 30 40 50 0.05 0.1 0.2 0.3 0.4 0.5
Actual renewable generation (kw) Bid
(a) Utility varying with actual renew- (b) Utility varying with bid b∗ when
able generation χ when bid b∗ = 0.2. actual renewable generation χ = 50kw.
Figure 3.5: Comparison on the station’s utility with different values of variance.
Note that σ = 0 is the resulting utility based on the actual renewable generation
and others are based on the predicted generation.
utility first increases and then decreases as χ rises. The reason is that
the renewable’s uncertainty has little impact on the station’s strategy
when either the standard deviation is low or the renewable genera-
tion is small, and vice versa. One key insight here is that aggressively
increasing the capacity of the co-located renewable will cause utility re-
duction to the station if the prediction of renewable is inaccurate. For
example, as to the present setting, the best capacity of the renewable
is around 25kw when σ = 0.3, and it is around 15kw when σ = 0.4.
Second, by Fig. 3.5(b), the station’s utility rises as the bid grows. This
observation quite follows the intuition that higher bids results in larger
payment by EVs and thus higher utility for the station. Moreover, the
utility increases faster when the standard deviation is low than when
it is high. This is also because higher standard deviations have more
impact on the station’s strategy.
Sensitivity to prediction errors. We have already provided the
impact of renewable’s uncertainty on the station’s utility in the above.
We now use Fig. 3.6 and briefly analyze how the competitive ratio
varies with the standard deviation of the prediction error. First, as the
standard deviation rises, the inaccuracy of the prediction impacts more
on the station’s strategy. Hence, by both of the figures, the competitive
ratio, i.e., the distance to the actual optimal solution, increases accord-
ingly. Second, by Fig. 3.6(a), the competitive ratio rises as the renew-
3.6. Summary 57

1.5
b*=0.5
&RPSHWLWLYHUDWLR
Competitive ratio
$
~ = 100kw
$
~ = 80kw 1.4
b*=0.4
$
~ = 60kw
$
~ = 40kw 1.3 b*=0.3
$
~ = 20kw b*=0.2
1.2 *
b =0.1
1.1
1
0 0.2 0.4 0.6 0.8 1
6WDUQGDUGGHYLDWLRQV Standard deviation (V)
(a) Competitive ratios with different (b) Utility varying with different val-
values of predicted renewable genera- ues of bids b∗ when predicted renew-
e when bid b∗ = 0.2.
tion χ able generation χe = 100kw.
Figure 3.6: Comparison on the charging station’s utility with varying standard
deviation σ.
able generation increases. This observation also discourages the station

to aggressively increase the renewable capacity. Third, by Fig. 3.6(b),
the competitive ratio increases as the bid rises. This seems to be a mo-
tivation that the station should impose lower unit prices to EV users
to obtain better performance. However, as illustrated by Fig. 3.5(b),
higher bids still benefit the station as to the utility maximization.
Hence, there is no need for the charging station to impose some unit
prices that are lower than the bid at the Nash equilibrium.
3.6 Summary
This chapter focuses on hierarchical demand response for EVs via

charging stations. We propose a distributed deadline-aware two-level
market mechanism for this scheme. The proposed mechanism has a hi-
erarchical structure, where the charging station leads the market and
EVs follow and determine their strategies accordingly. The mechanism
provably converges to a unique equilibrium solution, where no play-
ers can improve their individual performance gains by changing only
their own strategies. We further present a stochastic optimization based
algorithm and demonstrate its robust performance guarantee for the
charging station at the equilibrium. The trace-driven simulation re-
sults demonstrate the efficacy and also validate the theoretical analysis
for the designed market mechanism.
3.7 Appendix
Proof for Lemma 3.1.
Proof. (i) The strict convexity of disutility ui (bi , b−i , y) with respect
to bi . Given Eqn. (3.1)(3.2)(3.3), we have
ui (bi , b−i , y) = U (zi ), where

T y bi b2
zi (bi , b−i , y) = Ωi − P + λy P i . (3.26)
P i bi i bi
The first and second derivative of zi with respect to bi are

2bi −i bj + b2i
P P
∂zi T y −i bj
=− + λy , (3.27)
∂bi P ( i bi )2 ( i bi )2
P P
∂ 2 zi ( −i bj )2
P P
2T y −i bj
= + 2λy , (3.28)
∂b2i P ( i bi )3 ( i bi )3
P P
P
where −i bj is the sum of all bids except bi . The second derivatives
of ui (bi , b−i , y) with respect to bi is
2
∂ 2 ui ∂2U ∂zi ∂U ∂ 2 zi

2 = + . (3.29)
∂bi ∂zi2 ∂bi ∂zi ∂b2i
∂2U
Since U (·) is convex and strictly increasing, we have ∂zi2
≥ 0 and
∂U ∂2z
∂zi > 0. Moreover, by Eqn. (3.28), ∂b2i
i
is greater than zero. Hence, the
right-hand side of Eqn. (3.29) is positive, and the disutility ui (bi , b−i , y)
is strictly convex on the strategy set Bi with respect to bi .
(ii) The closed-form best response. Due to the strict convexity of
ui (bi , b−i , y), we derive the best response of EV i by setting its first
derivative to zero for the minimization. Since ∂u ∂ui ∂zi ∂ui
∂bi = ∂zi ∂bi and ∂zi > 0,
i
setting ∂u ∂zi
∂bi = 0 is equivalent to setting ∂bi = 0. Based on Eqn. (3.27)
i
and solving the equation ∂z ∂bi = 0, we have the best response of EV i as

i
s
X 2 T X X
Ri (b−i , y) = ( bj ) + ( bj ) − ( bj ). (3.30)
−i Pλ −i −i
3.7. Appendix 59
(iii) Standard function properties. We show that Ri (b−i , y) as a

function of b−i is a standard function, i.e., satisfies all the three prop-
erties in Definition 3.3.
• Positivity. Since PTλ ( −i bj ) > 0 in Eqn. (3.30), we have
P
r
X 2 X
Ri (b−i , y) > ( bj ) − ( bj ) = 0.
−i −i
• Monotonicity. The first derivative of Ri (b−i , y) with respect to

bj , ∀j 6= i, is
PT
∂Ri −i bj + 2P λ
=qP −1
∂bj ( −i bj )2 + PTλ ( −i bj )
P
r 2
P 2 T P T
( −i bj ) + Pλ
( −i bj ) + 2P λ
= q P −1
2 T P
( −i bj ) + Pλ
( −i bj )
q P
2 T P
( −i bj ) + Pλ
( −i bj )
>qP − 1 = 0.
( −i bj )2 + T P
Pλ
( −i bj )
Hence, Ri (b−i , y) is monotone with respect to b−i .
• Scalability. Using Eqn. (3.30) and a scale factor α > 1, we have
the following deduction.
s
X 2 α2 T X X
αRi (b−i , y) = α2 ( bj ) + ( bj ) − α( bj )
−i Pλ −i −i
s
X 2 αT X X
> α2 ( bj ) + ( bj ) − α( bj )
−i Pλ −i −i
= Ri (αb−i , y).
Hence, all three properties are satisfied. The lemma holds.
Proof. (i) Existence. By definition, the strategy set Bi for EV i is closed,

bounded and convex. Moreover, the disutility function ui (bi , b−i , y) is
continuous and, by Eqn. (3.28), strictly convex on Bi , ∀i. Hence, accord-
ing to Theorem 4.3 from [78], the noncooperative game among EVs has
a Nash equilibrium.
(ii) Uniqueness. Since Ri (b∗−i , y) = Ri (b∗i , y), ∀i, by Definition 3.2,

the Nash equilibrium is a vector b∗ that satisfies b∗ = R(b∗ , y), where
R(b∗ , y) = (Ri (b∗i , y))∀i . Based on Lemma 3.1, R(b∗ , y) is a standard
function. Hence, according to Theorem 1 from [93], the vector b∗ must
be unique, i.e., the Nash equilibrium is unique. The theorem holds.
Proof. The best response function given by Eqn. (3.30) shows the sym-
metry that all EVs have the same form. Thus, the unique Nash equi-
librium is obtained when b∗i = b∗j , ∀i 6= j. Based on Eqn. (3.6)(3.30),
we have
s
T ∗
b∗i = (N − 1)2 b∗i 2 + (N − 1) b − (N − 1)b∗i . (3.31)
Pλ i
Projecting the solution of Eqn. (3.31) into the feasible set Bi , we
have Eqn. (3.7). To prove that it is still the Nash equilibrium af-
ter the projection, by Definition 3.1, we need to show ui (b∗i , b∗i , y) ≤
ui (bi , b∗i , y), ∀bi ∈ Bi ∧ bi 6= b∗i . Since U (·) in Eqn. (3.2) is strictly in-
creasing, it only needs to demonstrate zi (b∗i , b∗i , y) ≤ zi (bi , b∗i , y), ∀bi ∈
Bi ∧ bi 6= b∗i . We consider the following two cases.
(i) When projected to B, we have
T (N − 1) T B(2N − 1)
<B ⇒ < . (3.32)
λP (2N − 1) Pλ N −1
For a bid bi > B, based on Eqn. (3.32), we have
zi (B, (B)−i , y) − zi (bi , (B)−i , y)

!!
T bi 1 b2i B

=λy − − −
λP (N − 1)B + bi N (N − 1)B + bi N
−λyN (bi − B)2
< ≤ 0.
N ((N − 1)B + bi )
(ii) When projected to B, we have
T (N − 1) T B(2N − 1)
>B ⇒ > . (3.33)
λP (2N − 1) Pλ N −1
3.7. Appendix 61
For a bid bi < B, based on Eqn. (3.33), we have
zi (B, (B)−i , y) − zi (bi , (B)−i , y)

−λyN (B − bi )2
< ≤ 0.
N ((N − 1)B + bi )
Therefore, after the projections, it still achieves the Nash equilibrium
and the theorem holds.

P
Proof. (i) The lower bound L. We relax the constraint w = i bi in
Eqn. (3.11) using N B ≤ w ≤ N B. This relaxation means that w
and bi , ∀bi are fully decoupled. Thus, the optimal value of the new
problem can be seen as a lower bound of the original problem. We
use Fe to denote the total disutility of the new problem. Since w and
bi , ∀i are fully decoupled for the new problem, its optimum can be
derived analytically. Fe is convex with respect to bi , and thus letting
h iB
∂zi T
∂bi = 0 derives the optimal solution ebi = 2P λ B
, ∀i. Then, the optimal
solution w
e is obtained as follows. If B ≤ PTλ , then ∂∂wF
e
≥ 0, i.e., Fe is
non-decreasing as to w, given ebi , ∀i. Thus, w
e = N B and Fe (N B, (e
bi )∀i )
T ∂F
is the first part of the lower bound L. If B > Pλ
, then e
∂w < 0, i.e., Fe
is decreasing as to w. Thus, w
e = N B and Fe (N B, (e
bi )∀i ) is the second
part of the lower bound L.
(ii) The upper bound Q. We add an additional constraint to
Eqn. (3.11), which is b1 = b2 = · · · = bN . Then, the optimal value
of the new problem can be seen as an upper bound of the original
problem. The new problem is
X Ty λybi
min U (Ωi − + ).
bi ∈Bi ,∀i i NP N
The objective function is strictly increasing with respect to bi . Thus,
the optimal optimum of the new problem is achieved when bi = B, ∀i,
and the corresponding optimal value is right the upper bound Q.
Proof. Note that = + + − , hence we have that E[] = 0 implies

E[+ ] = −E[− ]. Then, it follows that
σ 2 = E[2 ] = E[(+ + − )2 ] = E[(+ )2 ] + E[(− )2 ] + 2E[+ − ]

= E[(+ )2 ] + E[(− )2 ]
E[+ ]2 E[− ]2
≥ + (3.34)
P( ≥ 0) P( < 0)
1 1

+ 2
= E[ ] +
P( ≥ 0) 1 − P( ≥ 0)
1

+ 2
= E[ ]
P( ≥ 0)(1 − P( ≥ 0))
1
≥ 4E[+ ]2 , {∵ equality attains when P( ≥ 0) = }
2
where P(·) is the probability density function. Rearranging the in-
equality, we have E[+ ] ≤ σ2 and thus E[− ] ≥ − σ2 . The inequality
of Eqn. (3.34) holds because
Z ∞ Z ∞
E[(+ )2 ]P( ≥ 0) = h2 dF (h) 1dF (h)
0 0
Z ∞ 2
≥ hdF (h) = E[(+ )]2 ,
0
where F (·) is the cumulative distribution function. Rearranging the

E[+ ]2
inequality, we have E[(+ )2 ] ≥ P(≥0) . Using similar deduction, we can
E[− ]2
also have E[(− )2 ] ≥ P(<0) . The proof is parallel to that in [95].
3.7. Appendix 63
Proof. Based on the definition of competitive ratio in Eqn. (3.25), we

have the following deduction:
1 h
∗ ∗ ∗ NP
i
η= E G(N P − y ) + b [y + χ]
E [g(b∗ , ye∗ )]
1 h
∗ ∗ ∗ NP
i
= E G(N P − y − ϕ) + b [ y + ϕ + χ]
E [g(b∗ , ye∗ )]
e e
{∵ Lemma 3.7}
1 h
≤ E G(N P − ye∗ ) − G(N P − ye∗ ) + b∗ (ye∗ + χ)
E [g(b∗ , ye∗ )]
h i i
+G(N P − ye∗ − ϕ− ) + b∗ ϕ − ye∗ + ϕ + χ − N P
0
∵ Since G(·) is nondecreasing, replacing ϕ with ϕ− .

1 h
=1+ E G(N P − ye∗ − ϕ− ) − G(N P − ye∗ )
E [g(b∗ , ye∗ )]
h i i
+b∗ ϕ − ye∗ + ϕ + χ − N P
0
1 h
− 0 ∗ ∗
i
≤1+ E (−ϕ )G (N P − y ) + b ϕ {∵ G(·) is concave}
E [g(b∗ , ye∗ )]
e
1
0 ∗ − ∗
h
+ −
i
=1+ G (N P − y )E (−ϕ ) + b E ϕ + ϕ
g(b∗ , ye∗ )
e
1
0 ∗ − ∗
h
+
i
≤1+ G (N P − y )E − χ + b E χ
g(b∗ , ye∗ )
e e e

e G0 (N P − ye∗ ) + b∗
χ
≤1+ σ{∵ Lemma 3.6}.
2g(b∗ , ye∗ )
where the comments inside {} are the main reasons why the correspond-
ing equalities or inequalities hold. We now have finished the proof and
the theorem holds.
4
Charging Load Scheduling
Due to the relatively long and frequent charging cycles, one promis-
ing operation mode, named park-and-charge, has been recently pro-
posed for electric vehicle charging [10, 11, 12, 100, 56]. By this mode,
parking lots are equipped with charging points to function with both
parking and charging services. Electric vehicles can receive charging
during the period of parking. One key advantage of this mode is that
customers can directly follow their agendas after parking their vehicles
with no need to care about the charging process. Potential applications
of this park-and-charge mode include parking lots at office buildings,
shopping malls, and airports. Recently, several field experiments have
been conducted to explore the feasibility of the park-and-charge mode.
For example, several universities in Europe together with The Bosch
Group carry out the project V-Charge, which aims to develop an au-
tomated valet parking and charging system to support autonomous
local transportation [100]. General Motors (GM), OnStar and Timber-
Rock collaborate to perform a pioneering experiment at GM’s E-Motor
Plant, whose objective is to coordinate the charging demand of parked
electric vehicles with the co-located renewable generation [56]. These
small-scale experiments present positive feedbacks on the potential of
64
65
the park-and-charge mode. An important step to deploy this mode to

large-scale charging stations is to develop effective and efficient charg-
ing load scheduling methods.
Although researchers have begun to take efforts towards this step,
most existing works solely focus on the charging service and thus con-
fine to the time-driven scheduling policy, e.g., [101, 41, 26, 102, 103,
39, 104, 105, 106]. By this policy, the time line is equally slotted and
scheduling decisions are made for each time slot. A key assumption in
these works is that the arrival time and deadline (i.e., the time when
the user picks up her/his EV) of an EV are right at the beginning or the
end of a time slot. Under this assumption, one major dilemma for ap-
plying this policy to park-and-charge systems is to determine the length
of time slots. Long time slots lead to few charging mode switchings but
cause under-utilized charging points at the stations, while short time
slots improve charging point utilization but cause many mode transi-
tions for electric vehicles. Frequent mode switchings can significantly
compromise the lifetime and capacity of EV batteries. Hence, to avoid
such dilemma, we propose to explore the event-driven scheduling policy
to schedule electric vehicle charging in park-and-charge systems in this
chapter.
Event-driven scheduling is widely recognized in real-time commu-
nity, and thus many real-time scheduling methods such as earliest-
deadline-first (EDF) and least-laxity-first (LLF) have been designed
along with this policy. Further, these scheduling methods have been ex-
tensively studied in various kinds of systems such as embedded systems,
communication systems [107], and even server farms/data centers [108].
A clear vision is, however, still elusive regarding the feasibility and ben-
efits of applying real-time scheduling to park-and-charge systems. Un-
derstanding and analyzing such application is a non-trivial task because
of three key challenges. First, modeling charging tasks should capture
both the parking and charging functions. Second, charging tasks usu-
ally have different temporal constraints and may also have different
values (e.g., utilities to measure user satisfaction) due to different en-
ergy prices paid by the customers. Choosing between temporality-based
and value-based scheduling methods is involved. Third, the scheduling
66 Charging Load Scheduling
methods should be online, which need to determine vehicles for charg-

ing based only on the information of electric vehicles that have already
arrived at the system.
To address these challenges, we first propose to adopt a metered
model, by which the system gains in proportion to the amount of
already served demand of a charging task. This metered model bet-
ter fits to park-and-charge, compared with the classical deterministic
model (detailed model discussion in Section 4.2.2). Second, we study
the adaption of several widely used on-line scheduling methods to the
metered model. They include: (i) temporality-based algorithms such
as EDF and shortest-job-first (SJF); (ii) a value-based algorithm, i.e.,
highest-utility-first (HUF). Their pros and cons regarding park-and-
charge are thoroughly discussed using: (i) theoretical analysis via the
competitive analysis and resource augmentation techniques; (ii) exper-
imental validations via detailed simulations. To be specific, our major
contributions are as follows.
• We novelly adapt several well-known on-line scheduling algo-
rithms to schedule EV charging in park-and-charge systems. As to
maximizing EV user satisfaction, we demonstrate (i) the non-constant
performance bound of EDF and the constant bound of HUF; (ii) a gen-
eralized theoretical result, i.e., the performance bound for the whole
class of work-conserving scheduling algorithms.
• We conduct extensive simulations to make comparison on opti-
mizing EV user satisfaction. The results show the near-optimality and
performance consistency of the value-based algorithm (i.e., HUF), and
the sub-optimality and performance anomaly of temporality-based al-
gorithms (e.g., EDF and SJF).
Note that this chapter targets the case of multiple charging points,
which is equivalent to the multiple processor case as to the scheduling
theory. From this perspective, our work is an extension to [109].
Work [109] carries out competitive analysis for HUF solely, while our
work performs both competitive analysis and resource augmentation
analysis for several well-known algorithms (including HUF) and
further for the whole class of work-conserving algorithms.
The rest of the chapter is organized as follows. Section 3.1 presents

recent developments. Section 4.2 presents the system model and prob-
lem statement. Sections 4.3 and 4.4 analyze algorithms using competi-
tive analysis and resource augmentation respectively. Section 4.5 eval-
uates the studied on-line algorithms. Section 4.6 presents the summary
of this chapter. 1
Since time-driven methods for charging rate control has been reviewed
in Section 2.1, this section focuses on investigating existing works using
event-driven methods. Specially, we review two group of works: one
group is applying event-driven scheduling to EV charging and the other
is event-driven scheduling with the metered task model.
EV Charging Using Event-Driven Scheduling. The authors
of [80] model power loads as tasks, characterized by arrival time, dead-
line, energy demand, and power rate limit. They then apply real-time
scheduling policies such as EDF and LLF to schedule deferrable loads,
and further demonstrate the optimality of EDF. However, the discus-
sion here demonstrates several different results as follows. First, LLF is
unsuitable to schedule charging loads due to its frequent mode switch-
ing, which will cause battery degradation. Second, EDF is not optimal
for charging scheduling in presence of heterogenous user satisfaction.
Recognizing the applicability of real-time scheduling, the authors of
[10] combine admission control and pricing with scheduling approaches
such as EDF and FCFS. The authors of [110] address minimizing
charging waiting time of EVs in a highway scenario and apply FCFS
based queueing theory to schedule vehicle charging. The authors of
[12, 111] discuss the V-Charge project of park-and-charge, whose vi-
sion is to provide autonomous parking and charging to improve user
comfort. V-Charge presents a solution that efficiently utilizes charging
resources and increases the feasibility of EV users. One key enabler
to autonomous park-and-charge in V-Charge is the application of real-
time scheduling to coordinate EVs in a parking lot.
1
Although using similar event-driven schedule methods, there are

several major differences between the discussion in this chapter and
the above existing works. First, our work employs the metered task
model, which is more suitable to park-and-charge. Second, besides
temporality-based methods such as EDF and FCFS, we also study
value-based methods and further the whole class of work-conserving
methods. Third, our work carries out rigorous analysis based on both
competitive analysis and resource augmentation.
Metered Model Based Event-Driven Scheduling. Multiple
charging points can be seen as multiple processers as to the schedul-
ing theory, and thus we review technical developments of metered
model based scheduling methods. The authors of [112] study preemp-
tive scheduling for a set of metered tasks to maximize the overall value.
They propose an efficient algorithm to improve the lower and upper
bounds of the competitive ratio. The authors of [113] address the same
setting as in [112] and design another algorithm to further improve the
lower and upper bounds. These two existing works mainly focus on
uni-processor scheduling, that is, they consider charging scheduling for
only one charging point.
The discussion in this chapter is most related to [109], which studies
metered model based scheduling for multiple processors. Compared to
[109], we make significant technical contributions towards this schedul-
ing problem. The authors of [109] perform competitive analysis only for
HUF, i.e., no competitive analysis for other algorithms and no resource
augmentation analysis. In other words, the only overlapping between
our work and [109] is HUF. By contrast, our work first extends compet-
itive analysis for other well-known algorithms including EDF, FCFS,
and SJF, and further carries out generalized analysis about the entire
class of work-conserving algorithms for the multiprocessor case. Sec-
ond, our work performs resource augmentation to the multiprocessor
case for both temporality and value based algorithms.
4.2. Problem Description 69
CentralController
...
TheParkingandChargingZone
PowerLine ControlLink ChargingPoint
(a) A combined zone
CentralScheduler
...
...
TheParkingZone
TheChargingZone
(b) Separate zones
Figure 4.1: The schematics of two different implementations of park-and-charge.
4.2 Problem Description
There are usually two different implementations for a park-and-charge

system. The first one, a combined zone (Fig. 4.1(a)), is that the parking
zone and charging zone are in the same area, and each parking space
is equipped with one charging point [7, 8]. A centralized controller
controls power on and off for each charging point. The second one,
separate zones (Fig. 4.1(b)), is that the parking zone and charging
zone are at different areas [100, 111]. The parking zone has no charging
infrastructure. Charging points only concentrate in the charging zone
and always have power supply. A centralized scheduler decides which
electric vehicles to charge at the charging zone and which to park at
the parking zone. This implementation needs valets to assist moving
vehicles or allows fully autonomous driving between the two separate
zones [100].
For the customers, they check in and drop off their electric vehi-
cles at the parking zone. Then, they can leave and directly proceed to
follow their original agendas, instead of taking care of charging their ve-
hicles. During their absence, the park-and-charge system is responsible

to schedule vehicles for charging. Upon the returns, they pick up their
vehicles from the parking zone, where the vehicles are already wait-
ing for them. For the system operator, the checked-in vehicles need to
be scheduled according to users’ charging requirements. This work fo-
cuses on the charging scheduling problem: to determine when to charge
which electric vehicles. The scheduling decisions can be realized by (i)
switching power on or off for the corresponding charging points as to
the combined-zone case in Fig 4.1(a), and (ii) moving the correspond-
ing vehicles into or out of the charging zone as to the separate-zone
case in Fig. 4.1(b).
In this section, we first present a general system model that applies
to both of the cases. Then, we use model differentiation to show that
the model better fits the park-and-charge scenario, and lastly provide
problem statement.
4.2.1 System Modeling
We consider a park-and-charge system that supports at most M charg-

ing points turned on currently. This limit derives from the physical con-
straints such as respecting the capacity of upstream transformers [9],
or is imposed by the system operators such as capping the peak power
of the systems [3, 30]. Without loss of generality, we assume that M is
less than the number of parking spots [7, 111, 12]. Each charging point
supports a charging power of p. Each electric vehicle is modeled by a
charging task EVi , which is characterized by a four-tuple (ai , di , ei , ui ).
The following first details the definition of each parameter and then
provides the accommodation of this model to realistic charging behav-
ior of EV battery.
Parameter Definition. (i) The arrival time ai is when the cus-
tomer drops off his/her vehicle at the parking zone. At this time, the
vehicle is also ready for charging.
(ii) The charging demand ei is the amount of energy that the user
expects after the deadline. It can be estimated by (soc0i − soci ) · Bi ,
where soci is the state of charge (SOC) at the arrival time ai ; soc0i
is the expected SOC after deadline di ; Bi is the capacity of the vehi-
cle’s battery. SOC is usually denoted as a percentage value. Existing

methods such as [114, 115] can be used to estimate SOC of EV battery.
(iii) The deadline di is a user-specified time, after which he/she will
come back to pick up his/her vehicle. The deadline is soft, that is, after
it, the system stops charging the vehicle and switch off the power supply
or park it at the parking zone. If with partially completed demand, the
system gains the partial value. Further, it uses arbitrary deadline. That
is, there is no constraint between the deadline di and charging time ei /p,
i.e., the former can be larger than, less than, or equal to the latter.
This setting can largely avoid cheating as to customers specifying their
charging requirements, e.g., deadline. Since an EV receives no energy
after the deadline, the vehicle may receive little energy in the end if
the customer deliberately reports a much shorter deadline than ei /p.
(iv) The parameter ui denotes unit utility. If an EV receives an
amount x of energy, the gained utility is ui · x. The utility is a mea-
surement of user satisfaction or the value of a charging task, which de-
pends on both the energy price and state-of-charge. First, to differenti-
ate users’ charging requirements, the system may employ some market
mechanisms such as pricing or biding to determine energy price for in-
dividual charging tasks2 [84, 19]. Thus tasks may have different energy
prices. Customers willing to pay more usually come with higher value
about the charging service [84, 19]. Second, user satisfaction is usually
assumed to follow the law of diminishing utility [116, 39], where the
marginal utility decreases as state-of-charge increases. Hence, the unit
utility can be modeled as a function that is non-decreasing as to energy
price and non-increasing as to state-of-charge. This kind of modeling
approach has been widely used in EV charging studies [116, 39, 59, 5]
and network economics literature [117, 88, 91].
Model Accommodation for Realistic Non-linear Charging
Behavior. In reality, SOC of EV battery is non-linear with time. For
accommodating to this, we can split a charging task into a sequence of
subtasks that arrive sequentially at the system. We use an approxima-
2
No matter what market mechanisms are adopted, the system needs to schedule
charging tasks after deciding the energy price. This scheduling step is very important
to satisfy customers and is the right focus of this chapter.
10 Price = 2
Unit Utility
Price = 1
5
0
0 20 40 60 80 100
SOC (%)
Figure 4.2: An example piecewise linear function for the unit utility, which in-
creases as to the energy price and decreases as to the SOC.
tion: for each subtask, SOC or the received energy of an EV is linear

with time. Further, the unit utility function can be approximated by
a piecewise-constant function that respects the above non-decreasing
and non-increasing features. Each subtask corresponds to a piece and
thus has a constant unit utility. Fig. 4.2 shows a numerical example
for this accommodation. By this example, for a subtask with SOC in
[0, 20%], the unit utility is 10 if energy price is 2 (the dotted lines) and
it is 5 if price is 1 (the solid lines); for that in (20%, 40%], it is 8 if price
is 2 and 4 if price is 1; and so on. The gained utility of a charging task
EVi is thus equal to the sum of each subtask’s utility, i.e., j uij · xij ,
P
where uij and xij are the unit utility and received energy of subtask
EVij respectively. Note that the following results remain true with this
accommodation. For ease of presentation, we associate each electric ve-
hicle with one charging task and omit the subtask index in the following
algorithm analysis.
The system allows preemption of the charging process of electric ve-
hicles, and each preemption corresponds to one charging mode switch-
ing or mode transition. That is, the system is able to suspend an electric
vehicle’s charging process by turning power off for its charging point as
to the combined-zone case or re-parking the vehicle at the parking zone
as to the separate-zone case. Later on, it can also be resumed from the
suspension and continue for charging. We assume that the time over-
head incurred by charging mode switching is negligible compared with
the long charging time of electric vehicles.
4.2.2 Model Differentiation
Model Categorization. Real-time task models can be classified based

on different metrics such as flexibility on parallelism [118] and abstrac-
tion levels [119]. This chapter uses a categorization according to the
relation between the gained utility and the amount of finished work-
load (charging demand here). Thus, there are three categories: deter-
ministic, metered and hybrid task models. By the deterministic model,
the utility is gained (and thus the customer pays) only if the charg-
ing demand of a charging task is fully fulfilled, otherwise the gained
utility is zero (and the customer pays nothing)[10, 12, 80]. This model
follows the classical all-or-nothing setting for hard real-time tasks. The
metered task model allows the gained utility (and payment) to be a
strictly increasing function of the amount of finished workload. The
hybrid task model can be seen as a combination of the above two mod-
els. That is, each hybrid task consists of two subtasks. One uses the
deterministic model, i.e., the utility of the subtask is gained only if it
is fully completed. The other uses the metered model, i.e., the gained
utility of the subtask increases as more workload is finished. This hy-
brid model is closely related to the task model adopted in the imprecise
computation literature [120].
Model Suitability. EV charging loads in different application sce-
narios, such as home charging and park-and-charge, have different fea-
tures about the temporality and user satisfaction. Thus, it needs differ-
ent task models (with respect to the above categorization) to capture
the distinct features of different kinds of charging loads. However, exist-
ing works that use event-driven scheduling mainly confine to the deter-
ministic model regardless of the application scenarios, e.g., [10, 80, 12].
The deterministic model is not suitable to the park-and-charge scenario
due to reasons as follows.
With the deterministic task model, customers may intend to submit
stringent charging requirements. The system either accepts them and
risks at economic loss due to the potential unfulfilled requirements,
or simply rejects them and much dissatisfies these customers. This
dilemma may also happen if with the hybrid task model. Thus, both
deterministic and hybrid models are not suitable for park-and-charge.
By contrast, the metered model is compatible with such strict require-

ments. Further, it will not happen that even though there are enough
parking spots, the system rejects electric vehicles only due to their
tough requirements. Hence, the metered task model is more suitable
to characterize the charging load in park-and-charge systems. To be
specific, the task model proposed in the previous subsection belongs to
the category of metered model, and thus well fits to park-and-charge .
As mentioned above, existing works, e.g., [10, 80, 12], use real-time
scheduling methods such as EDF and LLF, but they confine to the
deterministic model. One key question is how these methods perform
when adapted to the metered model. The following sections answer this
question and further analyze more event-driven scheduling methods.
4.2.3 Problem Statement
This work studies the on-line charging scheduling problem with the
objective of maximizing the social welfare. A charging schedule is to
decide when to charge which vehicles. The social welfare is defined as
the sum utility of electric vehicles. This charging scheduling problem
pursues maximized social welfare, which is different from that of hard
real-time scheduling problems whose goal is to maximize the number
of tasks meeting their deadlines. Electric vehicles arrive in an on-line
manner, i.e., the system knows no details of the charging tasks before
the vehicles’ arrival. We assume that after arrival, all parameters of
a charging task are known to the system. On-line algorithms need to
make scheduling decisions based only on the parameters of electric
vehicles that have already arrived.
4.3 Algorithm Analysis Based on Competitive Analysis
In this section, we present the theoretical results based on competitive

analysis. We first demonstrate the non-constant and constant compet-
itive ratios for EDF and HUF (highest-utility first) respectively, and
then provide a much stronger result, i.e., the performance upper bound
of the whole class of work-conserving algorithms.
4.3. Algorithm Analysis Based on Competitive Analysis 75
4.3.1 Analysis for EDF and LLF

Impracticability of LLF. Earliest-deadline-first (EDF) and Least-
laxity-first (LLF) are two classic temporality-based algorithms. EDF
always first charges electric vehicles with the earliest deadlines (di );
while LLF always first processes vehicles with the least laxity time
(di − ai − e0i /p, where e0i is the remaining charging demand). Recently,
LLF has been used to schedule deferrable loads such as charging tasks
[80]. However, LLF is impractical for the charging scheduling problem.
The reason is that the algorithm may cause much frequent preemption
or mode switching. For example, suppose there are two charging points
and three vehicles. The charging power is set as p = 1 and the three
vehicles’ setting (by (ai , di , ei , ui )) is: EV1 = (0, 1, 1, 1), EV2 = (0, 1, 1−
ε, 1), EV3 = (0, 1 + ε, 1 − ε, 1), where 0 < ε < 1. By LLF, the number
of mode transitions for EV2 and EV3 is both O(1/ε). Thus, if ε is
small, the number will be large. Actually, this number is unbounded,
i.e., approaches infinity as ε → 0.
Most electric vehicles in the current generation are equipped with
Li-ion batteries. The lifetime and capacity of Li-ion batteries are signif-
icantly compromised by the frequent mode switching [121]. Although
the next-generation battery for future electric vehicles would have bet-
ter resilience to the mode switching, it is still dangerous to employ such
an algorithm with unbounded number of mode transitions. Hence, we
focus on the performance analysis for EDF in the following.
No constant competitive ratio for EDF. Competitive analysis
provides the performance gap between an on-line algorithm and the
off-line optimal algorithm. The performance gap is expressed by a def-
inition of competitive ratio as follows. Note that the definition is for
maximization problem, not for minimization problem such as [122].
Definition 4.1 (Competitive ratio). An on-line algorithm A is α-

competitive if Opt(I)
A(I) ≤ α for any feasible input instance I, where Opt(I)
and A(I) are the social welfare gained by the off-line optimal algorithm
and the algorithm A on input instance I, respectively.
Based on this definition, we demonstrate that there is no constant

competitive ratio for EDF. Suppose there are two charging points and
‫ܸܧ‬ଶ ‫ܸܧ‬ଵ ‫ܸܧ‬ଵ

Port1 t Port1 t
0 ͳെߝ ͳ 0 ͳെߝ ͳ
‫ܸܧ‬ଷ ‫ܸܧ‬૛
Port2 t Port2 t
0 ͳെߝ 0 ͳെߝ
EVarrivalܽ௜ EVdeadline݀௜
(a) EDF (EV1 misses deadline) (b) Opt (EV3 misses deadline)
Figure 4.3: Illustration for the example showing no constant bounds for EDF.
three electric vehicles. The charging power is set as p = 1 and the three
vehicles’ settings are as follows: EV1 = (0, 1, 1 − ε, 1/ε), EV2 = (0, 1 −
ε, 1 − ε, 1), EV3 = (0, 1 − ε, 1 − ε, ε), where 0 < ε < 1. As illustrated in
Fig. 4.3, EV2 and EV3 are first charged to finish by the EDF schedule;
while EV1 and EV2 are first charged to finish in the optimal schedule.
The utility gained by EV1 toEV3 is (i) EDF: 1, 1 − ε, (1 − ε) · ε; (ii)
Opt: (1 − ε)/ε, 1 − ε, 0. Hence, the social welfares derived by the two
schedules are:
1
U (EDF ) = 2 − ε2 , U (Opt) = − ε.
ε
Letting ε → 0 makes the competitive ratio arbitrarily approach the
positive infinity: lim UU(EDF
(Opt)
) → +∞. Therefore, EDF has no constant
ε→0
performance bound based on competitive analysis.
The non-constant competitive ratio. Before giving this per-
formance bound, we present several definitions and lemmas that will
be used in the following theoretical proofs. They are the definition of
canonical schedule, a shared property among off-line optimal schedules,
and a utility mapping scheme.
Definition 4.2 (Canonical schedule). A schedule S is called canonical if

for any two time points t1 < t2 , the following is satisfied: if z1 = S(t1 ),
and z2 = S(t2 ) 6= ∅, then either (i) mini∈z2 (ai ) > t1 , or (ii) z1 6= ∅
and maxi∈z1 (di ) < mini∈z2 (di ), where S(t) is the set of vehicles being
charged in schedule S at time t.
‫ݏ‬௜ ൅ ݂௝ െ ‫ݏ‬௝
‫ܸܧ‬௜ ‫ܸܧ‬௝ ‫ܸܧ‬௝ ‫ܸܧ‬௜ ‫ܸܧ‬௜

t t
‫ݏ‬௜ ݂௜ ‫ݏ‬௝ ݂௝ ݀௝ ݀௜ ‫ݏ‬௜ ݂௜ ‫ݏ‬௝ ݂௝ ݀௝ ݀௜
(a) Original schedule (b) Canonical schedule
Figure 4.4: Illustration of reschedule for Lemma 4.1.
Intuitively, Definition 4.2 indicates that at any time, schedule S

processes either the electric vehicles with the earliest deadlines, or dis-
card them forever. With this definition, we have Lemma 4.1 for the
off-line optimal schedules.
Lemma 4.1. At least one of the off-line optimal schedules is a canonical

schedule.
Proof. We prove this lemma by demonstrating that any off-line opti-

mal schedule can be converted to a canonical schedule without affect-
ing its optimality. Consider an off-line optimal schedule S, and EVi
with a later deadline and EVj with a earlier deadline. We use si and
fi (sj and fj ) to denote the starting and ending time of some con-
secutive portion of processing EVi ’s (EVj ’s) charging demand, where
si ≥ max(ai , aj ), sj ≥ max(ai , aj ). Suppose that EVi ’s portion is sched-
uled before EVj ’s portion, i.e., si < sj . As illustrated in Fig. 4.4, we
can reschedule and make EVj ’s portion processed ahead of EVi ’s with-
out changing their gained utility. For EVj , it is scheduled earlier than
the original, and thus its utility keeps unchanged. For EVi , though it
is postponed to a later time, its new ending time is still earlier than
max{fi , fj }, which is less than deadline di . Thus, EVi ’s utility also stays
unchanged. If using the above reschedule to all non-canonical portions,
we convert schedule S to a canonical schedule. The lemma holds.
To bound the competitive ratio of EDF, we further employ a sched-

ule mapping scheme proposed in [123]. We adapt this scheme to our sys-
‫ͳݐ݁݉݅ݐ‬ ‫ݐ݁݉݅ݐ‬Ϯ
KƉƚ ‫ܸܧ‬௜ ‫ܸܧ‬௝
‫݃ݎܥ‬ை௣௧ ‫ܸܧ‬௜ ǡ ‫ͳݐ‬ ‫݃ݎܥ‬ை௣௧ ሺ‫ܸܧ‬௝ ǡ ‫ʹݐ‬ሻ

൒ ‫݃ݎܥ‬஺ ሺ‫ܸܧ‬௜ ǡ ‫ͳݐ‬ሻ ൌ ‫݃ݎܥ‬஺ ሺ‫ܸܧ‬௝ ǡ ‫ͳݐ‬ሻ
‫ ͳݐ ܨ‬ൌ ‫ͳݐ‬ ‫ ʹݐ ܨ‬ൌ ‫ͳݐ‬
‫ܸܧ‬௝
Figure 4.5: The utility mapping scheme.
tem model and call it as utility mapping scheme. The adapted scheme
is detailed as follows.
The Utility Mapping Scheme. The scheme is described by a
function F : R → R, which maps each time point in the off-line optimal
schedule Opt to a time point in the schedule of an on-line algorithm
A. We use Crgx (EVi , t) to denote the amount of energy that has been
fueled into EVi by algorithm x by time t. For any time t, suppose EVi
is a vehicle being charged in Opt. If CrgOpt (EVi , t) ≥ CrgA (EVi , t),
F (t) = t. Otherwise, let F (t) = t0 < t, where t0 is the maximum time
that CrgOpt (EVi , t) = CrgA (EVi , t). In both cases the unit utility is
ui . The utility mapped to algorithm A is equal to that of Opt. Fig. 4.5
illustrates this utility mapping scheme. For example, at time t1 , since
CrgOpt (EVi , t1 ) ≥ CrgA (EVi , t1 ), the mapping is F (t1 ) = t1 .
We define the utility ratio at any time as the sum utility mapped
to A over the utility obtained by A at that time. The utility ratio at
time t0 in Fig. 4.5 is calculated by (p·ui + p·uj )/p·uj . Note that at any
time, there are at most two utilities mapped from Opt to A: one is p·ui
from time t0 and the other is p·uj from time t later than t0 . Hence, if we
can bound the utility ratio for all time points, then this offers a bound
on the competitive ratio of algorithm A.
We now show the non-constant performance bound, which is de-
pendent on the dynamic range of unit utility (see Theorem 4.2). This
dynamic range is described by the ratio of maximum to minimum unit
utility, β = max(ui )/ min(ui ), β ≥ 1.
Theorem 4.2. EDF is β-competitive and the bound is tight.

Proof. Bounding competitive ratio. Suppose that at time t,

EV1 , ..., EVM are the M vehicles being charged by EDF, and
EV10 , ..., EVM 0 are the M vehicles being charged by an off-line optimal
canonical schedule Opt. Some of the EVi s may be identical to some

of the EVi0 s. Thus, without loss of generality, we assume EVi = EVi0
for i ∈ I ⊂ {1, ..., M }. We discuss the two individual cases (whether
vehicles belong to set I) as follows.
Case 1 (i ∈ I): The utility of EVi0 by Opt can be mapped to EDF
at time t at most once. The reason is that according to the utility
mapping scheme described above, the EVi0 utility is either mapped to
time t or a time before t. The total utility mapped in this case is at
P
most i∈I p·ui .
Case 2 (i ∈ {1, ..., M } \ I): First, also according to the scheme, the
utility of the set of EVi0 is mapped to EDF at most once. The sum
utility is at most i∈{1,...,M }\I p·u0i . Second, the set of EVi is not being
P
charged in Opt at time t, and they will not be mapped to time t from a
later time. The reason is that EDF always first schedules vehicles with
earliest deadlines, and thus it must hold that di < mini (d0i ). Moreover,
since Opt is canonical, these EVi s will be dropped by Opt forever. Thus,
the mapped utility is zero.
Note that the utility obtained by the EDF schedule at time t is
P
i∈{1,...,M } p·ui . Hence, the competitive ratio α of EDF is bounded by
p·u0i
P P
i∈I p·ui + i∈{1,...,M }\I M ·u
α≤ ≤ = β, (4.1)
M ·u
P
i∈{1,...,M } p·ui
where u = max(ui ) and u = min(ui ). EDF is β-competitive.

Tightness of the bound. Consider a system with two charging
points and four vehicles for charging. Charging power is set as 1. Two
vehicles are set as (0, 1, 1, u1 ), and the other two are set as (0, 1 +
ε, 1, u2 ), where u1 < u2 . For EDF, as depicted in Fig. 4.6(a), the gained
utility of the first two vehicles is 2u1 , and that of the other two is 2ε·u2 .
For Opt, as illustrated in Fig. 4.6(a), they are 0 and 2u2 respectively.
Thus, the competitive ratio α = 2u2 /(2u1 + 2ε · u2 ), which approaches
to β = u2 /u1 as ε → 0. The bound is tight and the theorem holds.
‫ݑ‬ଵ ‫ݑ‬ଶ ‫ݑ‬ଶ

Port1 t Port1 t
0 ͳ ͳ൅ߝ 0 ͳ ͳ൅ߝ

Port2 t Port2 t
0 ͳ ͳ൅ߝ 0 ͳ ͳ൅ߝ
(a) EDF schedule (b) Opt schedule
Figure 4.6: Illustration of tightness for Theorem 4.2. Text in rectangles: unit utility.
Theorem 4.2 confirms that the competitive ratio of EDF is non-

constant. Furthermore, the performance gap between EDF and the
optimal schedule is unpredictable if the dynamic range of unit utility
is unbounded. Thus, an on-line algorithm with constant competitive
ratio may be more appealing. We will analyze an algorithm in first-fit
style and demonstrate its constant competitiveness in Section 4.3.2.
We have the following corollary based on Theorem 4.2.
Corollary 4.3. EDF is optimal if all electric vehicles have the same unit
utility, i.e., β = 1.
It is well-known that as to global multiprocessor (or multiple-

charging-point here) scheduling, EDF is not optimal for meeting dead-
lines. By contrast, this corollary indicates that if all customers can be
equally satisfied, EDF is optimal with respect to maximizing social wel-
fare or optimizing user satisfaction. In other words, EDF can process
charging demand as much as the off-line optimal algorithm does.
4.3.2 A 2-Competitive Scheduling Algorithm

We analyze a first-fit style algorithm called highest-utility-first (HUF).
The algorithm always first charges EVs with the highest unit utilities.
As we know, the algorithm is similar to the efficiency (profit to weight
ratio) first algorithm of the knapsack problem [124], which first packs
items with highest efficiencies. However, our system model is different
from that of the knapsack problem. The difference includes the metered
model, deadline and preemption. Hence, the analysis for the efficiency
first algorithm can not be applied to our problem. To bound the com-
petitive ratio of HUF, we also leverage the utility mapping scheme
described above. Using a similar proof skeleton to Theorem 4.2, we
prove a 2-competitiveness for HUF, i.e., HUF always gives a welfare at
least half of the optimal.
Theorem 4.4. HUF is 2-competitive and the bound is tight.
Proof. Bounding competitive ratio. Suppose that at time t,
EV1 , ..., EVM are the M vehicles being charged by HUF with u1 ≥
u2 ≥ ... ≥ uM , and EV10 , ..., EVM 0 are the M vehicles being processed
by an off-line optimal schedule Opt. Some EVi s may be identical to

some EVi0 s. Thus, without loss of generality, we assume EVi = EVi0 for
i ∈ I ⊂ {1, ..., M }. We discuss the two individual cases as follows.
Case 1 (i ∈ I): The utility of EVi0 by Opt can be mapped to HUF
at time t at most once, since the EVi0 utility is either mapped to time t
P
or a time before t. Thus, the total utility mapped is at most i∈I p·ui .
Case 2 (i ∈ {1, ..., M }\I): First, according to the scheme, the utility
of the set of EVi0 either is not mapped to HUF at time t, or if it is, the
following condition holds. We must have p·u0i ≤ p · uM because these
vehicles are not finished for charging and not chosen by HUF. The
utility (of EVi0 ) mapped is at most (M − |I|) · p·uM , where |I| is the
number of vehicles in set I. Second, the set of EVi is not being charged
in Opt at time t, and they may be mapped to time t from a later time
P
in Opt. Thus, the utility (of EVi ) mapped is at most i∈{1,...,M }\I p·ui .
Note that the utility obtained by the HUF schedule at time t is
P
i∈{1,...,M } p·ui . The competitive ratio (α) of HUF is thus bounded by
p·ui + (M − |I|) · p·uM

P P
i∈I p·ui + i∈{1,...,M }\I
α≤ P
i∈{1,...,M } p·ui
+ (M − |I|) · p·uM
P
i∈{1,...,M } p·ui
= P
i∈{1,...,M } p·ui
+ M · p·uM
P
i∈{1,...,M } p·ui
≤ P (4.2)
i∈{1,...,M } p·ui
P P
i∈{1,...,M } p·ui + i∈{1,...,M } p·ui
≤ P = 2. (4.3)
i∈{1,...,M } p·ui
ͳ൅ߝ ͳ ͳ൅ߝ
Port1 t Port1 t
0 ͳ ʹ 0 ͳ ʹ
ͳ൅ߝ ͳ ͳ൅ߝ
Port2 t Port2 t
0 ͳ ʹ 0 ͳ ʹ
(a) HUF schedule (b) Opt schedule
Figure 4.7: Illustration of tightness for Theorem 4.4. Text in rectangles: unit utility.
Eqn. (4.2) is true because of that the set I may be empty, i.e., |I| may
be equal to zero. As mentioned above, we must have u1 ≥ ... ≥ uM
according to HUF. Thus, M · p·uM ≤ i∈{1,...,M } p·ui and thus we have
P
Eqn. (4.3).
Tightness of the bound. Consider a system with two charging
points and four vehicles for charging. Charging power is set as 1. Two
vehicles are set as (0, 2, 1, 1 + ε), and the other two are set as (0, 1, 1, 1),
where ε > 0. As shown in Fig. 4.7(a), HUF only gains the utility of the
first two vehicles and none for the other two due to deadline miss. As
depicted in Fig. 4.7(b), the off-line optimal algorithm obtains all four
vehicles’ utility. The competitive ratio α = 4/(2+2ε), which approaches
to 2 as ε → 0. Hence, the bound is tight and the theorem holds.
As defined in Section 4.2.1, the unit utility of a charging task is a

piecewise function and each subtask corresponds to one piece. That is,
the unit utility varies between subtasks and is fixed for each individual
subtask. Thus, by HUF, subtasks being served always have higher unit
utilities than do suspended charging (sub)tasks. The former subtasks
can never be preempted by the latter ones. Hence, this utility definition
guarantees that HUF will not behave in the same way as LLF with
unbounded number of preemptions or mode transitions.
4.3.3 Bounding All Work-Conserving Schedules
We present a stronger result through showing that any work-conserving

algorithm is bounded. It is worth to note that the performance bound is
applicable to the whole class of work-conserving algorithms, not just to

a specific algorithm. In the following, we first define a work-conserving
schedule and then analyze the upper bound.
Definition 4.3 (Work-conserving schedule). A schedule S is called work-
conserving if the following is satisfied: at any time t, if there are unfin-
ished vehicles with ai ≤ t < di , then S(t) 6= ∅, where S(t) is the set of
vehicles being charged in schedule S at time t.
Definition 4.3 means that the system can not be idle if there are still
unfinished and unexpired charging demands. Note that the algorithms
discussed in this chapter are all work-conserving. Moreover, the optimal
off-line schedules must be work-conserving, otherwise a larger social
welfare can be obtained by rescheduling vehicles to fill the idleness.
Theorem 4.5. Any work-conserving algorithm is (1 + β)-competitive,
i.e., 1 + β is the upper bound.
Proof. The proof is similar to that of Theorem 4.2. Also suppose that at
time t, EV1 , ..., EVM are the M vehicles in a work-conserving schedule
S, and EV10 , ..., EVM
0 are in Opt. The difference lies in Case 2. The set
of EVi that is not in Opt but in S, may be mapped to time t from a

later time. Hence, the competitive ratio of schedule S is
p·ui 0 +
P P P
i∈I p·ui + i∈{1,...,M }\I i∈{1,...,M }\I p·ui
α≤ P
i∈{1,...,M } p·ui
p·ui 0
P
i∈{1,...,M }\I M ·u
=1+ P ≤1+ = 1 + β, (4.4)
i∈{1,...,M } p·u i M ·u
where u = max(ui ) and u = min(ui ). The theorem holds.
The ratio 1 + β is a performance upper bound for the whole class of

work-conserving algorithms. EDF and HUF are also work-conserving
algorithms and they have better performance bounds. For example, the
competitive ratio of HU F is 2 ≤ 1 + β. Based on Theorem 4.5, we can
obtain the performance bound for a batch of on-line algorithms such as
first-come-first-serve (FCFS) and shortest-job-first (SJF). FCFS always
first processes electric vehicles with the earliest arrival times; while SJF
always charges vehicles with the least charging demands.
Table 4.1: Summarization of theoretical results using competitive analysis.
Algorithm EDF HUF FCFS/SJF work-conserving

Bound β 2 1+β 1+β
Tight Yes Yes Yes No
Corollary 4.6. Both FCFS and SJF are (1 + β)-competitive and the
bound is tight.
Proof. The bound directly follows Theorem 4.5 and we only need to
prove the tightness. Consider a system with two charging points and
four vehicles for charging. The charging power is set as 1. For FCFS,
two electric vehicles are set as (0, 2, 1, u1 ), and the other two are set as
(ε, 1, 1 − ε, u2 ), where u1 < u2 . The gap between Opt and FCFS is α =
(2·u1 +2·u2 )/2·u1 = 1+β. For SJF, two vehicles are set as (0, 2, 1, u1 ),
and the other two are set as (0, 1 + ε, 1 + ε, u2 ), where u1 < u2 . The gap
between Opt and SJF is α = (2(1+ε)·u1 +2(1−ε)·u2 )/(2·u1 +2ε·u2 ),
which approaches to 1 + β as ε goes to zero. Therefore, the corollary
holds.
Table 4.1 summarizes the theoretical results based on competitive

analysis in this section. As mentioned above, multiple charging points
can be abstracted as the multi-processer case as to the scheduling the-
ory. Thus from a technical perspective, the theoretical results in this
section can be seen as (i) a generalization to [123], which only focuses
on the uni-processor case, and (ii) an extension to [109], which solely
studies competitive analysis for HUF.
4.4 Algorithm Analysis Based on Resource Augmentation
In this section, we leverage an analysis technique called resource

augmentation for performance analysis. Resource augmentation as a
method for analyzing on-line algorithms, first appears in [125, 126].
An on-line algorithm is given more resource than the optimal off-line
schedule with which it is compared. The idea is to analyze how much
additional resource the on-line algorithm is needed to achieve the op-
timality. For a park-and-charge system, the additional resource means
4.4. Algorithm Analysis Based on Resource Augmentation 85
either higher charging power or more charging points. We first present

the analysis for EDF and propose the results for the whole class of
work-conserving algorithms. In the following theorems, p and M de-
note the charging power and the number of charging points for an
optimal off-line schedule respectively.
4.4.1 Analysis for EDF

In Section 4.3.1, we discuss EDF using competitive analysis and prove
its non-constant competitive ratio. Because constant bounds are more
appealing, one pending question is whether EDF has a constant per-
formance bound if using other analysis techniques. We thus turn to
resource augmentation. The authors in [109] adopts resource augmen-
tation to analyze EDF for the uni-processor (or single charging point)
case. We make a significant generalization to that and deal with the
case of multiple charging points. We demonstrate that EDF also has no
constant performance bound according to the following two lemmas.
This result partly answers the pending question through eliminating
one more analysis technique.
Lemma 4.7. EDF is optimal with charging power β · p.
Proof. Bounding augmented resource. The proof is similar to that

of Theorem 4.2. Thus, we only highlight the difference as follows. Sup-
pose that EDF is with charging power of x · p and Opt is with power
p. The utility by EDF is thus x · i∈{1,...,M } p·ui . The mapped utili-
P
P
ties for Case 1 and Case 2 keep unchanged, which are i∈I p·ui and
P 0
i∈{1,...,M }\I p·ui respectively. Thus, we have
p·ui + i∈{1,...,M }\I p·u0i

P P
i∈I M ·u β
≤ = .
x · i∈{1,...,M } p·ui x·M ·u
P
x
Hence, to make EDF is 1-competitive (i.e., optimal), we let the above

equation less than or equal to 1. That is, if x = β, i.e., the charging
power is augmented to β · p, EDF is optimal.
Tightness of the bound. We prove this tightness using the same
example as that in Theorem 4.2. For EDF with augmented resource,
as depicted in Fig. 4.8(a), the gained utility of the first two vehicles is

Port1 t Port1 t
0 ͳΤߚ ͳ ͳ൅ߝ 0 ͳ ͳ൅ߝ

Port2 t Port2 t
0 ͳΤߚ ͳ ͳ൅ߝ 0 ͳ ͳ൅ߝ
(a) EDF schedule (b) Opt schedule
Figure 4.8: Illustration of tightness for Lemma 4.7. Text in rectangles is unit utility.
2u1 and that of the other two is 2u2 · (1 + ε − u1 /u2 ) respectively. For
Opt, as shown in Fig. 4.8(b), they are 0 and 2u2 respectively. Thus,
the competitive ratio is α = 2u2 /(2u2 + 2ε · u2 ), which approaches to 1
as ε → 0. The lemma holds.
Lemma 4.8. EDF is optimal with dM · βe charging points.
Proof. Bounding augmented resource. The proof is also similar

to that of Theorem 4.2 and we highlight the difference. Suppose that
EDF is with x charging points and Opt is with M charging points. The
P
utility by EDF is thus i∈{1,...,x} p·ui . The mapped utilities for Case 1
and Case 2 keep unchanged. Thus, we have
p·u0i
P P
i∈I p·ui + i∈{1,...,M }\I M ·u M ·β
≤ = .
x·u
P
i∈{1,...,x} p·ui x
Hence, to make EDF is 1-competitive, we let the above equation less

than or equal to 1. That is, if x = M · β, i.e., the number of charg-
ing points is augmented to dM · βe (the number is integral), EDF is
optimal.
Tightness of the bound. This tightness can be proved by con-
structing two different schedules as follows. One is Opt, where all sched-
uled vehicles have the highest unit utilities. The other is the EDF with
more charging points, where all scheduled vehicles have the lowest unit
utilities. For example, consider a system with two charging points. Sup-
pose that there are two vehicles with unit utility u2 and later deadlines,
and 2β vehicles with u1 and earlier deadlines, where u2 > u1 and
β = u2 /u1 . For Opt, the welfare is 2 · u2 . For EDF with 2β charging

points, the welfare is 2β · u1 . It is easy to verify that the competitive
ratio is 1. Hence, the lemma holds.
4.4.2 Analysis for Work-Conserving Schedules

Using similar proofs to those in the above lemmas, we prove that HUF
has a constant bound based on resource augmentation, as illustrated
in Lemma 4.9. We further have a stronger result, Theorem 4.10, which
is an upper bound of augmented resource for the whole class of work-
conserving algorithms. We omit their proofs to avoid repetition.
Lemma 4.9. HUF is optimal with charging power 2p or with 2M charg-

ing points.
Theorem 4.10. Any work-conserving algorithm is optimal with charg-

ing power (1 + β) · p or dM · (1 + β)e charging points.
Corollary 4.11. Both FCFS and SJF are optimal with charging power
(1 + β) · p or with dM · (1 + β)e charging points.
Table 4.2 summarizes the theoretical results based on resource aug-

mentation in this section. One application of these results is when a
park-and-charge system plans to upgrade its charging infrastructure.
These results can answer questions such as how much charging power
and how many charging points need to be added in order to satisfy
customers to a certain degree. From a technical perspective, the theo-
retical results in this section can be seen as generalization to that of the
uni-processor case studied in [123, 109]. Further, as mentioned above,
our work adopts the metered task model and has a goal of welfare
maximization. This differentiates it from those existing works such as
[125, 126] that use deterministic task model and optimize pure tempo-
ral metrics such as response time and deadline miss.
We evaluate the performance of four different algorithms including

HUF, EDF, FCFS and SJF. As mentioned in Section 4.3.1, LLF is
Table 4.2: Summarization of theoretical results based on resource augmentation.

p (M ) denotes the charging power (the number of charging points) for an optimal
off-line schedule.
Algorithm EDF HUF FCFS / SJF work-conserving

Charging Power β·p 2·p (1 + β) · p (1 + β) · p
Charging Point No. dβ · M e 2·M d(1 + β) · M e d(1 + β) · M e
Tight Yes Yes Yes No
impractical for park-and-charge system and thus is excluded in our

evaluation. We carry out comparisons between them and the off-line
optimal solution (Opt). To derive the optimal, we formulate an integer
linear program (as shown in Eqn. (4.5) to Eqn. (4.9)) for the charg-
ing scheduling problem, which is then solved using IBM ILOG CPLEX
[127]. We study the performance of these algorithms under different
ranges of unit utility (i.e., different values of β), and different scales of
charging resource (i.e., different values of p and M ), different charging
load of vehicles (i.e., different numbers of EVs and charging demand).
The following formulates the integer linear program that derives
the optimal value:
X X X
max ui · xijt , (4.5)
i j t
X X
s.t. ∀i, p · xijt ≤ ei , (4.6)
j t
X X
∀t, xijt ≤ M, (4.7)
i j
X
∀i, t, xijt ≤ 1, (4.8)
j
∀i, j, {t|t ≤ ai − 1 ∨ t ≥ di + 1}, xijt = 0. (4.9)
where xijt = 1 if EVi is being charged by charging point j at time
t; otherwise, xijt = 0. The ILP uses a discrete time model where t =
1, 2, 3, .... Objective Eqn. (4.5) is to maximize social welfare. Constraint
Eqn. (4.6) guarantees that the total charged energy of a vehicle is not
more than its charging demand. Constraint Eqn. (4.7) guarantees that
at any time, there are at most M vehicles being charged concurrently.
Constraint Eqn. (4.8) guarantees that no vehicle can be charged in
parallel. Constraint Eqn. (4.9) guarantees that that vehicles can only
be charged between its arrival time and deadline. The ILP is formulated
in a discrete time manner, i.e., variables are defined for each time unit.
However, the discussed on-line algorithms have no such limitation and
apply to both the continuous time and discrete time cases.
4.5.1 Experimental Setup
We consider a park-and-charge system with M charging points and

each charging point supports a charging power of p kw. We compare
the social welfare in a time duration of T = 24h between the above
algorithms and Opt. To make the integer linear program solvable, the
time unit is set as 0.1h and all parameters except unit utility are set
to be integral. Note that there is no such limitation for the algorithms
discussed in this chapter. During the 240 time units, there are N electric
vehicles that come to the system for charging. The arrival time ai is
uniformly generated from [1, 240] time unit. The deadline di , charging
demand ei and unit utility ui are chosen from uniform distributions in
[10, D] time unit, [5, C] kW h and [1, β], respectively. Note that each
figure in the following plots the results that are already normalized to
the maximum value in the corresponding setting. The following shows
the simulation results as well as the corresponding result analysis.
4.5.2 Experimental Results
Welfare Comparison
Fig. 4.9 and Fig. 4.10 demonstrate the welfare comparison under differ-
ent settings between the four algorithms and the off-line optimal result.
First, HUF significantly outperforms the other three algorithms on wel-
fare maximization across nearly all settings. The reason is that HUF
always first schedules electric vehicles with larger unit utilities, while
other three algorithms schedule vehicles based on their temporal char-
acteristics. Furthermore, HUF performs nearly as well as Opt across all
of the settings. In practice, the performance gap between HUF and Opt
is much smaller than the theoretical bound, and so do the other three
algorithms. For example, when β = 8 (in Fig. 4.9(c)), the theoretical
bound of EDF is 8 according to Theorem 4.2, but the experimental per-
formance gap between EDF and Opt is about 1.4. Another observation
is that EDF, FCFS and SJF have nearly the same performance (i.e., the
three curves overlap with each other), though their theoretical perfor-
mance bounds are quite different (β- and 1 + β-competitiveness). This
observation infers that only differentiating vehicles’ temporal charac-
teristics helps little on welfare maximization with the metered charging
model, when the unit utility range β is large.
Sensitivity Analysis
Impact of charging point number. Fig. 4.9(a) plots social welfare

varying with the number of charging points. First, the welfare of all four
algorithms as well as Opt increases as the number of charging points
rises. This observation follows the intuition that more charging points
can process more vehicles and thus results in higher welfare. Second,
an interesting observation is that as the charging point number grows,
the difference on welfare between Opt (or HUF) and the other three
algorithms first increases and then decreases. The reason is two-fold.
On one hand, all algorithms and Opt can process only a small amount
of vehicles when charging points are scarce (e.g., M = 1), and they
can finish almost all charging demands when charging points become
abundant (e.g., M = 20). On the other hand, in-between (e.g., M =
5 to 15) is when much difference occurs to the set of vehicles that these
algorithms can schedule, and so does the welfare. Third, the augmented
resource (charging points) in practice is less than the theoretical bound.
For example, the bound for EDF is dβ · M e. However, the EDF value at
M = 20 is much larger than the Opt value at M = 5. Similar remarks
can be also made for HUF, FCFS and SJF.
Impact of charging power. Fig. 4.9(b) depicts how welfare varies
with the charging power. For this setting, we have similar observations
and analysis to those for Fig. 4.9(a). We thus only outline them as
follows. First, the welfare of all four algorithms and Opt increases as
the charging power rises. Second, the difference on welfare between
Opt (or HUF) and the other three algorithms first goes up and then
declines. Third, the augmented resource (charging power) in practice is
less than the theoretical bound, e.g., the EDF value at p = 20 is larger
than the Opt value at p = 5, where charging power grows 4 times.
Normalized Social Welfare

0.8
0.6
Opt
EDF
0.4
HUF
0.2 FCFS
SJF
0
1 5 10 15 20
Charing Point Number M
(a) p = 10, β = 4
1
0.8
0.6
Opt
EDF
0.4
HUF
0.2 FCFS
SJB
0
1 2 5 10 20
Charging Power p
(b) M = 10, β = 4
1
0.8
0.6
Opt
0.4 EDF
HUF
0.2 FCFS
SJF
0
1 2 4 6 8
Utility Range E
(c) M = 10, p = 10
Figure 4.9: Experimental results for varying charging point number (M ), charg-
ing power (p) and unit utility range (β). Parameters on vehicle number, charging
demand and deadline are set as N = 400, C = 20, D = 60.

0.8
0.6
Opt
0.4 EDF
HUF
0.2 FCFS
SJF
0
100 200 400 600 800
Vehicle Number N
(a) C = 20, D = 60, β = 4
1
0.8
Opt
0.6
EDF
HUF
0.4 FCFS
SJF
0.2
10 15 20 25 30
Charging Demand C
(b) N = 200, D = 60, β = 8
1
Normalized Socail Welfare
0.9
Opt
0.8 EDF
HUF
0.7 FCFS
SJF
0.6
10 30 60 90 120
Deadline D
(c) N = 200, C = 20, β = 8
Figure 4.10: Experimental results for varying parameters on vehicle number (N ),

charging demand (C) and deadline (D). Charging point number and charging power
are set as M = 10, p = 10.
Impact of unit utility range. Fig. 4.9(c) demonstrates welfare

varying with the range of unit utility. As the utility range grows, the
welfare of all four algorithms and Opt increases, and the performance
gap becomes wider and wider. As to flat unit utility (i.e., β = 1), the
welfare derived by EDF is equal to the Opt value. This observation
validates Corollary 4.3. The other three algorithms also perform nearly
the same as Opt. This is because that since the unit utility is flat, the
resulting gap between them and Opt is only caused by temporality,
which is thus rather small. When the range of unit utility is β = 8,
the performance difference between Opt and EDF, FCFS, or SJF be-
comes to about 40%. By contrast, HU F still performs nearly as well
as Opt. This observation indicates the robustness of HUF on welfare
maximization.
Impact of vehicle number. Fig. 4.10(a) plots welfare varying
with the number of electric vehicles. First, the welfare of all four
algorithms and Opt increases as the number of vehicles rises. This
is because there are more vehicles scheduled for charging. However,
the increasing rate on welfare decreases. For example, as to Opt and
HUF, the welfare increases more than 75% from N = 100 to N = 200;
while it is less than 6% from N = 600 to N = 800. For the other three
algorithms, there is nearly no change on welfare from N = 200 to
800. The reason is as follows. When the number of vehicles is small,
the system is lightly loaded and all charging demand can be fulfilled.
As the number grows, the system’s charging load becomes heavier
and heavier. After it reaches the load limit that an algorithm can
schedule (e.g., N = 200 for EDF), the welfare by the algorithm stays
nearly unchanged. Second, because of the same reason, the difference
on welfare between Opt or HUF and the other three algorithms
grows as the number of vehicles rises, especially when N ≥ 200. This
observation indicates that HU F is more suitable to systems that tend
to be overloaded. As to the charging service, the park-and-charge
system is such a system. In the real-life scenario, the charging point
or power supply in one system is limited compared with the parking
space and is an even scarcer resource compared with the scale of
electric vehicles.
Impact of charging demand. Fig. 4.10(b) demonstrates welfare

varying with the charging demand of vehicles. In reality, it seldom
happens that all vehicles have the same charging demand. Thus for the
charging demand range [5, C] kW h, we set that C varies from 10 kW h.
For this setting, we can make similar observations and analysis to those
for Fig. 4.10(a). For example, the welfare by Opt and HUF increases as
the charging demand rises, and the increasing rate decreases. Instead of
repeating them, we only highlight a different observation as follows. The
social welfare by EDF and FCFS even declines as the charging demand
grows, which is rather counter-intuitive. The reason for EDF is that
vehicles with earlier deadlines may be with lower unit utility. Increasing
the charging demand of these vehicles may make the system schedule
less vehicles with higher unit utility. Similarly, the reason for FCFS is
that increasing the charging demand of vehicles with earlier arrival time
may reduce the serving time for vehicles with higher unit utility. This
observation reveals the performance anomaly of the temporality-based
methods on welfare maximization.
Impact of deadline. Fig. 4.10(c) shows welfare varying with the
deadline of vehicles. First, the welfare of all four algorithms and Opt
increases as the deadline rises. The reason is that longer deadlines make
more charging demand be processed and thus results in higher welfare.
Second, the increasing rate on welfare decreases as the deadline grows.
For example, the welfare by HUF grows about 10% from D = 10 to
D = 30; while it stays nearly uncharged from D = 90 to D = 120.
This is because that the welfare benefits little from deadline increasing
after the deadline is large enough to make all charging demand fully
satisfied.
Event-driven vs. Time-driven
The above comparisons (Section 4.5.2 and 4.5.2) confine to event-driven

algorithms. Here shows the comparison between event-driven and time-
driven policies. Time-driven policy has been widely used for general
EV charging problems in existing works such as [101, 41, 26, 102, 103,
39], but it is inadequate for the park-and-charge scenario compared to
4.6. Summary 95
Social Welfare
Social Welfare
1 1
0.5 HUF(TT) 0.5 EDF(ET)

HUF(ET) EDF(TT)
0 0
1 10 20 30 40 50 60 1 10 20 30 40 50 60
Time Slot Length Time Slot Length
(a) Comparison on HUF (b) Comparison on EDF
Figure 4.11: Comparisons about event-driven and time-driven versions of HUF

and EDF. ET: event-driven; TT: time-driven.
event-drive policy. This fact is validated by Fig. 4.11, which depicts

exemplary comparisons using event-driven and time-driven versions of
HUF and EDF. We can see that the event-driven versions perform
better than the time-driven versions. Their difference becomes larger
as the time slot length increases. The reason is that by time-driven
policy, the scheduler is activated only at the beginning of each time
slot. If a charging task arrives within a time slot, it needs to wait
for being scheduled until the beginning of next time slot. The longer
the time slot is, the more waiting time the task would have. For the
extreme case that a charging task arrives just after a time slot begins,
it will delayed for an entire time slot. Thus time-driven policy can much
compromise the social welfare as to park-and-charge.
4.6 Summary
This chapter studies a park-and-charge system and focuses on the on-

line charging scheduling problem for maximizing social welfare or EV
user satisfaction. We propose to adopt the metered model to char-
acterize charging tasks and employ event-driven methods to schedule
charging tasks. Our goal is to perform both theoretical and experi-
mental analysis for event-driven algorithms adapted to the metered
model. Theoretical results show the varying performance bound of
temporality-based algorithms such as EDF, FCFS, and SJF, and the
constant performance bound of the value-based algorithm, HUF. Simu-
lation results further demonstrate the near-optimality and performance
consistency of HUF, and the sub-optimality and performance anomaly

of the temporality-based algorithms. These results indicate that HUF
has better robustness than those temporality-based algorithms do in
terms of maximizing social welfare. Hence, adopting HUF in park-and-
charge systems can provide better charging service to customers and
thus make them better satisfied.
5
Charging Demand Balancing
Public availability of fast charging stations is intended to shorten charg-

ing time and thus to advocate the popularity of EVs. However, com-
petition for charging resources degrades quality of service and can sig-
nificantly compromise this intent. For example, EVs may swam into a
charging station with favorable parameters such as lower energy price
and shorter distance [128, 129]. Their choices would cause the charg-
ing station heavily overloaded and make some users leave to search
again for unoccupied charging space. Energy and time wasted by the
re-searching counteract their original intents of saving money and re-
ducing charging duration. This chapter studies the upper level of our
solution framework, which is how to balance charging demand among
multiple public charging stations.
EV user behavior is hardly controllable in reality and thus it is im-
practical to assume users exactly following scheduling decisions. For the
charging system, the scheduler should involve the uncertainty of user
behavior when making scheduling decisions. For EV users, the schedul-
ing decisions can be seen as recommendation that aid them to carry
out decision making. We design two approximate demand balancing al-
gorithms that are compatible with the above facts. With the objective
97
98 Charging Demand Balancing
of optimizing the expected user satisfaction, they are 3-approximate

and (2 + ε)-approximate respectively, that is, their solutions are at
least 1/3 and 1/(2 + ε) of the optimal respectively. We evaluate our
algorithms using realistic data traces and simulation tools. The results
demonstrate that our approaches perform close to the optimal in prac-
tice in spite of rather large theoretical performance bounds. They are
much better than myopic or user-self-determined approaches regarding
overall QoS improvement.
The rest of the chapter is organized as follows. Section 5.1 presents
recent development related to charging demand balancing. Section 5.2
presents the system model and problem statement. Section 5.3 pro-
poses charging demand balancing algorithms. Section 5.4 evaluates the
proposed algorithms. Section 5.5 gives the summary of this chapter.1
There are usually two different ways that can address charging demand
balancing. One way carries out demand balancing from a global view,
while the other way performs decision-making with a scope of each
individual vehicle. This section provides reviews of existing works that
follow these two ways respectively.
Global Demand Balancing. The authors of [129] present an al-
gorithm that allocates EVs among multiple stations to improve quality
of service. They carry out performance analysis through modelling a
set of charging stations using a multi-queue system. One analysis result
is that the quality of service can be maximized if the user behavior can
be controlled. To achieve such good performance in reality where EV
users choose charging stations spontaneously, the authors leverage price
control to provide incentives and motivate users to follow the optimal
allocations. Similarly, the authors of [110] also model charging stations
as multiple queues. They consider a different application scenario from
[110], which is a highway scenario where charging stations are built at
the entrances/exits of highways. EVs need multiple times of recharg-
ing during their trips. The problem studies is how to schedule EVs
1
both temporally and spatially to these charging stations. Based on the

status information from charging stations, the reservation decisions of
EVs are periodically made in order to minimize their waiting time. The
authors of [14] also use queueing model but differently, they propose
a decentralized stochastic demand balancing algorithms. Basically, the
algorithm running on the charging station end is to broadcast status
information with some probability, while the algorithm executing on
the EV end is to listen the broadcast status information with some
probability.
There are several key differences between our work and the above
existing works. First, we explicitly consider the uncertainty of user
behavior into the algorithm design. Considering uncertainty makes the
charging demand balancing much harder by changing it to a NP-hard
problem. Second, we propose approximate algorithms with provable
performance guarantees. Further, we evaluate the proposed algorithms
with realistic traces and a simulation tool, which makes our result and
its analysis closer to the reality.
EV Routing. Traditional vehicle routing problems focus on find-
ing shortest paths in road networks. Due to the unique characteristics
of EVs such as limited cruising range, long charging times, and the
capability of energy regeneration, it requires new routing algorithms
to determine the most economical routes rather than just the shortest
ones. Thus, most routing algorithms revise existing shortest path
algorithms through considering the above characteristics. For example,
the authors of [44] adapt A∗ search algorithm and the authors of
[128, 130, 131] present Dijkstra-like algorithms. Although EV routing
algorithms seem to provide EV users with optimal solutions, the
myopic or self-determined decision making process of these algorithms
can cause unbalanced demand allocation, i.e., some charging stations
are overloaded while some are under-utilized. To overcome this
potential negative result, our work leverages the way of global demand
balancing.
5.2 System Model
We consider a time snapshot of a charging system with M stations,

which belong to one entity. At the time, there are N EVs that ex-
pect to use the charging service. They report their information such
as charging demand and positions to the system, based on which the
scheduler in the upper level decides how to schedule these EVs to the
charging stations. Assuming users exactly following the scheduling deci-
sion is questionable, because EV user behavior is hardly controllable in
reality. Thus, as to the charging system, the scheduler should consider
the uncertainty of user behavior when making scheduling decisions. As
to EV users, the scheduling decisions can be seen as recommendation
aiding them to perform decision making. This means that users have
full control on when and where to charge their vehicles, that is, users
may or may not choose to follow the decisions made by the scheduler of
the system. Our work is compatible with the above facts. The follow-
ing provides models for the user behavior uncertainty and the charging
cost, which are adopted by our charging demand balancing methods.
Charging cost. Each EVi ’s charging cost consists of three parts.
The first part is the energy demand cost: Cij dem = B (SOC 0 − SOC )q ,
i i i j
where qj is the energy price of station CSj , SOCi and SOCi0 are the
current and demanded state of charge respectively, Bi is the battery
capacity of EVi [25]. The energy price can be different across charg-
ing stations, and after an EV chooses a charging station, the station’s
energy price for the EV stays unchanged until its charging is finished
[129, 132]. The second part is the detour cost, which is incurred by the
detour made to a user’s destination because of charging. The cost is
calculated as Cij dtr = β dtr q , where EV travels dtr more distance
i ij j i ij
by the detour than by the original route to its destination and βi is
its unit energy consumption (e.g., kW h/km). Emerging smart charg-
ing systems allow users to reserve a charging point and pay for the
reservation [12]. Thus, the third part is this reservation cost, which de-
pends on the time length lij of the reservation and the energy price:
rsv = c (l , q ). Our methods do not relay on any specific realization
Cij j ij j
of cj (·). In addition, discussion of the detailed pricing policy for the
reservation, i.e., the realization of function cj (·), is out of the scope of
this chapter. Hence, the EVi ’s charging cost if going to charging station
CSj is the sum: Cij = Cij dem + C dtr + C rsv .
ij ij
We further introduce a valuation function for each EVi as Vi =
vi (SOCi , SOCi0 ), which defines the gained value when charging the
EV from SOCi to SOCi0 . This function is usually defined to be non-
decreasing as to the difference between SOCi0 and SOCi [133, 134, 39].
The utility for EVi is thus denoted as Wij = Vi − Cij if choosing to
charge at station CSj . This utility can be seen as how much the EV
user is satisfied if choosing station CSj . This modeling approach has
been widely-used in EV charging studies such as [133, 134, 39]. Simi-
larly, our methods are independent on any specific form of the valuation
function vi (·).
User behavior uncertainty. We assume that all EV users have
the same preferences on selecting which charging station to charge.
These preferences are as follows. (i) EV users prefer lower energy price
to higher energy price. (ii) EV users prefer shorter detour distance
(and shorter reservation time) to longer detour distance (and longer
reservation time). (iii) EV users are less sensitive to the energy price
and more sensitive to the detour distance when their SOC are low. That
is, the price and distance elasticity is smaller in EVs with lower SOC
than those with higher SOC. (iv) EV users are only concerned about
the absolute energy price and detour distance instead of relative price
and detour distance between different charging stations. The rationale
behind this preference is that when all stations are with too high energy
price or too long detour distance, EV users may choose none of them
and give up charging at that time. The charging price may vary within
a regulated range, and at some time, users may not choose to charge
their EVs due to high price. They may wait till the charging price
becomes lower than what they expect, and then choose to charge their
vehicles. These are common assumptions on EV users’ preferences in
EV charging studies, e.g., [129, 135, 132]. We can use a function Fij =
fij (qj , dtrij , SOCi ) to denote the probability on EVi choosing charging
station CSj . The probability on EVi not choosing station CSj is thus
1−Fij . Note that it is not necessary to satisfy that Fi1 +· · ·+FiM = 1, ∀i
according to the fourth preference. According to these preferences, fij (·)
is a non-increasing function as to the energy price qj , detour distance

dtrij and the current SOC SOCi . Some realizations of the probability
fi (·) include weighted linear and quadratic functions [129, 132]. By
contrast, our work depends on no specific realization of fi (·).
5.2.1 Problem Formulation

We use a notion of social welfare to define the sum of EVs’ utility.
The social welfare definition here is different from that given in Sec-
tion 2.2.1. The objective of this charging demand balancing problem
is to maximize the expected social welfare, and the problem can be
formulated as the following optimization problem.
X X
max Zij xij , (5.1)
i j
X
s.t. xij ≤ 1, ∀i, (5.2)
j
X
Fij xij ≤ Aj , ∀j, (5.3)
i
where Zij = Wij Fij ; xij s are binary variables: xij = 1 if the system
assigns station CSj to EVi ; otherwise, xij = 0; Aj is the remaining
capacity, i.e., the number of unoccupied charging points, at the time. In
the above optimization problem, the objective Eqn. (5.1) is for expected
social welfare maximization, where Zij is the expected utility when
assigning EVi to charging station CSj . The constraint Eqn. (5.2) limits
that each EV can be assigned at most one charging station. We use
P
“≤ 1" instead of “= 1", i.e., j xij may equal zero for some EVi s,
because the total remaining capacity of charging stations may be not
enough to serve all N EVs. The constraint Eqn. (5.3) restricts that the
expected number of EVs assigned to charging station CSj is not larger
than its remaining capacity.
5.3 Algorithm Design for Demand Balancing
The expected welfare maximization problem (EWMP) given by

Eqn. (5.1)-(5.3) can be interpreted as the generalized assignment prob-
lem (GAP), which is well known to be NP-hard. Note that if not con-
sidering the uncertainty of user behavior (i.e., Fij = 1, ∀i, j), problem
5.3. Algorithm Design for Demand Balancing 103
EWMP becomes very trivial. The optimal solution of this case can be
derived by first sorting utility Wij in a decreasing order and then from
the beginning of the sorted list, scheduling EVs as many as possible un-
til stations become no space. Many approximate algorithms exist for the
GAP problem [136, 137, 138]. We propose new algorithms to problem
EWMP or GAP. Our algorithms not only possess the same theoretical
performance guarantee as state-of-the-art, but also have much better
average performance in practice.
Approximate Algorithms
Algorithm M-X (Algorithm 2) deals with charging demand balanc-
ing for M ultiple stations. It uses Algorithm X as a subroutine, which
handles how to schedule EVs as to a single station. We will give two
examples for Algorithm X later on. The following is the definition of
approximation ratio, used in our algorithm analysis.
Definition 5.1 (α-approximation). A solution is defined to be an α-

approximation, if the solution is feasible and α·Z(g) ≥ Z(g ∗ ), where g ∗
is the optimal solution. An algorithm is defined as an α-approximation
if it always gives α-approximate solutions.
Description for Algorithm M-X. In line 3, Algorithm M-X uses

Algorithm X as a subroutine to schedule EVs to station CSj in each
jth recursive call. By line 4, the algorithm splits utility matrix ωj . The
operation by line 4 implies that ωj1 is identical to ωj regrading station
CSj . Moreover, if i ∈ Ψ e j , then ω 1 is assigned with the same utility
j
ωj [i, j] for all the remaining stations. EVi can be rescheduled to an-
other charging station later if the new utility (by scheduling it to the
new station) is larger than the current utility. Line 5 updates the utility
matrix for the next recursive call. The updated utility matrix may con-
tain negative entries. All the negative entries are not considered when
calling Algorithm X. Line 6 calls the next recursion and pushes the call
stack. Line 7 and 8 pops the call stack and returns the final assignment.
Note that line 7 guarantees that assignments with larger welfare cannot
be overwritten by ones with smaller welfare. Algorithm M-X is adapted
from that presented in [136], which is regarded as state-of-the art [137].
Algorithm 2: M-X. Charging Demand Balancing

Input : ∀i, j, Zij , Fij , Aj ;
Output: An assignment of charging stations to EVs;
I Initialization as follows:
1 Let w be an N × M utility matrix, and the entry w[i, j] is the
utility Zij when assigning charging station CSj to EVi ;
2 Let ωj be the utility matrix at jth recursive call, and ω1 ← w;
I Invoke the procedure Next-Station(j) with j = 1:
Procedure Next-Station(j)
3 Run Algorithm X on station CSj using utility matrix ωj ,
and let Ψe j be the returned set of EVs scheduled to
charging station CSj in the jth recursive call;
// Algorithm X represents an algorithm for assigning a single charging
station to EVs.
4 Split utility matrix ωj into two utility matrices ωj1 and ωj2
such that, ∀k(∈ [j, M ], ∀i ∈ [1, N ],
ωj [i, j], if (i ∈ Ψe j ) or (k = j)
ωj1 [i, k] ← ,
0, otherwise
ωj2 ← ωj − ωj1 ;
if j < M then
5 ωj+1 ← ωj2 ;
6 Do Next-Station(j+1), and let Ψj+1 be the returned
assignment set;
return Ψj ← {Ψj+1 ∪ (Ψ e j \ SM
7 k=j+1 Ψk )};
8 else
return Ψj ← Ψ
e j;
// Ψj is the set of EVs scheduled to station CSj .

5.3. Algorithm Design for Demand Balancing 105
Thus, its approximation ratio is (1 + α), where α is the approximation

ratio of Algorithm X, as stated by Claim 2.1 in [136]. In addition, the
time complexity of Algorithm M-X is O(M · Λ(N ) + N · M 2 ), where
Λ(N ) is the complexity of Algorithm X.
According to Algorithm M-X, for each recursion call, an EV could
be rescheduled to a new station even if it has been scheduled to a
station by previous calls. The reason is that the algorithm schedules
all EVs with positive utility by line 6, which includes two kinds of
EVs. The first is those EVs that have not considered yet. The second is
those EVs that have been scheduled and whose updated utility is still
positive. Since the algorithm goes through each station only once, this
scheduling operation may cause that the first M − 1 stations still have
space to take in more EVs. Hence, scheduling the remaining EVs to the
remaining space can further improve the welfare. In the following, we
propose an enhanced algorithm, which can further improve the average
performance of Algorithm M-X.
Description for Algorithm EM-X. Algorithm 3 contains a loop
of M iterations. In the kth iteration, the algorithm calls Algorithm M-
X to schedule the remaining EVs to the remaining space of stations
CS1 , ..., CSk . Charging station CSk is the final station considered by
Algorithm EM-X in the kth iteration. This station does not have suf-
ficient space to accept any of the EVs that are still unscheduled. Thus,
Algorithm EM-X removes this station and considers CS1 , ..., CSk−1
in the next iteration. In total, the algorithm calls Algorithm M-X
M times, and thus its time complexity is O(M 2 · Λ(N ) + N · M 3 ),
where Λ(N ) is the complexity of Algorithm X. By Theorem 5.1, Algo-
rithm EM-X has the same performance bound as Algorithm M-X. How-
ever, we expect that Algorithm 3 has better average performance than
Algorithm 2. Experimental results in Section 5.4 verify this expectation.
Theorem 5.1. Algorithm EM-X is a (1 + α)-approximation and the
approximation ratio is tight, where α is the approximation ratio of
Algorithm X.
Proof. The proof consists of two steps. The first step is to prove the per-
formance bound of the first iteration (when ΠCS = {CS1 , ..., CSM }).
The bound is (1 + α) because the algorithm employs Algorithm M-X.
Algorithm 3: EM-X. Enhanced Demand Balancing

Input : ∀i, j, Zij , Fij , Aj ;
Output: An assignment of charging stations to EVs;
1 Initialize station set: Π
CS ← {CS , CS , ..., CS };
1 2 M
2 Initialize EV set: Π
EV ← {EV1 , EV2 , ..., EVN };
3 Initialize the assignment: ∀j ∈ [1, M ], Ψj ← ∅;
4 for k ← M to 1 do
5 Run Algorithm M-X on ΠCS and ΠEV , and denote EVs
scheduled to CSj∈[1,k] as Ψ0j ;
∀j ∈ [1, k], Ψj = Ψj Ψ0j ;
S
6
// Combine EVs scheduled to CSj in each

iteration;
EV ← ΠEV \ Ψj 0 ;
S
7 Π
j∈[1,k]
// Remove EVs that has been scheduled;
8 ΠCS ← ΠCS \ CSk ;
// Remove CSk , which is the final station
considered by Algorithm M-X in each iteration;
9 ∀j ∈ [1, k − 1], update Aj as the remaining space;
10 return ∀j ∈ [1, M ], Ψj ;
The second step is to prove that the rest iterations give no im-
provement to this performance bound and further the bound is tight.
Suppose there are two single-point stations and two determinate EVs
(with probability F = 1 for both stations). The utility of EV1 to the
two stations is α > 1 and 0; while that of EV2 is 1 for both stations.
Since Algorithm X is an α-approximation, Algorithm E-X may re-
turn the solution that EV2 is scheduled to the first station. However,
the optimal solution is to schedule EV1 to the first station and EV2
to the second station. The welfare ratio between the two assignments
is (1 + α). Moreover, there is no rescheduling for this case and thus
iterations except the first one make no change to the assignment. The
bound keeps unchanged and is tight.
Examples for Algorithm X. We now return to the examples for

Algorithm X, which handle how to schedule EVs to a single station.
It is easy to know that this single-station EV scheduling problem can
be interpreted as the knapsack problem [124]. Any algorithm for the
knapsack problem can be the subroutine to Algorithm EM-X. We con-
sider two well-recognized types of algorithms for this problem, which
include greedy algorithms and fully polynomial approximation schemes
(FPTAS). A classic example for the first type schedules EVs with high-
est utility Wij as much as possible, which is 2-approximation. Details
on this algorithm can be found on [124, p. 34]. If using this example
algorithm as a subroutine, Algorithm 2 and Algorithm 3 are denoted
as M-Greedy and EM-Greedy, and they are thus 3-approximation by
Theorem 5.1. The second type mainly adopts dynamic programming.
A classic example for this type can be found on [124, p. 166], which
is (1 + ε)-approximation. Using this example algorithm as a subrou-
tine, Algorithm 2 and Algorithm 3 are denoted as M-FPTAS and EM-
FPTAS, and thus they are (2 + ε)-approximation by Theorem 5.1.
To this point we have introduced our algorithm for charging demand

balancing, and also provided its theoretical guarantee. To get a better
picture of our algorithm in practical settings, it is important to evaluate
Figure 5.1: The street map of Eichstätt, Germany, from OpenStreetMap [139],
the distribution of 100 EVs generated by SUMO [140] and 2 different settings of
charging stations. Yellow triangles: electric vehicles; red rectangles: stations.
it using real data, which is the goal of this subsection. We conduct

extensive simulations based on realistic data traces and simulation tools
to compare our algorithm with the current practice.
Experiment Setup and Methodology

Map and Traffic Data. Fig. 5.1 depicts a real-world street map from
OpenStreetMap [139]. The street map covers a 4.5km × 3km area and
has 2962 edges/streets. We use a simulation tool called SUMO [140]
to generate the traffic data of 100 electric vehicles including their cur-
rent positions, their destinations, the corresponding paths, and their
travelling speeds (ϑi ). These 100 EVs are distributed on the map ac-
cording to their current coordinates. Initially, there are three charging
stations evenly distributed on the map, as illustrated by the red rectan-
gles (CS1 , CS2 , and CS3 ). The default capacity of each charging station
is set as 10 fast charging points. The detour distance (dtrij ) and the
direct distance (dirij ) of each EV to each charging station can be then
calculated accordingly. The detour distance here is the distance differ-

ence between the detour path and the original route; while the direct
distance is the distance from the current position to a charging station.
Charging Cost. These charging stations have different energy
prices. We use the average on-peak spot electricity prices of Califor-
nia, Oregon and Illinois over 2012 [141] to represent the energy price
(qj ) for the three stations, which are 0.035$/kW h, 0.027$/kW h and
0.032$/kW h. They correspond to station CS1 , CS2 and CS3 respec-
tively. Each station has 10 charging points by default. Each EV’s bat-
tery capacity is set to 24kW h, the same as that of Nissan Leaf [96].
Their charging demand (i.e., required SOC minus the current SOC) is
randomly picked from the range (0, 80%]. The setting about the unit
energy consumption is βi = 0.15kW h/km, estimated as the maximum
range divided by the batter capacity of Nissan Leaf [96]. The travelling
time from an EV’s current position to station CSj is calculated by the
direct distance divided by its travelling speed. This time is treated as
the reservation time, during which station CSj may have to let one
charging point idle. The charging station may waste some energy for
the idling, and this energy is estimated by the charging power multi-
plied by the reservation time. The unit reservation cost is smaller than
the energy price, and its initial value is set as 0.5∗qj . Thus, the reserva-
rsv = 50∗0.5∗q ∗ dirij , where a charging point can support
tion cost is Cij j ϑi
50kW power. We set the valuation function as the maximum charging
cost among all charging stations. The utility for each EV is then its
cost saving if charging at station CSj . We use these settings for the
sake of simplicity, but our proposed method applies to any reservation
schemes and valuation functions.
EV User Behavior. For the probability on EVi choosing charging
station CSj , we take a weighted exponential distribution as an example
function in this evaluation, which is Fij = exp(1 − SOCi−1 ))∗gj (dtrij )∗
hj (qj ). Here, gj (dtrij ) ∈ {0.8, 0.95, 1} corresponds to charging stations
with the longest, medium, and shortest detour distance, respectively,
and hj (qj ) ∈ {0.8, 0.95, 1} corresponds to stations with the highest,
medium, and lowest energy price, respectively. This example function
satisfies all user preferences given in Section 5.2.
Experimental Methodology. Our evaluation includes two differ-

ent kinds of analysis. The first one, demonstrated in Section 5.4, is from
comparing eight different methods as follows. (i) Opt. This method is
by solving the Integer Linear Programming (ILP) problem given by
Eqn. (5.1)-(5.3) using a mathematical solver IBM ILOG CPLEX Op-
timizer [127]. It is seen as the optimal solution here. (ii) M-Greedy,
EM-Greedy, M-FPTAS and EM-FPTAS. Approximate algorithms pro-
posed in Section 5.3. The inaccuracy is initially set as ε = 0.1 for FP-
TAS algorithms. (iii) Min-Cost. EV users make decisions by themselves
and choose to charge their EVs at stations with the minimum charg-
ing cost. (iv) Min-Dir. EV users make myopic decisions and choose
to charge their EVs at stations with the minimum direct distance. (v)
Min-Dtr. EV users make decisions by themselves and choose to charge
their EVs at stations with the minimum detour distance. Min-Dtr is
usually included by EV routing algorithms such as [128, 44]. Note that
the above three methods may cause some charging stations overloaded
and others lightly loaded. The other kind of analysis is the sensitivity
analysis conducted in Section 5.4. We analyze the impact of different
parameters (including inaccuracy ε, station capacity, and energy price)
on the welfare and demand balancing. Suggests on how to set these
parameters in practice are also given accordingly.
Comparison among Different Methods
Min-Cost, Min-Dtr and Min-Dir may violate the constraint of stations’

capacity; while the other methods can always satisfy this constraint.
To do a fair comparison, after EV users make their decision on which
station to charge, the three methods, i.e., Min-Cost, Min-Dtr and Min-
Dir, let the charging station accept as many as possible EVs in an
increasing order of their direct distances to the station. Then, the three
methods calculate the expected welfare and number of EVs in each
station just before violating its capacity constraint.
Welfare comparison. Fig. 5.2(a) and 5.2(c) demonstrate the com-
parison on the expected welfare for two different charging station set-
tings: CS = {CS1 , CS2 , CS3 } and CS 0 = {CS10 , CS2 , CS30 }. CS is the
initial setting, where charging stations are evenly distributed on the
20 60
Expected Welfare
CS1 CS2 CS3
Expected EV No.
CS1 CS2 CS3
15
40
10
20
5
0 0
Opt EMG MG EMF MF Cost Dir Dtr Opt EMG MG EMF MF Cost Dir Dtr
(a) Welfare-CS (b) Expected EV No.-CS
20 60
Expected Welfare
Expected EV No.
CS1 CS2 CS3 CS1 CS2 CS3
15
40
10
20
5
0 0
Opt EMG MG EMF MF Cost Dir Dtr Opt EMG MG EMF MF Cost Dir Dtr
(c) Welfare-CS’ (d) Expected EV No.-CS’
Figure 5.2: Welfare and demand balancing comparison for station setting CS =
{CS1 , CS2 , CS3 } and CS 0 = {CS10 , CS2 , CS30 } among the eight methods. Their
locations are depicted in Fig. 5.1. EMG: EM-Greedy; MG: M-Greedy; EMF: EM-
FPTAS; MF: M-FPTAS; Cost: Min-Cost; Dir: Min-Dir; Dtr: Min-Dtr.
map; while CS’ replaces station CS1 and CS3 with station CS10 and
CS30 , which are near the border of the map. First, EM-Greedy and EM-
FPTAS significantly outperform Min-Cost, Min-Dir or Min-Dtr as to
welfare maximization. For example, for station setting CS, EM-FPTAS
achieves 17.7%, 13.0% and 14.5% more welfare than the three meth-
ods respectively; for station setting CS’, the difference becomes even
larger, which is 82.3%, 34.1% and 80.9% respectively. Second, by the
theoretical bound proved in Theorem 5.1, EM-Greedy and EM-FPTAS
(with ε = 0.1) are far from Opt. However, in practice, both of the two
methods perform rather close to Opt. For example, EM-Greedy and
EM-FPTAS performs 7.7% and 7.0% worse than Opt for setting CS,
and both are 2.3% worse for setting CS’. The reason is that the theoret-
ical bound is analyzed based on the worst case, which does not usually
happen in practice.
Third, although EM-Greedy and M-Greedy (EM-FPTAS and M-

FPTAS) have the same approximation ratio, the former has better
average performance in practice. As shown in Fig. 5.2(a), EM-Greedy
(EM-FPTAS) achieves 4.5% (3.7%) more welfare than M-Greedy (EM-
FPTAS). Fourth, an interesting observation is that Min-Cost has the
worst performance among all three self-decision methods, even though
by this method, each EV user chooses to charge at stations with the
minimum charging cost. This is because the unbalanced charging de-
mand dominates EV users’ initial intent. To be specific, there are few
EVs scheduled to charging station CS1 (CS10 ) by Min-Cost. Most EVs
swam into station CS2 (CS20 ) and make it overloaded. Much fewer
EVs would be serviced if using this method. Fifth, we further ob-
serve that the welfare by station setting CS is greater than that of
CS’. For instance, bar Opt in Fig. 5.2(a) is up to 13.4% better than
that in Fig. 5.2(c). This observation also indicates the importance of
well-planned charging station deployment about welfare maximization
and user satisfaction.
Demand balancing comparison. Fig. 5.2(b) and 5.2(d) plot the
comparison on the expected number of EVs scheduled to stations. First,
EM-Greedy and EM-FPTAS balance EVs almost as well as Opt when
stations are either evenly distributed as setting CS or ill-deployed as
setting CS’. Second, a counter-intuitive result is that self-decision mak-
ing methods such as Min-Dir and Min-Dtr also come with rather good
load balance with respect to setting CS. The reason is two-fold. The
first reason is that the three stations are evenly deployed on the map;
the other one is that the EVs’ current positions and their trips are also
uniformly distributed on the map by SUMO. However, if the stations
are ill-deployed such as setting CS’, these self-decision making methods
cause very unbalanced demand scheduling. For example, as shown in
Fig. 5.2(d)5.2(b), the expected number of EVs assigned to station CS10
and CS20 significantly decreases compared to that of station CS1 and
CS2 . Third, station CS2 are fully occupied over all methods and all
settings. This is because this station has the cheapest energy price and
also in the middle of trips of most EVs.
Sensitivity Analysis
We vary one parameter and keep other parameters unchanged (using

initial values) when analyzing the impact of the parameter on the wel-
fare and demand balancing.
Impact of inaccuracy ε. Fig. 5.3(a) shows the trade-off between
performance and running time for M-FPTAS and EM-FPTAS with
different inaccuracies, where ε varies from 0.5 to 0.01. First, M-FPTAS
(EM-FPTAS) becomes more and more close to Opt as inaccuracy ε
decreases, that is, the performance ratio of M −FOpt Opt
P T AS ( EM −F P T AS )
varies from 1.15 to 1.11 (from 1.10 to 1.07) according to the yellow bars
(green bars) and left y-axis. Moreover, the ratio keeps nearly unchanged
when inaccuracy ε becomes less than 0.1. By contrast, the running time
(the blue and red curves, right y-axis) grows in a much larger scale,
that is, ε = 0.01 runs more than 40 times longer than ε = 0.5. The
running time is normalized according to that for M-FPTAS when ε =
0.5 (whose running time is the minimum among all). Hence, we suggest
that ε = 0.1 is a good balance between performance and running time
in practice. Second, EM-FPTAS has nearly the same running time as
M-FPTAS. This is because that the number of EVs is usually much
larger than that of charging stations. The extra loop of EM-FPTAS is
by the station number and thus increases running time only by little,
while it achieves better welfare.
Impact of energy price. Fig. 5.3(b) depicts the welfare
comparison by varying energy price. We consider the six different
permutations of the three different energy prices, as shown in Table 5.1.
Each method’s bars are plotted in an order of s1, s2, ..., s6. First, we
observe that the lower price station CS2 has, the larger the total
welfare is. For example, s1 and s6 are the two tallest bars, for the
price of station CS2 is the lowest among the three stations; s4 and
s5 are the two shortest bars, for the price of station CS2 becomes
the highest in this case. Most of the EV trips generated by SUMO
passes the middle of the map, and thus more EVs have shorter detour
distance to the middle station (CS2 ) than to either the left (CS1 ) or
right (CS3 ) station. EV users can benefit more if charging at station
CS2 than charging at the other two stations. However, this benefit
1.3 45 20
Expected Welfare
Opt/EM-FPTAS EM-FPTAS CS1 CS2 CS3
Running Time
Opt/M-FPTAS M-FPTAS
1.2 30 Opt EM-Greedy EM-FPTAS
Ratio
10
1.1 15
1 0 0
0.5 0.2 0.1 0.05 0.02 0.01 s1 s6 s1 s6 s1 s6
Inaccuracy H Price permutation
(a) Welfare vs running time (b) Welfare-varying price

Expected Welfare
30
CS1 CS2 CS3
Opt EM-FPTAS
20 EM-Greedy
10
0
10 20 30 40 50 10 20 30 40 50 10 20 30 40 50
Station Capacity
(c) Welfare-varying station capacity
120
Expected EV No.
CS1 CS2 CS3

Opt EM-FPTAS
80
EM-Greedy
40
0
10 20 30 40 50 10 20 30 40 50 10 20 30 40 50
Station Capacity
(d) Expected number of EVs-varying station capacity
Figure 5.3: Sensitivity analysis of inaccuracy (a), energy price (b), and station
capacity (c)(d).
Table 5.1: Six permutations of energy price.
s1:{0.035,0.027,0.032} | s2:{0.035,0.032,0.027} | s3:{0.027,0.032,0.035}

s4:{0.027,0.035,0.032} | s5:{0.032,0.035,0.027} | s6:{0.032,0.027,0.035}
5.5. Summary 115
will be offset if CS2 has larger energy price. This is the reason why
we have the above observation. Second, the lower price a charging
station has, the higher welfare it can give. For example, station CS1
(CS3 ) has the largest welfare for bar s3 and s4 (s2 and s5), where the
station has the lowest price. This observation implies that varying
energy price according to the traffic density can help balance charging
demand across charging stations.
Impact of station capacity. Fig. 5.3(c) and 5.3(d) plot the com-
parison about welfare and demand balancing for different station ca-
pacities. Stations’ capacity is set to be identical and varies from 5 to
50 (in a step of 5) simultaneously. First, the welfare by all the three
methods increases as the station capacity grows. The difference among
them is that Opt’s welfare stays with little change after capacity turns
to 20 ∼ 25. A similar observation can be also made from Fig. 5.3(d),
where the expected number of EVs nearly keeps unchanged. From the
above observations, we can make a rough estimation on the ratio be-
tween EV number and charging point number for a general case. For
example, if charging load is uniformly distributed during 8am to 12pm
and other hours has no charging load, the estimated ratio of EV num-
ber to charging point number is 64 : 1 ∼ 80 : 1, which is calculated
8 (load duration)
as: 100 (current setting) × 0.5 (charging time) : 20 ∼ 25. Second, the
expected EV number of station CS1 and CS2 first increases and then
decreases as the station capacity grows. The increase happens when the
total expected number of EVs is larger than the total station capacity.
At this time, assigning more EVs results in larger welfare. The decrease
occurs when the total station capacity becomes excessive. More EVs are
assigned to station CS2 because of its lower energy price and the mid-
dle position on the map. This also explains why the EV number of
station CS2 always rises with its capacity growth.
5.5 Summary
This chapter addresses the competition on public charging points

among electric vehicles. We first propose charging demand balancing
algorithms with provable theoretical performance guarantees. These al-

gorithms take account of user behavior uncertainty and the scheduling
decisions are seen as recommendations that aid EV users to perform
decision making. We then conduct extensive simulations using realistic
data traces and simulations tools. The result validates our theoreti-
cal results and further highlights the benefits and importance of our
algorithms. We finally present recent developments about charging de-
mand balancing, which analyze relationships and differences between
our work and existing ones.
6
Conclusions
In this article, we investigate smart electric vehicle charging. Our focus

is to jointly address quality of service for electric vehicle users and the
stability and reliability of the power grid. Following the proposed solu-
tion framework, the article proposes solutions for several key problems
arising from both the lower and upper levels. The proposed solutions
are developed mainly using techniques from the optimization, game
theory, algorithmic, and scheduling fields.
In Chapter 1, we first introduce the ecosystem of electric vehicles
and discuss motivations for managing charging load. EVs keep prolifer-
ating due to their low oil dependency and environmental friendliness.
The popularity of EVs causes heavy load impact onto the power grid
and urges the deployment of various charging facilities such as pub-
lic available charging stations. In order to jointly address these two
consequences, we then propose a hierarchical solution framework that
exploits the elasticity of charging load and existing communication net-
works to coordinate and control EV charging. The framework consists
of two different levels: the lower level solves three crucial problems
about how to charge vehicles within a charging station, while the up-
per level solves the problem of how to balance charging demand among
117
118 Conclusions
multiple charging stations. The following chapters thoroughly study

each of these four problems.
In Chapter 2, we investigate charging rate control, which handles
how to allocate power supply to electric vehicles served within a charg-
ing station. We design an efficient charging rate control algorithm that
results in provably optimal power allocation among EVs. The algorithm
not only maximizes quality of service for EV users but also is able to
guarantee the reliability of the power grid. We then carry out extensive
simulations using realistic data traces. The results and the correspond-
ing analysis demonstrate the benefits and importance of the proposed
algorithm. Finally, we review recent developments about charging rate
control, which have different focuses or different optimization goals such
as power load flattening, charging cost minimization, etc..
In Chapter 3, we address electric vehicle demand response, which
is how to make electric vehicles follow the power supply of charging
stations and the power grid. The focus is hierarchical demand response
for EVs via charging stations. We propose a distributed deadline-aware
two-level market mechanism, which is compatible with both economic
and temporal aspects of EV demand response. The analysis for the
proposed mechanism employs Stackelberg game and demonstrates it
convergence to a unique equilibrium solution. We further provide a
stochastic optimization based algorithm to incorporate renewable en-
ergy into the market mechanism. It shows that the impact of renew-
able’s prediction error on the mechanism is bounded. The trace-driven
simulation results demonstrate the efficacy and also validate the theo-
retical analysis for the designed market mechanism. Finally, we review
recent developments about EV demand response, which analyzes rela-
tionships and differences between our work and existing ones.
In Chapter 4, we study electric vehicle scheduling, which discusses
how to schedule electric vehicles to multiple charging points within a
charging station. We adopt the metered model to characterize charging
tasks and leverage event-driven methods to schedule these tasks. Both
theoretical and experimental analysis are performed for the adaption
of event-driven algorithms to the metered model. Theoretical results
show the varying performance bound of temporality-based algorithms
119
including EDF, SJF, and FCFS, and the constant performance bound
of the value-based algorithm, HUF. Simulation results further demon-
strate the near-optimality and performance consistency of HUF, and
the sub-optimality and performance anomaly of the temporality-based
algorithms, in terms of optimizing user satisfaction. These results in-
dicate adopting HUF in park-and-charge systems can provide better
charging service to customers and thus make them better satisfied.
In Chapter 5, we study charging demand balancing, which deals
with how to balance electric vehicles among multiple charging sta-
tions. We propose approximate charging demand balancing algorithms,
which take account of user behavior uncertainty. The scheduling deci-
sions made by the algorithms are seen as recommendations that aid
EV users to perform decision making. We then carry out extensive
simulations based on realistic data traces and a simulations tool. The
result validates our theoretical analysis and highlights the benefits of
our algorithms. We finally review recent developments about charging
demand balancing and EV routing.
Due to time and space limitations, we could not discuss all the re-
search and discoveries related to smart electric vehicle charging in this
article. We apologize for any possible negligence and omissions. Inter-
ested readers can find the most recent works in tutorial and survey
papers such as [2, 6], as well as in the related conferences and journals
including International Conference on Future Energy Systems (ACM e-
Energy), ACM/IEEE International Conference on Cyber-Physical Sys-
tems (ICCPS), International Green and Sustainable Computing Con-
ference, IEEE International Conference on Smart Grid Communica-
tions (SmartGridComm), IEEE Power & Energy Society (PES) Gen-
eral Meeting, IEEE PES Innovative Smart Grid Technologies (ISGT),
and various ACM and IEEE Transactions.
Acknowledgements
We would like to thank our editor, Prof. Marija D. Ilic, and the anony-
mous reviewers of this article for their constructive criticism of the
earlier manuscripts. This work was supported in part by the NSERC
Discovery Grant 341823, NSERC Collaborative Research and Develop-
ment Grant CRDPJ418713, Mitacs Fund IT06703, and McGill Tom-
linson Scientist Award.
120
References
[1] International Energy Agency. Co2 emissions from fuel combustion high-
lights 2016. http://www.iea.org/. [Online accessed Feb., 2017].
[2] Liansheng Liu, Fanxin Kong, Xue Liu, Yu Peng, and Qinglong Wang.
A review on electric vehicles interacting with renewable energy in smart
grid. Renewable and Sustainable Energy Reviews, 2015.
[3] Kristien Clement-Nyns, Edwin Haesen, and Johan Driesen. The impact
of charging plug-in hybrid electric vehicles on a residential distribution
grid. IEEE Transactions on Power Systems, 2010.
[4] João A Peças Lopes, Filipe Joel Soares, and Pedro M Rocha Almeida.
Integration of electric vehicles in the electric power system. Proceedings
of the IEEE, 2011.
[5] Fanxin Kong, Xue Liu, Zhonghao Sun, and Qinglong Wang. Smart rate
control and demand balancing for electric vehicle charging. In ICCPS.
IEEE, 2016.
[6] Qinglong Wang, Xue Liu, Jian Du, and Fanxin Kong. Smart charging
for electric vehicles: A survey from the algorithmic perspective. IEEE
Communications Surveys & Tutorials, 2016.
[7] Jing Huang, Vijay Gupta, and Yih-Fang Huang. Scheduling algorithms
for phev charging in shared parking lots. In ACC. IEEE, 2012.
[8] ChargePoint. We are chargepoint. http://www.chargepoint.com. [On-
line accessed June, 2015].
[9] Omid Ardakanian, Catherine Rosenberg, and Srinivasan Keshav. Dis-
tributed control of electric vehicle charging. In e-Energy. ACM, 2013.
121
122 References
[10] Shiyao Chen, Yuting Ji, and Lang Tong. Large scale charging of electric
vehicles. In PES General Meeting. IEEE, 2012.
[11] Wikipedia. Park & Charge. http://en.wikipedia.org/wiki/Park_
%26_Charge. [Online accessed Apr., 2015].
[12] Julian Timpner and Lars Wolf. Design and evaluation of charging sta-
tion scheduling strategies for electric vehicles. IEEE Transactions on
Intelligent Transportation Systems, 2014.
[13] Zhipeng Liu, Fushuan Wen, and Gerard Ledwich. Optimal planning of
electric-vehicle charging stations in distribution systems. IEEE Trans-
actions on Power Delivery, 2013.
[14] Florian Hausler, Emanuele Crisostomi, Arieh Schlote, Ilja Radusch, and
Robert Shorten. Stochastic park-and-charge balancing for fully electric
and plug-in hybrid vehicles. IEEE Transactions on Intelligent Trans-
portation Systems, 2014.
[15] Robert C Green, Lingfeng Wang, and Mansoor Alam. The impact of
plug-in hybrid electric vehicles on distribution networks: A review and
outlook. Renewable and sustainable energy reviews, 2011.
[16] Luis Pieltain Fernandez, Tomás Gomez San Roman, Rafael Cossent,
Carlos Mateo Domingo, and Pablo Frias. Assessment of the impact of
plug-in electric vehicles on distribution networks. IEEE Transactions
on Power Systems, 2011.
[17] Ulrich Eberle and Rittmar Von Helmolt. Sustainable transportation
based on electric vehicle concepts: a brief overview. Energy & Environ-
mental Science, 2010.
[18] Qiuming Gong, Shawn Midlam-Mohler, Vincenzo Marano, and Giorgio
Rizzoni. Study of pev charging on residential distribution transformer
life. IEEE Transactions on Smart Grid, 3(1):404–412, 2012.
[19] Qiao Xiang, Fanxin Kong, Xi Chen, Linghe Kong, Xue Liu, and Lei Rao.
Auc2charge: An online auction framework for electric vehicle park-and-
charge. In e-Energy. ACM, 2015.
[20] Fanxin Kong, Qiao Xiang, Linghe Kong, and Xue Liu. On-line event-
driven scheduling for electric vehicle charging via park-and-charge. In
RTSS. IEEE, 2016.
[21] International Energy Agency. Global ev outlook 2016: Beyond one mil-
lion electric cars. https://www.iea.org/. [Online accessed May, 2017].
[22] U.S. Department of Energy. Electric vehicle charging station locations.
http://www.afdc.energy.gov/. [Online accessed May, 2017].
References 123
[23] SAE EV Charging Systems Committee . Sae surface vehicle recom-

mended practice j1772, sae electric vehicle conductive charge coupler.
http://www.arb.ca.gov/. [Online accessed May, 2017].
[24] Jason Taylor, Arindam Maitra, Mark Alexander, Daniel Brooks, and
Mark Duvall. Evaluation of the impact of plug-in electric vehicle loading
on distribution system operations. In IEEE Power & Energy Society
General Meeting (PES). IEEE, 2009.
[25] Lingwen Gan, Ufuk Topcu, and S Low. Optimal decentralized protocol
for electric vehicle charging. IEEE Transactions on Power Systems,
2013.
[26] Lingwen Gan, Ufuk Topcu, and Steven H Low. Stochastic distributed
protocol for electric vehicle charging with discrete charging rate. In PES
General Meeting. IEEE, 2012.
[27] Lingwen Gan, Adam Wierman, Ufuk Topcu, Niangjun Chen, and
Steven H Low. Real-time deferrable load control: handling the uncer-
tainties of renewable generation. In e-Energy. ACM, 2013.
[28] Zhongjing Ma, Duncan Callaway, and Ian Hiskens. Decentralized charg-
ing control for large populations of plug-in electric vehicles. In CDC.
IEEE, 2010.
[29] Eric Sortomme, Mohammad M Hindi, S.D. James MacPherson, and
S.S. Venkata. Coordinated charging of plug-in hybrid electric vehicles
to minimize distribution system losses. IEEE transactions on smart
grid, 2011.
[30] Sara Deilami, Amir S Masoum, Paul S Moses, and Mohammad AS
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, 2011.
[31] Sekyung Han, Soohee Han, and Kaoru Sezaki. Estimation of achievable
power capacity from plug-in electric vehicles for v2g frequency regu-
lation: Case studies for market participation. IEEE Transactions on
Smart Grid, 2011.
[32] Sekyung Han, Soohee Han, and Kaoru Sezaki. Development of an op-
timal vehicle-to-grid aggregator for frequency regulation. IEEE Trans-
actions on smart grid, 2010.
[33] Willett Kempton and Jasna Tomić. Vehicle-to-grid power implemen-
tation: From stabilizing the grid to supporting large-scale renewable
energy. Journal of power sources, 2005.
124 References
[34] Chenye Wu, Hamed Mohsenian-Rad, Jianwei Huang, and Juri Jatske-
vich. Pev-based combined frequency and voltage regulation for smart
grid. In IEEE PES Innovative Smart Grid Technologies (ISGT). IEEE,
2012.
[35] Hongming Yang, CY Chung, and Junhua Zhao. Application of plug-
in electric vehicles to frequency regulation based on distributed signal
acquisition via limited communication. IEEE Transactions on Power
Systems, 2013.
[36] Yifeng He, Bala Venkatesh, and Ling Guan. Optimal scheduling for
charging and discharging of electric vehicles. IEEE Transactions on
Smart Grid, 2012.
[37] Casey Quinn, Daniel Zimmerle, and Thomas H Bradley. An evaluation
of state-of-charge limitations and actuation signal energy content on
plug-in hybrid electric vehicle, vehicle-to-grid reliability, and economics.
[38] MF Shaaban, Muhammad Ismail, EF El-Saadany, and Weihua Zhuang.
Real-time pev charging/discharging coordination in smart distribution
systems. IEEE Transactions on Smart Grid, 2014.
[39] Zizhan Zheng and Ness B Shroff. Online welfare maximization for elec-
tric vehicle charging with electricity cost. In e-Energy. ACM, 2014.
[40] Niklas Rotering and Marija Ilic. Optimal charge control of plug-in hy-
brid electric vehicles in deregulated electricity markets. IEEE Transac-
tions on Power Systems, 2011.
[41] Jonathan Donadee and Marija D Ilie. Stochastic optimization of grid
to vehicle frequency regulation capacity bids. IEEE Transactions on
Smart Grid, 2014.
[42] Tarek Abdelzaher and etc. Green gps : A participatory sensing fuel-
efficient maps application. http://green-way.cs.illinois.edu/.
[Online accessed June, 2015].
[43] Moritz Baum, Julian Dibbelt, Thomas Pajor, and Dorothea Wagner.
Energy-optimal routes for electric vehicles. In SIGSPATIAL. ACM,
2013.
[44] Martin Sachenbacher, Martin Leucker, Andreas Artmeier, and Julian
Haselmayr. Efficient energy-optimal routing for electric vehicles. In
AAAI, 2011.
[45] Albert G Boulanger, Andrew C Chu, Suzanne Maxx, and David L
Waltz. Vehicle electrification: status and issues. Proceedings of the
IEEE, 2011.
References 125
[46] Navid Rahbari-Asr and Mo-Yuen Chow. Cooperative distributed de-

mand management for community charging of phev/pevs based on kkt
conditions and consensus networks. IEEE Transactions on Industrial
Informatics, 2014.
[47] Kevin Mets, Tom Verschueren, Wouter Haerick, Chris Develder, and
Filip De Turck. Optimizing smart energy control strategies for plug-in
hybrid electric vehicle charging. In IEEE/IFIP Network Operations and
Management Symposium Workshops (NOMS Wksps). IEEE, 2010.
[48] Amir S Masoum, Sara Deilami, Ahmed Abu-Siada, and Mohammad AS
Masoum. Fuzzy approach for online coordination of plug-in electric ve-
hicle charging in smart grid. IEEE Transactions on Sustainable Energy,
2015.
[49] Chenye Wu, Hamed Mohsenian-Rad, and Jianwei Huang. Vehicle-to-
aggregator interaction game. IEEE Transactions on Smart Grid, 2012.
[50] George R Parsons, Michael K Hidrue, Willett Kempton, and Meryl P
Gardner. Willingness to pay for vehicle-to-grid (v2g) electric vehicles
and their contract terms. Energy Economics, 2014.
[51] Emil B Iversen, Juan M Morales, and Henrik Madsen. Optimal charging
of an electric vehicle using a markov decision process. Applied Energy,
2014.
[52] Omid Ardakanian, Srinivasan Keshav, and Catherine Rosenberg. Real-
time distributed control for smart electric vehicle chargers: From a static
to a dynamic study. IEEE Transactions on Smart Grid, 2014.
[53] Lijun Chen, Na Li, SH Low, and JC Doyle. Two market models for
demand response in power networks. In SmartGridComm. IEEE, 2010.
[54] Lazaros Gkatzikis, Iordanis Koutsopoulos, and Theodoros Salonidis.
The role of aggregators in smart grid demand response markets. IEEE
Journal on Selected Areas in Communications, 2013.
[55] Zhenhua Liu, Iris Liu, Steven Low, and Adam Wierman. Pricing data
center demand response. In SIGMETRICS. ACM, 2014.
[56] TimberRock Energy Solutions. http://timberrockes.com/. [Online
accessed Feb., 2015].
[57] UGE (Urban Green Energy). http://www.urbangreenenergy.com/.
[Online accessed Feb., 2015].
[58] Wikipedia. Charging station. https://en.wikipedia.org/wiki/
Charging_station. [Online accessed Feb., 2015].
126 References
[59] Fanxin Kong and Xue Liu. Distributed deadline and renewable aware
electric vehicle demand response in the smart grid. In RTSS. IEEE,
2015.
[60] Jian Yao, Shmuel S Oren, and Ilan Adler. Two-settlement electricity
markets with price caps and cournot generation firms. European Journal
of Operational Research, 2007.
[61] Gregor Giebel, Richard Brownsword, George Kariniotakis, Michael Den-
hard, and Caroline Draxl. The state-of-the-art in short-term prediction
of wind power: A literature overview. Technical report, ANEMOS. plus,
2011.
[62] Dong C. Park, M.A. El-Sharkawi, R.J. Marks, L.E. Atlas, and M.J.
Damborg. Electric load forecasting using an artificial neural network.
IEEE Transactions on Power Systems, 1991.
[63] Yurii Nesterov, Arkadii Nemirovskii, and Yinyu Ye. Interior-point poly-
nomial algorithms in convex programming. SIAM, 1994.
[64] National Renewable Energy Laboratory. National solar radiation data
base - 1991- 2010 update. http://rredc.nrel.gov. [Online accessed
Fep., 2015].
[65] Southern California Edison. Regulatory information - sce load profiles.
http://www.sce.com/. [Online accessed Fep., 2015].
[66] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cam-
bridge university press, 2009.
[67] Eilyan Bitar, Pramod P Khargonekar, and Kameshwar Poolla. Systems
and control opportunities in the integration of renewable energy into
the smart grid. In IFAC World Congress, 2011.
[68] U.S. Department of Energy. The smart grid: An introduction. http:
//www.energy.gov/. [Online accessed March, 2015].
[69] Na Li, Lijun Chen, and Steven H Low. Optimal demand response based
on utility maximization in power networks. In Power and Energy Society
General Meeting. IEEE, 2011.
[70] Shengnan Shao, Manisa Pipattanasomporn, and Saifur Rahman. Grid
integration of electric vehicles and demand response with customer
choice. IEEE Transactions on Smart Grid, 2012.
[71] Zhao Tan, Peng Yang, and Arye Nehorai. An optimal and distributed
demand response strategy with electric vehicles in the smart grid. IEEE
Transactions on Smart Grid, 2014.
References 127
[72] JS Vardakas, N Zorba, and CV Verikoukis. A survey on demand re-

sponse programs in smart grids: Pricing methods and optimization al-
gorithms. IEEE Communications Surveys & Tutorials, 2015.
[73] Mohamed H Albadi and EF El-Saadany. A summary of demand re-
sponse in electricity markets. Electric Power Systems Research, 2008.
[74] Pierluigi Siano. Demand response and smart gridsâĂŤa survey. Renew-
able and Sustainable Energy Reviews, 2014.
[75] U.S. Department of Energy. Evaluating electric vehicle charging impacts
and customer charging behaviors. http://www.energy.gov/. [Online
accessed March, 2015].
[76] Fanxin Kong, Chuansheng Dong, Xue Liu, and Haibo Zeng. Blowing
hard is not all we want: Quantity vs quality of wind power in the smart
grid. In INFOCOM. IEEE, 2014.
[77] Fanxin Kong, Chuansheng Dong, Xue Liu, and Haibo Zeng. Quantity
versus quality: Optimal harvesting wind power for the smart grid. Pro-
ceedings of the IEEE, 2014.
[78] Tamer Basar and Geert Jan Olsder. Dynamic noncooperative game
theory. SIAM, 1995.
[79] Zhonghao Sun, Fanxin Kong, Xue Liu, Xingshe Zhou, and Xi Chen.
Intelligent joint spatio-temporal management of electric vehicle charging
and data center power consumption. In International Green Computing
Conference. IEEE, 2014.
[80] A Subramanian, M Garcia, A Dominguez-Garcia, D Callaway, K Poolla,
and P Varaiya. Real-time scheduling of deferrable electric loads. In
ACC. IEEE, 2012.
[81] Ricardo J Bessa, Manuel A Matos, Filipe Joel Soares, and João A Peças
Lopes. Optimized bidding of a ev aggregation agent in the electricity
market. IEEE Transactions on Smart Grid, 2012.
[82] Habiballah Rahimi-Eichi and Mo-Yuen Chow. Auction-based energy
management system of a large-scale phev municipal parking deck. In
Energy Conversion Congress and Exposition. IEEE, 2012.
[83] Matteo Vasirani, Ramachandra Kota, Renato LG Cavalcante, Sascha
Ossowski, and Nicholas R Jennings. An agent-based approach to vir-
tual power plants of wind power generators and electric vehicles. IEEE
Transactions on Smart Grid, 2013.
[84] Eilyan Bitar and Yunjian Xu. On incentive compatibility of deadline
differentiated pricing for deferrable demand. In Annual Conference on
Decision and Control. IEEE, 2013.
128 References
[85] Li Zhang, Faryar Jabbari, Tim Brown, and Scott Samuelsen. Coordi-
nating plug-in electric vehicle charging with electric grid: Valley filling
and target load following. Journal of Power Sources, 2014.
[86] Anna Nagurney. Network economics: A variational inequality approach.
Springer Science & Business Media, 1998.
[87] Xue Liu, Qixin Wang, Wenbo He, Marco Caccamo, and Lui Sha. Op-
timal real-time sampling rate assignment for wireless sensor networks.
ACM Transactions on Sensor Networks, 2006.
[88] Weihuan Shu, Xue Liu, Zonghua Gu, and Sathish Gopalakrishnan. Op-
timal sampling rate assignment with dynamic route selection for real-
time wireless sensor networks. In Real-Time Systems Symposium. IEEE,
2008.
[89] Daniel O’Neill, Marco Levorato, Andrea Goldsmith, and Urbashi Mitra.
Residential demand response using reinforcement learning. In Interna-
tional Conference on Smart Grid Communications. IEEE, 2010.
[90] Steven H Low and David E Lapsley. Optimization flow controlâĂŤi:
basic algorithm and convergence. IEEE/ACM Transactions on Net-
working, 1999.
[91] Srinivas Shakkottai and Rayadurgam Srikant. Network optimization
and control. Now Publishers Inc, 2008.
[92] Drew Fudenberg and Jean Tirole. Game theory. MIT Press, 1991.
[93] Roy D. Yates. A framework for uplink power control in cellular radio
systems. IEEE Journal on Selected Areas in Communications, 1995.
[94] Noam Nisan, Tim Roughgarden, Eva Tardos, and Vijay V Vazirani.
Algorithmic game theory. Cambridge University Press, 2007.
[95] Zhenhua Liu, Adam Wierman, Yuan Chen, Benjamin Razon, and Ni-
angjun Chen. Data center demand response: Avoiding the coincident
peak via workload shifting and local generation. Performance Evalua-
tion, 2013.
[96] Nissan Leaf. Charging & range. http://www.nissanusa.com/. [Online
accessed Fep., 2015].
[97] Tesla Motors. Model S. http://www.teslamotors.com/models. [On-
line accessed March, 2015].
[98] U.S. Department of Energy. Developing infrastructure to charge plug-
in electric vehicles. http://www.afdc.energy.gov/. [Online accessed
March, 2015].
References 129
[99] Roberto Cominetti, José R Correa, and Nicolás E Stier-Moses. Network

games with atomic players. In Automata, Languages and Programming.
Springer, 2006.
[100] V-charge: Automated valet parkging and charging for e-mobility col-
laborative project no. fp7-269916. http://www.v-charge.eu/. [Online
accessed Apr., 2015].
[101] Julian de Hoog, Tansu Alpcan, Marcus Brazil, Doreen Anne Thomas,
and Iven Mareels. Optimal charging of electric vehicles taking distri-
bution network constraints into account. IEEE Transactions on Power
Systems, 2014.
[102] Olle Sundstrom and Carl Binding. Flexible charging optimization for
electric vehicles considering distribution grid constraints. IEEE Trans-
actions on Smart Grid, 2012.
[103] Shizhen Zhao, Xiaojun Lin, and Minghua Chen. Peak-minimizing online
ev charging. In Allerton, 2013.
[104] Lei Zhang and Yaoyu Li. Optimal management for parking-lot elec-
tric vehicle charging by two-stage approximate dynamic programming.
[105] Miadreza Shafie-khah, Ehsan Heydarian-Forushani, Gerardo J Osório,
Fábio AS Gil, Jamshid Aghaei, Mostafa Barani, and João PS Catalão.
Optimal behavior of electric vehicle parking lots as demand response
aggregation agents. IEEE Transactions on Smart Grid, 2016.
[106] Nilufar Neyestani, Maziar Yazdani Damavandi, Miadreza Shafie-Khah,
Anastasios G Bakirtzis, and João PS Catalão. Plug-in electric vehicles
parking lot equilibria with energy and reserve markets. IEEE Transac-
tions on Power Systems, 2017.
[107] Bogdan Doytchinov, John Lehoczky, and Steven Shreve. Real-time
queues in heavy traffic with earliest-deadline-first queue discipline. An-
nals of Applied Probability, 2001.
[108] Vivek Sharma, Arun Thomas, Tarek Abdelzaher, Kevin Skadron, and
Zhijian Lu. Power-aware qos management in web servers. In RTSS.
IEEE, 2003.
[109] Francis YL Chin and Stanley PY Fung. Improved competitive algo-
rithms for online scheduling with partial job values. Theoretical com-
puter science, 2004.
[110] Hua Qin and Wensheng Zhang. Charging scheduling with minimal wait-
ing in a network of electric vehicles and charging stations. In VANET.
ACM, 2011.
130 References
[111] Julian Timpner and Lars Wolf. Efficient charging station scheduling for
an autonomous parking and charging system. In VANET. ACM, 2012.
[112] Marek Chrobak, Leah Epstein, John Noga, Jiří Sgall, Rob van Stee,
Tomáš Tichỳ, and Nodari Vakhania. Preemptive scheduling in over-
loaded systems. In International Colloquium on Automata, Languages,
and Programming. Springer, 2002.
[113] Francis YL Chin and Stanley PY Fung. Online scheduling with partial
job values: Does timesharing or randomization help? Algorithmica, 2003.
[114] Mohammad Charkhgard and Mohammad Farrokhi. State-of-charge es-
timation for lithium-ion batteries using neural networks and ekf. IEEE
Transactions on Industrial Electronics, 2010.
[115] Yinjiao Xing, Wei He, Michael Pecht, and Kwok Leung Tsui. State of
charge estimation of lithium-ion batteries using the open-circuit voltage
at various ambient temperatures. Applied Energy, 2014.
[116] Enrico H Gerding, Valentin Robu, Sebastian Stein, David C Parkes,
Alex Rogers, and Nicholas R Jennings. Online mechanism design for
electric vehicle charging. In AAMAS. IFAAMAS, 2011.
[117] Anna Nagurney. Network economics: A variational inequality approach.
Springer Science & Business Media, 2013.
[118] Kwon Oh-Heum and Chwa Kyung-Yong. Scheduling parallel tasks with
individual deadlines. Theoretical Computer Science, 1999.
[119] Martin Stigge and Wang Yi. Graph-based models for real-time work-
load: A survey. Real-Time Systems Journal, 2015.
[120] Jane WS Liu, Wei-Kuan Shih, Kwei-Jay Lin, Riccardo Bettati, and J-Y
Chung. Imprecise computations. Proceedings of the IEEE, 1994.
[121] Bu-808: How to prolong lithium-based batteries. http://
batteryuniversity.com/. [Online accessed Apr., 2016].
[122] Elisabeth Günther, Olaf Maurer, Nicole Megow, and Andreas Wiese. A
new approach to online scheduling: Approximating the optimal compet-
itive ratio. In SODA. SIAM, 2013.
[123] Marek Chrobak, Leah Epstein, John Noga, JiřıÌĄ Sgall, Rob van Stee,
Tomáš Tichỳ, and Nodari Vakhania. Preemptive scheduling in over-
loaded systems. Journal of Computer and System Sciences, 2003.
[124] Hans Kellerer, Ulrich Pferschy, and David Pisinger. Knapsack problems.
Springer, 2004.
[125] Bala Kalyanasundaram and Kirk Pruhs. Speed is as powerful as clair-
voyance. Journal of the ACM, 2000.
References 131
[126] Cynthia A Phillips, Cliff Stein, Eric Torng, and Joel Wein. Optimal
time-critical scheduling via resource augmentation. In STOC. ACM,
1997.
[127] IBM. CPLEX optimizer. http://www.ibm.com/. [Online accessed Fep.,
2015].
[128] Jonathan D Adler, Pitu B Mirchandani, Guoliang Xue, and Minjun
Xia. The electric vehicle shortest-walk problem with battery exchanges.
Networks and Spatial Economics, 2016.
[129] Daehyun Ban, George Michailidis, and Michael Devetsikiotis. Demand
response control for phev charging stations by dynamic price adjust-
ments. In ISGT. IEEE, 2012.
[130] Myriam Neaimeh, Graeme A Hill, Yvonne Hübner, and Phil T Blythe.
Routing systems to extend the driving range of electric vehicles. IET
Intelligent Transport Systems, 2013.
[131] Jochen Eisner, Stefan Funke, and Sabine Storandt. Optimal route plan-
ning for electric vehicles in large networks. In AAAI, 2011.
[132] Hao Liang, Isha Sharma, Weihua Zhuang, and Kankar Bhattacharya.
Plug-in electric vehicle charging demand estimation based on queueing
network analysis. In PES General Meeting. IEEE, 2014.
[133] Valentin Robu, Enrico H Gerding, Sebastian Stein, David C Parkes,
Alex Rogers, and Nicholas R Jennings. An online mechanism for multi-
unit demand and its application to plug-in hybrid electric vehicle charg-
ing. Journal of Artificial Intelligence Research, 2013.
[134] Valentin Robu, Sebastian Stein, Enrico H Gerding, David C Parkes,
Alex Rogers, and Nicholas R Jennings. An online mechanism for multi-
speed electric vehicle charging. In AMMA. Springer, 2012.
[135] Yijia Cao, Shengwei Tang, Canbing Li, Peng Zhang, Yi Tan, Zhikun
Zhang, and Junxiong Li. An optimized ev charging model considering
tou price and soc curve. IEEE Transactions on Smart Grid, 2012.
[136] Reuven Cohen, Liran Katzir, and Danny Raz. An efficient approxima-
tion for the generalized assignment problem. Information Processing
Letters, 2006.
[137] Lisa Fleischer, Michel X Goemans, Vahab S Mirrokni, and Maxim Sviri-
denko. Tight approximation algorithms for maximum general assign-
ment problems. In SODA. ACM, 2006.
[138] Wikipedia. Generalized assignment problem. http://en.wikipedia.
org/wiki/Generalized_assignment_problem. [Online accessed Fep.,
2015].
132 References
[139] OpenStreetMap. http://www.openstreetmap.org/. [Online accessed

Fep., 2015].
[140] DLR-Institute of transportation systems. SUMO-Simulation of urban
mObility. http://dlr.de/ts/sumo. [Online accessed Fep., 2015].
[141] Federal Energy Regulatory Commission. Electric power markets: Na-
tional overview. http://www.ferc.gov/. [Online accessed Fep., 2015].

3100000016

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

3100000016

Загружено:

Авторское право:

Доступные форматы

Foundations and Trends

Sustainable Transportation with Electric Vehicles

Fanxin Kong Xue Liu

1 Introduction and Overview 3

2 Charging Rate Control 13

3 Charging Demand Response 32

4 Charging Load Scheduling 64

5 Charging Demand Balancing 97

F. Kong and X. Liu. Sustainable Transportation with Electric Vehicles. Foundations

(a) U.S. EV sales 2010-2015 (b) World-wide EV sales 2013-2020

Figure 1.1: Market research about EV sales.

with some eco-friendly areas such as neighbourhoods with higher pen-

latter method and studies smart EV charging. In the following, we first

1.1 Electric Vehicle Overview

(a) Nissan Leaf 2017 (b) Chevrolet Volt 2017

Figure 1.2: Examples of BEV and PHEV.

Most plug-in EVs of this generation use lithium-ion (li-ion) bat-

1.2 Charging Infrastructure Overview

Charging stations or charging points are a fundamental element in

Table 1.1: Charging Standard Specification [23].

[3, 18]. The second context is park-and-charge, i.e., charging while

1.3 Electric Vehicle Charging Overview

The increasing integration of electric vehicles has motivated many re-

sequence, which is the quality of service degradation caused by EVs’

1.4 A Systematic Solution Framework

This article considers an application scenario, as illustrated in Fig. 1.3,

case, called charging demand response, confines to the demand response

1.5 Road Map

This chapter investigates charging rate control, which addresses the

The rest of this chapter is organized as follows. Section 2.1 presents

2.1 Recent Developments

There is a batch of research efforts that study charging rate control.

2.1.1 Power Grid Oriented Rate Control

applicable. The reason is two-fold. First, the number of EVs in a charg-

The authors of [30] propose a real-time smart load control approach,

tion acquisition. By contrast, the application scenario discussed in this

2.1.2 EV User Oriented Rate Control

Charging Cost Minimization. A customer’s top interest may not

hourly regulation signals) and formulate the problem as a stochastic dy-

2.2 System Model

As depicted in Fig. 2.1, charging station CS has a controller to ad-

‫ܸܧ‬ଵ ‫ܸܧ‬௞ ‫ܸܧ‬௄

‫ݑ‬ଶ ሺ‫݌‬ሻ ൌ ݈‫݃݋‬௔ଶ ሺ‫݌‬ሻ

‫ݑ‬ଷ ሺ‫݌‬ሻ ൌ ݈‫݃݋‬௔ଷ ሺ‫݌‬ሻ

Figure 2.2: Illustration of utility functions of three different EVs.

Case (i): this case is called proportional fairness. To maximize the

2.2.1 Problem Formulation

where p = (pk )∀k is the vector of charging rates. The objective

2.3 Optimal Charging Rate Control

In this subsection, we design a charging rate control algorithm, which

Algorithm 1: Optimal Charging Rate Control

Theorem 2.1. Algorithm 1 optimally determines charging rates for the

Proof. The proof is in Appendix 2.6.

We see that Algorithm 1 efficiently and optimally solves the opti-

2.4 Performance Evaluation

To highlight the benefits of our design for charging rate control, we

Experimental Setup. We consider a station with 20 fast charging

2.4.1 Benefits of Integrating Renewable.

Fig. 2.4(a) compares the total EV charging rate or user satisfaction

(a) Solar trace (b) Base load trace

sources, we plot figures from Fig. 2.4(b) to Fig. 2.4(d) to demonstrate

2.4.2 Importance of Charging Control.

Fig. 2.4(e) shows the individual EV charging rate comparison between

Sum Crg Rate (MW)