You are on page 1of 14

886 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 7, NO.

2, APRIL 2016

Self Scheduling of Plug-In Electric Vehicle


Aggregator to Provide Balancing
Services for Wind Power
Marina Gonzlez Vay, Student Member, IEEE, and Gran Andersson, Fellow, IEEE

AbstractThis paper focuses on the self-scheduling problem by shifting charging in time, using unidirectional charging only
of an aggregator of plug-in electric vehicles (PEVs) purchasing [7], or by additionally discharging energy to the grid, which is
energy in the day-ahead market, and offering balancing services referred to as vehicle-to-grid (V2G).
for a wind power producer, i.e., committing to compensate the
forecast errors of wind power plants. The aggregated charging In this paper, we approach the self-scheduling problem of a
and discharging flexibility of the PEV fleet is represented by a PEV aggregator who bids in the day-ahead market to purchase
probabilistic virtual battery model, accounting for the uncertainty energy on behalf of a PEV fleet, and, at the same time, enters a
in the driving patterns of PEVs. Another source of uncertainty bilateral contract with a wind power producer to balance a frac-
is related to the balancing requests, which are a function of the tion of the producers wind power forecasting errors (respect
forecasted wind power output. A scenario-based robust approach
is used to tackle both sources of uncertainty in a tractable way. to the day-ahead forecast) during the next day. Therefore, the
The interdependency between the day-ahead market prices and aggregator, when deciding on its positions in the day-ahead
the aggregators bidding decisions is addressed using complemen- market, should leave enough charging flexibility to compen-
tarity models. A case study analyzes the capability of the PEV sate prediction errors in the wind power output throughout the
aggregation to provide balancing services, for different settings next day. A co-optimization trades off between the costs in
of the balancing contract, and both with and without the use of
vehicle-to-grid. the day-ahead market and the revenues from providing bal-
ancing services, i.e. in order to provide balancing services the
Index TermsPlug-in electric vehicles, demand-side flexibility,
balancing power, distributed storage, stochastic optimization.
aggregator may incur higher costs in the day-ahead market.
The clearing of the day-ahead market is modeled explicitly
I. I NTRODUCTION using complementarity models [8]. The aggregator must ensure
that the end-use constraints of the PEVs under its management

T WO recent development trends bring about new chal-


lenges, but also new opportunities for power systems. One
is an increasing penetration of plug-in electric vehicles (PEVs),
are not violated. These individual constraints are integrated
into an aggregated virtual battery representation, character-
ized by a set of time-varying power and energy constraints.
which represent at the same time new, mobile loads, and dis- These constraints are parameterized using individual driving
tributed storage resources. The other is the increasing share of patterns (arrival and departure times, trip energy consump-
fluctuating renewable energy sources (RES) in the generation tion) obtained from a transport simulation [9]. A scenario-based
mix, with a weather-dependent output that cannot be perfectly robust approach [10] ensures that the aggregator is able to com-
forecasted. The synergies between PEVs and RES have been ply with the balancing contract without impacting the PEVs
the subject of some studies [1], [2]. On one hand, RES can end-use, given uncertainty a) in driving patterns [11], affecting
provide electricity for vehicle charging at low marginal costs the availability and charging flexibility of the PEVs, and b) in
and CO2 emissions [1]. On the other hand, PEVs could cover the wind power production [12], and therefore in the balancing
part of the increasing need for storage to balance intermittent requests to be serviced.
RES [3]. Some studies [2], [4], [5] conclude that PEVs could A number of papers focus on the provision of ancillary ser-
help integrate wind power into power systems. In order to seize vices by PEVs, such as primary [13], secondary [14], [15] and
the potential of PEV fleets as renewable energy enablers, vehi- tertiary frequency control [16]. In this paper, we specifically
cle charging needs to be managed, either directly or indirectly consider the problem of balancing wind with aggregations of
[6]. The entity in charge of coordinating PEV charging is usu- PEVs, which has been addressed in [17][20]. These papers
ally called aggregator. The charging flexibility can be exploited focus on the balancing process [17][19] or determine poten-
Manuscript received April 14, 2015; revised August 13, 2015 and October tials [20], but do not address how to define the charging
21, 2015; accepted October 30, 2015. Date of publication December 03, baseline, i.e. there is no co-optimization of the goals of energy
2015; date of current version March 18, 2016. This work was supported by cost minimization and wind balancing revenue maximization.
the European Union Seventh Framework Programme (FP7/2007-2013) under
negotiation no. 608957. Paper no. TSTE-00284-2015. This type of co-optimization is found in [21], [22]. A three-level
The authors are with the Power Systems Laboratory, ETH Zurich, Zurich controller is proposed in [21], scheduling power plants, includ-
8092, Switzerland (e-mail: marina.gonzalezvaya@alumni.ethz.ch; andersson@ ing wind, (1st level), PEV demand (2nd level), and real-time
eeh.ee.ethz.ch).
PEV charging (3rd level). In [22], a common bidding strategy
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. for a PEV aggregator and a wind power producer in day-ahead
Digital Object Identifier 10.1109/TSTE.2015.2498521 and regulation markets is proposed.
1949-3029 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
GONZLEZ VAY AND ANDERSSON: SELF SCHEDULING OF PLUG-IN ELECTRIC VEHICLE AGGREGATOR 887

The uncertainty in the flexibility of the fleet is not considered tced cumulative energy deviation due to balancing
in [21]. In [22], PEVs are modeled individually in a centralized tced,V2G cumulative energy deviation due to balancing
optimization problem, which may not be a tractable approach when discharging
for large fleets. Moreover, none of the mentioned approaches
provides (probabilistic) guarantees on the ability of the aggre-
Individual vehicle parameters:
gator to fulfill the balancing service given the uncertainty in V,low
Evt lower energy trajectory
driving patterns and balancing requests. V,low,free
The novel contributions of this paper, which extends our pre- Evt free lower energy trajectory
V,up
vious work in [11], [12], are the following: 1) We propose Evt upper energy trajectory
an aggregated, stochastic representation of the fleets charging V,low
Pvt lower power trajectory
(and discharging) flexibility, and a robust formulation of the V,up
Pvt upper power trajectory
fleets capability to respond to stochastic balancing requests.
uV
vt connection status
2) We assess the impact of the availability of vehicle-to-grid,
and of the characteristics of the balancing contract, on the
ability of a PEV aggregator to offer balancing services. Exogenous inputs:
The paper is structured as follows: II Methodology, III t time step duration
Case study, IV Conclusion. chance constraint confidence parameter
 chance constraint violation parameter
virtual battery efficiency
II. M ETHODOLOGY
et normalized expected down reserve request
After introducing the notation in II-A, the aggregators co- e+t normalized expected up reserve request
optimization problem is described in II-B. Thereafter, the D
cdt market demand bid price
stochastic model representing charging and discharging flex-
cSst market supply bid price
ibility (virtual battery model) is introduced in II-C. Finally, bid,max
II-D describes how to generate scenarios of balancing requests c maximum bid price
deg
(wind power forecast errors) and driving patterns. c degradation costs per unit of processed energy
EtA,ad virtual battery energy contribution of arriving and
departing vehicles
A. List of Symbols A,max
Et virtual battery energy upper bound
Indices: EtA,min virtual battery energy lower bound
t time steps {1 . . . T } A,min,free
d demand bids Et virtual battery free energy lower bound
s supply bids ft normalized wind power forecast
v vehicles g cap, down reserve capacity price
k scenarios {1 . . . K} g cap,+
up reserve capacity price
g en, down reserve energy price
Variables: en,+
g up reserve energy price
E virtual battery energy shift D,max
t day-ahead market clearing price Pdt demand bid volume
S,max
cA
t aggregators demand bid price Pst supply bid volume
Ct down reserve capacity PtA,max virtual battery power upper bound
Ct+ up reserve capacity PtA,max,free virtual battery free power upper bound
E0A virtual battery initial energy content PtA,min virtual battery power lower bound
EtA virtual battery energy content PtA,min,free virtual battery free power lower bound
PdtD
accepted demand bid volume PtA,min,V2G virtual battery power lower bound with V2G
S
Pst accepted supply bid volume
PtA accepted aggregator demand bid volume
B. Co-Optimization Problem
PtA,char expected charging power
The self-scheduling of a PEV aggregator placing demand
PtA,dis expected discharging power
bids in the day-ahead market, and entering a balancing contract
with a wind power producer is formulated as a bilevel optimiza-
Random variables: tion problem [11], see Fig. 1. The upper level represents the
e
t normalized down reserve request co-optimization problem of the aggregator, whereas the lower
e+
t normalized up reserve request level represents the day-ahead market clearing process.The
tx random variable associated with virtual battery day-ahead market modeled here contemplates self-scheduling
parameter x agents submitting independent supply/demand bids for each
888 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 7, NO. 2, APRIL 2016

the capacity revenues from the balancing contract,


  cap,+ 
revcap = gt Ct+ + gtcap, Ct , (3)
t

and the energy revenues and payments linked to the expected


provided balancing energy,
  en,+ en, 
reven = gt e+ +
t Ct g t e t Ct , (4)
t

where e+
t and e
are the absolute values of the expected nor-
t
malized balancing energy requests. Since PEVs are modeled as
Fig. 1. Model overview.
an aggregation, and not individually, degradation costs in (2)
are computed based on the processed energy only, and not on
hour, as is the case in many European day-ahead markets, more complex models considering further parameters such as
such as the Swiss, and German-Austrian day-ahead markets. the depth of discharge [25], [26]. Results in [27] for realistic
Other papers focus on the Unit Commitment problem, includ- V2G utilization suggest that the processed energy is the most
ing PEVs, from the perspective of a system operator [23], [24]. important parameter to quantify battery degradation.The term
When the PEV aggregator is a market player important enough cdeg is computed as in [28, Eq. (13)], as the cost of the battery
to impact market prices, a model with endogenous prices, such divided by the total energy it can process until the end of life.
as the one proposed in this paper, is required. Because charging Since the aggregator groups a number of heterogenous PEVs,
demand is typically very flexible and can be shifted to a few with different battery types, in practice an average value for the
low-price hours, a visible increase in prices during those hours fleet would need to be estimated.
may already occur at modest PEV penetrations [11]. The aggregators minimization problem is subject to the
We assume that the balancing contract can potentially be intertemporal constraints that define the flexibility of the fleet,
separated into up/down balancing reserve components with denoted virtual battery constraints and described in the next
capacity Ct+ /Ct , respectively. Committing Ct+ means that, subsection (II-C). The upper level problem is given by
when the wind power realization at time step t is lower than
the forecasted value, the PEV aggregator has to compensate Min. costDA + costdeg revcap reven (5a)
UL ,LL
the forecast error corresponding to Ct+ installed wind capac- s.t. virtual battery constraints, (5b)
ity. This can be done by reducing the charging demand or by
using V2G. Similarly, down reserves are activated when the PtA + e+ +
t C t e t C t = Pt
A,char
PtA,dis , (5c)
wind outcome exceeds the forecast, and are serviced by increas- 0 PtA,char M uchar
t , (5d)
ing the charging demand. We assume a capacity remuneration,
with capacity prices gtcap,+ /gtcap, for up/down reserves, respec- M (ut 1) PtA,dis 0,
char
(5e)
tively. The energy delivered when providing up reserves is uchar
t {0, 1}, (5f)
remunerated at the price gten,+ and the energy drawn when pro- Ct 0, Ct+ 0, (7),

t, (5g)
viding down reserves is charged at the price gten, . The energy
t c
cA , t,
bid,max
and capacity prices of the balancing contract are considered (5h)
exogenous parameters, i.e. it is assumed that they have been with decision variables for the upper level problem
announced by the wind power producer in advance.1 These   
+
prices would depend in practice on the penalty that the wind UL = cA A,char
t , C t , C t , Pt , PtA,dis , uchar
t t; E0A , (6)
power producer has to pay for the imbalances of the wind power
A,char
plants under its management. where cA t are the prices of the aggregators bids, Pt and
A,dis
In the upper level problem, the aggregator minimizes its net Pt the expected aggregated charging and discharging power
costs, considering the costs of purchasing electricity in the day- values, respectively, defined by the integer variables uchar t
ahead market, through (5d), (5e), and E0A the initial energy content of the
 virtual battery.
costDA = t PtA t, (1) The market clearing price t and the accepted volume of
t the aggregators bid PtA stem from the solution of the market
clearing problem (7) for time step t
the costs from battery degradation when discharging energy  
(PtA,dis t) from the PEV batteries to the grid, Max. dt Pdt + ct Pt
cD D A A
cSst Pst
S
(7a)
LL
t
 
d

s
costdeg = cdeg PtA,dis t, (2) s.t. D
Pdt + PtA = S
Pst : t , (7b)
t s
d
1 Tomodel the energy and capacity prices of the balancing contract as 0 Pdt
D
Pdt
D,max
, d, 0 PtA PtA,max , (7c)
endogenous variables, the competition for the balancing contract could be
modeled by another set of lower level problems. 0 S
Pst S,max
Pst , s, (7d)
GONZLEZ VAY AND ANDERSSON: SELF SCHEDULING OF PLUG-IN ELECTRIC VEHICLE AGGREGATOR 889

with decision variables for the lower level problems 2) Determining the Virtual Battery Parameters: In the fol-

S
lowing, we briefly sketch how the parameters describing the
LL = LL t , t ; LL
t = Pst , s; Pdt , d; Pt
D A
. (8) virtual battery constraints are determined. For this purpose, a
In the lower level problem, the sum of consumer and pro- bottom-up approach is adopted, building on the driving patterns
ducer surplus is maximized, given the demand, aggregator and characteristics of individual PEVs. For each PEV, we define
V,up V,low
(demand), and supply bid prices cD A S upper and lower trajectories, Evt and Evt , for the energy
dt , ct and cst , respectively.
The market equilibrium is enforced by (7b). The associated dual in the battery of each vehicle v at time step t. These define
variable t is the market clearing price. The lower and upper the envelope of feasible energy trajectories. Since we do not
bounds of demand Pdt D
and aggregator demand bid volumes consider a fixed state of charge (SOC) for each vehicles bat-
S
(7c), as well as supply Pst bid volumes (7d), are also enforced. tery at the beginning of the optimization horizon, we compute
The upper level problem (5) and the lower level problems (7) these extreme trajectories under the assumption that the maxi-
need to be jointly solved. For this purpose, the bilevel problem mum SOC is reached at some point. Later, a possible shift from
is reformulated as an equivalent mixed-integer linear program. this reference is introduced. The upper trajectory of the energy
First, the lower level problems (7), which are convex (linear content is calculated assuming charging starts as soon as the
programming), are replaced by their Karush-Kuhn-Tucker con- vehicle parks. The lower energy trajectory corresponds to the
ditions. Second, linear reformulations are established for the case where charging is deferred as much as possible, under the
complementarity slackness conditions, and for the objective conditions that
function of the upper level problem. For brevity, we refer to (a) the vehicle should depart with enough energy in the
the derivation in [11], [29]. battery for the forthcoming trip,
To be able to compute the bilevel problem, the aggregator (b) the maximum SOC should be reached at some time of the
needs to estimate the supply and demand bids in the market. day (as mentioned above), and
Here, the aggregator uses a simple strategy, where the bids of (c) the total daily charge should be equal to the total daily
the previous week are the estimate of the bids of the current consumption.
week. Then, the aggregator uses the thereby computed accepted Because of conditions (b) and (c), the depth of discharge with
respect to the maximum SOC reference cannot be larger than
bid volumes as bid volumes, and sets cA t =c
bid,max
. To address
the uncertainty in market bids more accurately, bid curves can the normalized daily energy consumption.
be constructed by considering several scenarios of the lower When providing balancing services, it might be beneficial
level market clearing, see our work in [29]. to operate the battery around a lower SOC, i.e. to drop con-
dition (b). Moreover, due to random balancing requests, the
total daily charge can be somewhat higher or lower than the
C. Virtual Battery Constraints total daily consumption (violation of condition (c)). To ensure
1) Basic Equations: The charging/discharging flexibility of that the drivers energy needs are covered, only condition (a) is
the fleet is represented by a set of time-varying aggregated required. If conditions (b) and (c) are dropped, the SOC can
power and energy constraints. The basic equations of the virtual reach lower values than previously assumed. The corresponding
V,low,free
battery model are the following: lower energy profile is denoted Evt , and can be interpreted
as the minimum energy that should be in the battery at a given
energy dynamics EtA = E(t1)
A
+ PtA t + EtA,ad , (9a) time step t in order not to impact the end-use constraints of the
PEV driver.
energy bounds EtA,min EtA EtA,max , t, (9b)
As a next step, the upper and lower power trajectories of
V,up
power bounds PtA,min PtA PtA,max , t. (9c) each vehicle, Pvt V,low
and Pvt , are computed. Instead of sim-
V,max
ply assuming that a PEV can charge at maximum power Pvt
Variable EtA,ad represents the net effect of the increase in energy whenever it is plugged in, the fact that the charging power is
content due to the arrival of PEVs at a given time step, and the also constrained by the batterys energy content is taken into
reduction in energy content due to PEV departures. account, which yields
The aggregated parameters EtA,ad , EtA,min , EtA,max , PtA,min ,
   
and PtA,max are determined out of individual PEV parameters V,up
Pvt = min Pvt
V,max V,up
, Evt Ev(t1)
V,low
/ (v t) , (12)
and driving patterns (II-C2), the latter being subject to uncer-
tainty. For this reason, we introduce random variables, denoted    
V,up
, representing the uncertainty in each of the aggregated param-
V,low
Pvt = max 0, Evt
V,low
Ev(t1) / (v t) . (13)
eters, and replace (9) by the energy constraints
t 
  To determine the power trajectories without the full charge
min ad V,up,free
EtA,min + tE E0A + PA t + EA,ad + E (b) and daily energy recovery (c) conditions, denoted Pvt
V,low,free V,low V,low,free
=1 and Pvt , Ev is replaced by Ev in equations (12)
max
EtA,max + tE , t, (10) (13) above. If V2G is allowed, in addition to this change, the
lower bound of zero in (13) is replaced by Pvt V,max
. The cor-
V,low,V2G
and the power constraints responding lower trajectory is denoted Pvt . Note that
min max inelastic charging could be easily modeled by setting both the
PtA,min + tP PtA PtA,max + tP , t. (11) upper and lower power trajectories equal to a fixed trajectory.
890 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 7, NO. 2, APRIL 2016

The aggregated virtual battery parameters are determined out


of the individual PEV trajectories as follows
  V,up
EtA,min = uV V,low
vt Evt , Et
A,max
= vt Evt , t,
uV (14)
v v

and analogously for the other aggregated power and energy


bounds. Binary parameter uV vt is equal to one when PEV v is
connected at time step t, and zero otherwise. The contributions
of arriving and departing vehicles are computed as [11]
  
V,up
EtA,ad = Evt vt + uv(t1)
uV V
vt uv(t1) .
uV V
(15)

3) Providing Balancing Services: To respond to balancing


requests, the aggregator needs to deviate from the aggregated
power and energy setpoints, PtA and EtA respectively, which are
based on its commitments in the day-ahead market. For this
reason, in addition to the basic equations, further constraints
need to be considered to make sure that the perturbed power
and energy profiles stay within predefined bounds.
The aggregator can provide balancing services with unidirec-
tional charging only, i.e. by reducing charging with respect to a
predefined setpoint, or by additionally using V2G, i.e. allowing
discharging of energy from the PEV batteries to the grid. While Fig. 2. Illustration of the energy shift E and the different energy bounds.
only a limited amount of up balancing reserves can be pro-
vided with unidirectional charging, V2G comes at a cost, due
min,free
to the additional battery cycling and the corresponding battery EtA,min,free + tE
degradation [28]. t 
 
When balancing services are to be provided, it makes sense E0A + PA t + EA,ad + E
ad
+ tced E
to drop conditions (b) and (c) in II-C2, and introduce a shift =1
from the full-charge reference, denoted E. Therefore, for the max

perturbed energy profiles, we use an aggregated lower energy A,max


Et + tE , t. (19)
bound EtA,min,free associated with the lower energy trajectories
Constraint (17) ensures that the aggregator can increase the
that hold when assumptions (b) and (c) are dropped,
charging power to compensate the worst-case down-request
 event, which occurs when the wind power plants produce at
EtA,min,free = uV V,low,free
vt Evt , t. (16)
v
their maximum output instead of producing at the forecasted
value ft (normalized). Similarly, the aggregator should be able
Fig. 2 illustrates the different energy bounds and the role of to reduce the scheduled charging power - the worst-case being
the energy shift. Without offering balancing services (Fig. 2a), when no wind power is produced instead of the forecasted
only the region between EtA,min and EtA,max is utilized. When value - without violating its power lower bound (18). Since
balancing is to be provided (Fig. 2b), the batteries are operated this lower bound is nonnegative (no V2G), this means that
around a lower SOC thanks to the shift E. The scheduled the scheduled charging power cannot be lower than the con-
energy EtA should still stay within the bounds EtA,min and tracted up-reserve capacity times ft . Note that in (17)(18) the
EtA,max , represented by the area shaded with the lighter grey power bounds derived using the free lower energy trajectories
color. When balancing requests are serviced, the energy profile are used. Concerning the energy profile, it should stay within
will deviate from the scheduled one. These deviations should the energy bounds, see (19), coping with possible cumulative
stay within the limits EtA,min,free and EtA,max , represented by the energy deviations tced due to providing balancing services,
area shaded with the darker grey color.
Therefore, constraints (10)(11) still remain valid (unper- 
t 
t

turbed schedule). Additional constraints are described in the tced = C e


t C+ e+
t, (20)
following, to be satisfied by the perturbed energy and power =1 =1

profiles.
where e+
t /et is the absolute value of the normalized up/down
4) Constraints to Provide Balancing Services with
balancing energy requested at a given time step, and is a random
Unidirectional Charging: The following constraints are
variable.
added to constraints (10)(11):
Note that a probabilistic approach is chosen for the cumula-
PtA + Ct (1 ft ) PtA,max,free + tP
max,free
, t, (17) tive energy deviations, whereas a worst-case approach is used
for the power deviations. The idea is that, for a very short
Ct+ ft ,
min,free
PtA,min,free + tP PtA t, (18) period of time, the PEV aggregator might have to deal with
GONZLEZ VAY AND ANDERSSON: SELF SCHEDULING OF PLUG-IN ELECTRIC VEHICLE AGGREGATOR 891


the worst possible situation. It would also be possible to adopt a
P , | max a j px + j (y)
probabilistic approach for the power deviations, but only hourly j=1,...,J
wind data was available for this work.
5) Constraints to Provide Balancing Services with V2G: + j bj 0 1 , (26)
The following constraints are added to constraints (10)(11):
where is the vector of uncertainties related to the
PtA + Ct (1 ft ) PtA,max,free + tP
max,free
, t, (21) -variables, the vector of uncertainties related to the -
PtA,min,V2G + tP PtA Ct+ ft ,
min,V2G
t, (22) variables, and  the violation parameter. The function j (y) :
t   R2T R is linear.
 ad
E0A + PA t + EA,ad + E Because the constraints included in the joint chance con-
=1 straint are linear with respect to the uncertainty3 , one way to
max approach the problem is to enumerate all the vertices of the
+ tced E EtA,max + tE , t, (23)
min,free
hyper-rectangle B in which the uncertainty is confined [10],
EtA,min,free + tE E0A + tced,V2G E and enforce the original constraints at all the vertices. This
hr
 t 
ad
 would results in J2n constraints, where J is the number of
+ PA t + EA,ad + E , t. (24) constraints affected by the uncertainty (J = 8T ) and nhr is the
=1 dimension of the hyper-rectangle (specified below). Because of
Constraints (21)(24) are slightly modified compared with the structure of the constraints, and the fact that y is nonneg-
(17)(19). First, the lower bound PtA,min,V2G now allows for ative, it is possible to determine a-priori at which vertex each
negative power values.2 Second, since the up-reserve requests constraint can be potentially binding. Therefore, it is enough
could be (but not necessarily are) provided by discharging to enforce each of the J constraints at a particular vertex, as
batteries, the discharging efficiency is used in (24), i.e. explained in the following. Hence, the adopted robust approach
does not impact the number of constraints of the problem, it just

t 
t affects their specific parameters.
tced,V2G = C e
t C+ e+
t/ (25) 7) Reformulation of Joint Chance Constraint: Considering
=1 =1 a finite number of samples k = {1, . . . , K} of the uncertain
instead of the charging efficiency as in (19) (worst-case). variables, constraints (10)(11), (17)(19) can be reformulated
6) Addressing Uncertainty: The constraints above are as:
affected by two sources of uncertainty: the uncertainty in driv- 
t
E min E ad
ing patterns, represented by the -variables, and the uncertainty Et A,min
+ max kt k
k
related to the balancing requests, represented by the -variables. =1
Due to these uncertainties, we formulate the virtual battery con- 
t
 
straints as a joint chance constraint, that we reformulate into E0A + PA t + EA,ad
a tractable set of constraints based on a scenario-based robust =1

approach [10]. The joint chance constraint includes the basic max 
t
ad
constraints (10)(11), and also EtA,max + min kt
E
E
k , t, (27)
k
(17)(19) in the unidirectional case, or  min 
=1

(21)(24) in the V2G case. PtA,min + max ktP


PtA
k
Reformulating a joint chance constraint is typically compu-  max 
tationally intractable, except under some specific assumptions, PtA,max + min ktP
, t, (28)
k
e.g. for a multivariate normal distribution of the random vari- 
ables [30], [31]. To reformulate the joint chance constraint, the min,free 
t
ad
EtA,min,free + max kt
E
E
k
approach proposed in [10], offering probabilistic guarantees k
=1
on the constraint violation, is applied: A robust optimiza-
tion problem with bounded uncertainty is solved, where the 
t
   ced 
E0A + PA t + EA,ad + min kt E, t,
uncertainty bounds are obtained from a random scenario-based =1
k
program. The bounded uncertainty set is defined by a hyper- (29)
rectangle. This approach is similar to other robust optimization

t
   ced 
approaches, such as the one used in the Unit Commitment prob- E0A + PA t + EA,ad + max kt E
lem in [32], but offers the advantage that the uncertainty bounds =1
k
are computed by means of sampling scenarios. 

t
Defining x RT++1 as x = (E0A ; PtA , t), and y R2T + EtA,max + min kt
E max
E
k
ad
, t, (30)
as y = (Ct+ , t; Ct , t) the joint chance constraint can be k
=1
described generically as 3 Note that the inequality in (26) is linear in and (y), and that (y) is
j j j
2 To establish the added value of V2G for balancing purposes only, the power a linear function of the uncertain inputs e+
t and et .However, unlike in [11],
lower bound in the basic constraint (11) remains unchanged, i.e. discharging there is a product of optimization variables and random variables in the chance
is not used for price-arbitrage. This could be easily included by redefining the constraint, represented by j (y), which makes the reformulation of the joint
lower bounds in the basic constraints. chance constraint more complex.
892 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 7, NO. 2, APRIL 2016

 max,free 
PtA + Ct (1 ft ) PtA,max,free + min ktP
, t, (31) uncertainty.
 ced  This works similarly for the replacement of
 min,free 
k maxk kt by its upper bound.
PtA,min,free + max kt
P
PtA Ct+ ft , t. (32) The uncertain inputs whose upper and lower bounds define
k the hyper-rectangle are those corresponding to the variables
related to driving pattern uncertainty (-variables) and the
Note that tced is a function of optimization variables C
upper and lower bounds of et for all t. Therefore, the dimension
and C for = {1, . . . , t}, see (20),
 cedso
 a further step is of the hyper-rectangle is nhr = 8T and the number of bounds to
needed in the reformulation: mink kt is replaced by its
be computed is reduced to 9T , which imposes the condition
lower bound, i.e.
 
 1 e 1
 ced  t
  K 9T 1 + ln (36)
min kt + C min e
k t
e1
k k
=1
on the number of samples that needs to be extracted.
 t
 
C+ max e+
k t. (33) 10) Reformulation of Joint Chance Constraint - Constant
k Capacities: In the case of capacities
=1  ced  that are constant through-
out the time horizon, mink kt is replaced by the lower
This lower bound can be easily determined ced  since C+ and C bound
are nonnegative. Similarly, maxk kt is replaced by its upper t 
bound, which yields an analogous expression.  ced  
min kt + C min e
k t
Therefore, the uncertain inputs whose upper and lower k k
=1
bounds define the uncertainty hyper-rectangle are: 
Variables related to driving pattern uncertainty: Lower  t

max max,free min t ad min,free


+
C max e+
k t. (37)
bound of tP , tP , tE =1 E , tE k
=1
t E ad
P min
P min,free
=1 for all t. Upper bound of t , t ,
t The fact that the worst-case energy
max ad
tE =1 E for all t. t requests donot necessarily
come consecutively, e.g. maxk ( =1 e
t
) =1 maxk ek
Variables related to balancing request uncertainty: Upper k
is exploited
 ced  here. This works similarly for the replacement of
and lower bounds of e+
t , et for all t. maxk kt by its upper bound.
Therefore, the dimension of the hyper-rectangle is nhr = 9T
The uncertain inputs whose upper and lower bounds define
and the number of bounds to be computed is 11T . According
the hyper-rectangle are those corresponding to the variables
to [10], the number of samples K that needs to be extracted in
related to driving pattern uncertainty
t (-variables)
t and the
order to ensure that the joint chance constraint is violated with
upper and lower bounds of =1 e+ k and

=1 ek for all
probability lower than , with confidence 1 , is
t. Therefore, the dimension of the hyper-rectangle is nhr =
  9T and the number of bounds to be computed is 11T , as in
1 e 1
K 11T 1 + ln . (34) the original case without constraints on the reserve capacities,
e1
see (34).
With V2G, the constraints can be reformulated analogously, 11) Reformulation of Joint Chance Constraint - Symmetric
leading to the same number of samples to be extracted. and Constant Capacities: For constant and symmetric capac-
8) Constraints on Reserve Capacities: Further, it is inter- ities C, both of the simplifications above apply. The uncer-
esting to explore restrictions that possibly apply to the tain inputs whose upper and lower bounds define the hyper-
offered reserve capacities. Two such restrictions are typically: rectangle are those corresponding to the variables related to
Symmetric capacities : Ct := Ct = Ct+ , t. driving pattern uncertainty
t (-variables) and the upper and
/+ /+ .
Constant capacities : C /+ := C(t=1) = = C(t=T ) lower bounds of =1 ek for all t. Therefore, the dimension
Constraints (10)(11) and (17)(19) are reformulated in the of the hyper-rectangle is nhr = 8T and the number of bounds
following for these special cases, reducing the conservativeness to be computed is 9T , as in the case with symmetric reserve
of the adopted robust approach. The same procedure can be capacities, see (36).
followed for the constraints with V2G. 12) Constraint on the Final Energy Content: It is desirable
9) Reformulation of Joint Chance Constraint - Symmetric that the energy at the beginning of the optimization horizon
Capacities: In the case of symmetric capacities, only the nor- differs from that at the end of the horizon as little as possi-
malized energy requests ekt = e+
kt ekt need to be considered,
ble. Considering the expected deviations due to the balancing
which can take positive or negativevalues. requests, e+
t and et yields,
 In this case, a less   
conservative lower bound for mink kt ced
is 
PtA t Ct+ e+
t + C
e
t t + E A,ad
t
t
 ced  
t
 
min kt C max (ek ) t.
ad

k k
(35) = med tE . (38)
=1 t

Because maxk (e+ +


kt ekt ) maxk (ekt ) mink (ekt ) holds, If the balancing requests are biased, or if the aggregator tends to
the approach becomes less conservative with respect to the offer reserves asymmetrically, then this is to be compensated by
GONZLEZ VAY AND ANDERSSON: SELF SCHEDULING OF PLUG-IN ELECTRIC VEHICLE AGGREGATOR 893

TABLE I
V IRTUAL BATTERY C ONSTRAINTS W ITHOUT V2G

1 unconstrained, i.e. no constraints apply on the reserved capacities;


2 symmetric capacities;
3 constant capacities;
4 symmetric and constant capacities.

scheduling either more or less regular charging PtA . The aver- to obtain the scenarios: 1) The probability distributions of the
age energy requests e+ t and et 

can be computed out of the wind power output for each hour of the next day are estimated.
K extracted samples, i.e. et = k e+
+
kt /K, et =

k ekt /K.
A non-parametric distribution is obtained from quantile regres-
Due to the random nature of the balancing requests, the energy sions using the wind forecast as explanatory variable. Past pairs
drawn by the vehicles to charge the batteries will not exactly of data of the forecasted and realized power output are used in
match their daily energy consumption. However, the energy the regression. 2) The series of wind power outputs is trans-
needs of the drivers are not affected by this mismatch, since formed into a multivariate Gaussian random variable using the
enforcing the lower energy bound EtA,min,free ensures that there distributions determined in the first step. The corresponding
is enough energy in the batteries at all times. covariance matrix is calculated. 3) K samples of multivariate
13) Summary: An overview of the virtual battery con- Gaussian random variables are extracted and converted back to
straints for the different balancing contract settings is given in the original wind power output distribution. Out of the K wind
Table I, for the case without V2G. Note that all constraints are power realizations, the corresponding values of the forecast
linear and that (27), (28), (31), (32), (38) apply in all cases. error (balancing requests) are derived.
The corresponding constraints with V2G can be easily 2) Driving Pattern Scenarios: We obtain one driving pat-
obtained by tern per vehicle from a transport simulation. To generate K
replacing PtA,min,free by PtA,min,V2G in (32), samples from this reference sample, we assume that the order of
weighting e+ kt with 1/ instead of when considering the
performed activities (e.g., home-work-shop-home) given by the
energy lower bound. transport simulation is deterministic, but model departure times,
trip durations and trip energy as stochastic variables, see details
in [11]. With these assumptions, different realizations of driv-
D. Generating Scenarios of the Uncertain Inputs ing patterns are generated for each individual vehicle, which are
1) Wind Forecast Error Scenarios: To obtain samples of then used to obtain different samples of the aggregated virtual
the wind energy prediction error, the method in [33], capable battery parameters. Note that the method used to reformulate
of generating stochastic samples of short-term wind genera- the chance constraints (II-C7) can be applied independently of
tion, is used. This method takes the interdependence structure the assumed probabilistic distribution. However, the results, in
of the prediction errors and the predictive distribution of the particular the revenues obtained by the aggregation, are sen-
wind power production into account. The goal is to generate sitive to the assumed distribution: The higher the variability
K scenarios of possible wind power outputs given a day-ahead in driving pattern behavior, the more conservative the aggre-
wind power output forecast. The following steps are followed gator needs to be when placing bids in the day-ahead market
894 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 7, NO. 2, APRIL 2016

TABLE II
AVERAGE H OURLY C APACITY P OTENTIAL [MW]

1 unconstrained, i.e. no constraints apply on the reserved

capacities;
Fig. 3. Wind power profiles used in the case study. 2 symmetric capacities;
3 constant capacities;
4 symmetric and constant capacities.
and offering balancing reserves to the wind power producer.
However, when aggregating a large amount of PEVs, as it is the
case in this paper, a very high predictability can be achieved
in the aggregated behaviour, even when individual behavior is
quite variable [34].

III. C ASE S TUDY


A. Case Study Setup
To define the markets bids, aggregated bid curves of the
Swiss day-ahead power market, commercially available from
the European Energy Exchange (EEX), are used. In particular,
a week in 2013 is simulated (21-27.10.13). Each day is divided Fig. 4. Average hourly offered capacity as a function of the capacity price.
into hourly time steps, i.e. t corresponds to one hour.
A relatively high PEV penetration is assumed, 10%, which B. Results
would correspond to about 430000 PEVs in Switzerland.
The driving patterns used in the case study, obtained from 1) Wind Forecast Error Balancing Capacity Potential:
the agent-based transport simulation MATSim [9], are spe- First, the wind balancing capacity potentials are analyzed, i.e.
cific to Switzerland and represent typical mobility behavior the wind power capacity whose forecast errors the PEV aggre-
on weekdays. The transport simulation only provides data on gation is able to compensate independently of the costs. The
trip distances; the corresponding energy consumption is derived average hourly capacities, computed as
from these with the factor 0.2 kWh/km. The assumed maximum T  

P 
D 
charging power for vehicles is 3.5kW4 and the charging effi- +
Cpdt
+ Cpdt /(P DT ), (39)
ciency 90%. Each vehicle is assigned a battery with the capacity p=1 d=1 t=1
of 24kWh, operated between 20% and 80% SOC.
Wind output and day-ahead forecast data for one year was are displayed in Table II for the different configurations
obtained from NREL. The analyzed wind power forecast pro- (with/without V2G, and the different settings of the contract).
files are displayed in Fig. 3, together with the actual realization, The variable p indexes the wind profiles, d the days of the week,
and 1000 scenarios generated with the method described in II- and t the time steps.
D1. Seven different daily profiles (corresponding to one week) Given that the sum of individual charging rated powers of
are considered. Each of the profiles is combined with each of the PEVs in the fleet is 1505MW, the fleet would be capable of
the days of the simulated week (7x7 combinations).The balanc- balancing the forecast error of a group of power plants which
ing request scenarios e+
kt /ekt result from the difference between
represent 47%-72% of this capacity, which is substantial. The
the wind power output scenarios and the forecast. largest potential is obtained for the unconstrained case, both
For the joint chance constraint described in II-C7, the vio- with and without V2G, followed by the constant, symmetric,
lation parameter  was set to 0.1 and the confidence parameter and symmetric and constant cases.
to 1e7 . The number of samples of driving patterns and bal- Enabling V2G does not significantly increase the potential
ancing requests that need to be extracted are 4416 (or 3657 in in some cases, but has a clear impact on the costs of providing
the case of symmetric bids). balancing, as discussed in the following.
If not stated otherwise, the following values are used for the 2) Offered Capacity as a Function of Capacity
capacity and energy payments gtcap,+ = gtcap, = 10 C/MW/h, Remuneration: Fig. 4 shows the average hourly contracted
gten,+ = gten, = 10 C/MWh, and for the battery degradation capacity as a function of g cap,+ = g cap, . As expected, it
cdeg = 50 C/MWh. Sensitivity analysis on these variables are is possible to offer higher capacities with V2G. Moreover,
carried out. symmetry is a less constraining factor with V2G, since up
4 Here
reserves can be more flexibly offered.
a low charging rate is assumed for all PEVs. It is however possible
V,up
to assign different rated power values Pvt to the different vehicles. It is also
With V2G, the total capacity offered in the case of sym-
possible to define different rated power values for a given vehicle at different metric and constant capacity bids is higher than the one with
time steps, depending on the parking location. constant bids, when the capacity price is around 10 C/MW/h.
GONZLEZ VAY AND ANDERSSON: SELF SCHEDULING OF PLUG-IN ELECTRIC VEHICLE AGGREGATOR 895

Fig. 7. Average hourly offered capacity as a function of the degradation costs


(unconstrained case).
Fig. 5. Tradeoff between average hourly offered capacity and energy costs. The
energy costs are normalized by the costs without providing balancing services.

Fig. 8. Average hourly offered capacity for each of the wind profiles.

Fig. 6. Average hourly offered capacity as a function of the balancing energy


price (unconstrained case). unconstrained capacity offers is shown in Fig. 7. The battery
degradation model in [27] predicts that 5300 cycles at a depth-
This result is counter-intuitive: one would expect to obtain of-discharge of 95% are possible before end of life, which
the lowest offered capacity when all constraints are imposed means that a battery would be able to process 5035 kWh per
on the reserve capacities, since it is the most restrictive case. kWh of battery capacity. Using this assumption, the range of
However, as discussed in II-C, assuming symmetry means that cdeg depicted in Fig. 7 corresponds to battery costs of up to
less scenarios are required for the reformulation of the chance 503.5 C/kWh, which is a reasonable range in the near future
constraint, i.e. potentially less extreme samples need to be con- [35]. Both the average hourly offered capacity and the shares
sidered. The intuition is that, by eliminating some degrees of of up and down reserve capacity are almost not affected by the
freedom, the deviations due to balancing requests become more degradation costs. This is because it is possible to provide a sub-
predictable, and therefore we need to be less conservative about stantial amount of capacity without typically discharging much
the uncertainty related to the balancing requests. energy.
The tradeoff between the energy costs (costDA reven ) and 5) Offered Capacity as a Function of the Wind Profile: The
the offered capacity is illustrated in Fig. 5. The energy costs impact of the shape of the wind profile on the results is ana-
are normalized by the costs obtained when no balancing ser- lyzed in Fig. 8. The fourth wind profile seems to be the most
vices are provided. In some cases, the energy costs are actually favorable for the provision of balancing reserves. The least
reduced compared with that reference (normalized costs lower propitious profile is the sixth one. In fact, the ranking of the
than 1). In fact, a substantial amount of capacity can be offered average offered capacities (computed with gtcap,+ = gtcap, =
at no additional cost, specially for asymmetric capacities. This 10 C/MW/h) corresponds approximately to the inverse ranking
happens when the aggregator needs to buy less energy in the of the average daily forecasted wind output. This result suggests
day-ahead market, i.e. when additional charging is obtained at that, the lower the forecasted wind output, the more attractive
a low cost on average through the provision of down reserves. for the PEV fleet to balance the corresponding forecast error.
Through the additional flexibility provided by V2G, the cor- The reason is that, if the forecasted wind output is low, the
responding tradeoff curves tend to be lower (lower costs) and worst-case up reserve requests (which correspond to the fore-
further to the right (higher capacity) than those without V2G. casted value ft ) are also low. Since PEVs can more easily
The ends of the tradeoff curves represent the capacity that provide down reserves than up reserves, their capability to offer
would be provided if there is no capacity payment (left end), reserves is higher when the forecasted wind output is low. Down
and the maximum capacity that can be offered (right end). reserves are more attractive because they yield cheap additional
3) Offered Capacity as a Function of Energy Remuneration: charging energy through the balancing requests. Also, without
The offered capacity as a function of g en,+ = g en, for the case V2G, providing up reserves is more costly, since it requires
with unconstrained capacity offers is shown in Fig. 6. The aver- scheduling a certain amount of charging power that can be
age hourly offered capacity remains approximately the same reduced upon request. With V2G, energy deviations due to up
independently of the energy price, but the share of up reserve reserves are penalized more strongly, i.e. they are weighted with
capacity becomes larger as the price increases. the discharging efficiency (1/0.9 instead of 0.9).
4) Offered Capacity as a Function of Degradation Costs: 6) Typical Profiles: Some typical energy, charge and
The offered capacity as a function of cdeg for the case with reserve capacity profiles are analyzed, for the first day of the
896 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 7, NO. 2, APRIL 2016

Fig. 11. Impact of uncertainty in balancing requests, computed for the case
where no constraints are imposed on the reserve capacity.

In the cases where no constraints on the offered capacities


apply, the energy profile is shifted downwards almost as much
as possible through E (Fig. 10a, 10c), which allows to pro-
vide a substantial amount of down reserves, while some margin
is left for downward deviations due to up reserves. Only a small
amount of up reserves is provided, since down reserves are
more attractive. Because of this asymmetry, the scheduled final
energy content is lower than the initial one. The gap is expected
Fig. 9. Scheduled power profile (P A ), reference charge profile (P A,ref ) and to be bridged through balancing requests, see (38).
worst case deviations when balancing the wind forecast error. 7) Impact of Uncertainty in Balancing Requests: To estab-
lish the impact of uncertainty in balancing requests, results are
computed over a range of values of the violation parameter
, which are associated with different values of the number
of samples to be extracted, K. To isolate the effect of this
type of uncertainty, perfect information on driving patterns5
and market bids is assumed. Moreover, to make the analysis
more interesting very high capacity remuneration prices are
used gtcap,+ = gtcap, = 1000 C/MW/h, which pushes the aggre-
gator to offer as much capacity as possible and therefore makes
more of the virtual battery constraints binding. A fully robust
approach ( = 0) is introduced for comparison, by considering
the worst possible wind forecast errors, i.e. those corresponding
to no wind or maximum wind power production. The effective
violation probability is validated on a set of 100000 samples,
different to the set used in the optimization (out of sample
test). Fig. 11b shows that the effective violation probability is
several orders of magnitude lower than the nominal one, i.e.
the used scenario-based robust approach is quite conservative
in this context. The impact of the violation parameter on the
offered balancing capacity is displayed in Figure 11a. The curve
is quite steep for low  values, and becomes flatter when higher
violation probabilities are allowed.
8) Impact of Uncertainty in Driving Patterns: The aggre-
Fig. 10. Scheduled energy profile (E A ), reference energy profile (E A,ref ), worst gated driving behaviour can be very accurately predicted here:
case deviations, and initial energy content E0A .
As shown in Fig. 12, the values of the different virtual bat-
tery bounds are almost identical across samples. This is due
simulated week and wind profile 4, see Figs. 9 and 10. In
to the large size of the fleet, since predictability increases with
general, compared with the reference profile (no balancing)
the fleet size [34]. Therefore, we would not expect driving pat-
charging is partially shifted from the night to the afternoon
tern uncertainty to play a major role on its own in this case.
(Fig. 9), which is what leads to higher energy costs. In the case
However, driving pattern uncertainty has theoretically an indi-
without V2G and symmetric and constant capacities, the charge
rect impact on results, because it affects the number of samples
profile is almost flat (Fig. 9b): The scheduled demand needs to
of balancing requests that need to be extracted, see II-C7. The
be at least as high as the offered up reserve capacity times the
revenues are reduced by 46% in the worst case due to this. Since
normalized forecast.
In the case of symmetric and constant capacities, the flatter 5 Since there is no uncertainty in the variables related to driv-

charging schedule leads also to a flatter energy profile (Fig. 10b, ing
 pattern the number of samples is computed as K =
 uncertainty, 
1 e 1
10d), which allows to provide symmetric reserve capacities.  e1
4T 1 + ln
when capacity bids are allowed to be asymmetric.
GONZLEZ VAY AND ANDERSSON: SELF SCHEDULING OF PLUG-IN ELECTRIC VEHICLE AGGREGATOR 897

TABLE IV
AVERAGE N ET R EVENUES PER V EHICLE AND DAY A SSUMING
P ERFECTLY F ORECASTED E XOGENOUS P RICES [ C]

1 unconstrained, i.e. no constraints apply on the reserved

capacities;
2 symmetric capacities;
3 constant capacities;
4 symmetric and constant capacities.
Fig. 12. Energy and power bounds and their related uncertainty. The highest
and lowest values of the bounds (among 4450 samples) are almost identical. TABLE V
P ROBLEM S IZE AND E XECUTION T IME
TABLE III
AVERAGE N ET R EVENUES PER V EHICLE AND DAY [ C]

Table IV with those of Table III under market bid uncertainty,


1 unconstrained, i.e. no constraints apply on the reserved the aggregator does not do better with the exogenous price
capacities;
2 symmetric capacities; model. Note that a very simple model was used to estimate
3 constant capacities; market bids in the bilevel model (using the bids of the previ-
4 symmetric and constant capacities. ous week), so the benefit of using a bilevel model could be
5 perfect information;
larger. Similar results were obtained in a study considering a
6 market uncertainty.
more moderate PEV penetration of 2%, see [11]. However, the
simple model with exogenous prices could be more appropriate
the robust scenario-based approach used here is very conserva- at low PEV penetrations.
tive, as discussed above, from a practical perspective it could 11) Computation: The simulations were run with
make sense to neglect the uncertainty in driving patterns for MATLAB R2014b on a 2.20GHz Intel Xeon processor.
aggregations of large fleets. The optimization problems were solved using CPLEX v12.6.
9) Impact of Uncertainty in Market Bids: In the realistic The size of the problem changes with the day being simulated,
case that the aggregator does not know the market bids of since the number of supply and demand bids to be considered
other participants perfectly, it makes suboptimal bid choices changes. An overview of the problem size and execution times
that lead to lower revenues, see Table III. With the simple bid is given in Table V.
estimation strategy used here, the revenues can drop by up to
18%. However, this result could be improved by using a more
IV. C ONCLUSION
elaborate forecasting technique and/or including several market
bids scenarios [29]. Note that net revenues are positive, which A self-scheduling approach for a PEV aggregator purchas-
means the costs of purchasing energy in the day-ahead market ing energy in the day-ahead market and offering balancing
(and those of battery degradation when using V2G) are more services for a wind power producer was proposed. Results
than compensated by the revenues from providing balancing show that an aggregation of PEVs can compensate the forecast
services. errors of wind power plants of a comparable capacity. Because
10) Exogenous Price Model Benchmark: In this paper, a balancing reserves need to be provided with high reliability,
model with endogenous day-ahead market prices is used, based a scenario-based robust approach is proposed to address the
on complementarity modeling. An alternative would be to con- uncertainty related to balancing requests and driving patterns.
sider the aggregator as a price taker, i.e. assume exogenous Because this type of conservative approach is needed, in some
prices. The revenues that the aggregator would obtain using cases, restraining from using all available degrees of freedom in
such a strategy are displayed in Table IV, assuming the aggrega- the reserve offers can actually be beneficial, since thereby the
tor could perfectly forecast the market prices (more specifically predictability of the balancing requests is increased.
the prices before its impact on prices is taken into account). The proposed method yields aggregated charging profiles
Comparing these results with the results in Table III also only. The aggregator would need a further tool to schedule indi-
under perfect information shows that the assumption of exoge- vidual vehicles, tracking the aggregated profile as accurately as
nous prices is clearly not justified in this case because of the possible. To do this in a tractable way, decentralized algorithms,
high PEV penetration. Even when we compare the results of such as the one we propose in [36], could be used.
898 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 7, NO. 2, APRIL 2016

The presented framework only considers the bidding strategy in islanded electric grids, in Proc. Int. Conf. Clean Elect. Power, 2009,
of an aggregator, taking the strategies of other market partic- pp. 290295.
[20] A. Schuelke and K. Erickson, The potential for compensating wind fluc-
ipants as fixed. To model several strategic market participants tuations with residential load shifting of electric vehicles, in Proc. IEEE
requires extending the formulation in the paper to a more com- Int. Conf. Smart Grid Commun. (SmartGridComm), 2011, pp. 327332.
plex EPEC (equilibrium problem with equilibrium constraints) [21] L. Chiao-Ting, A. Changsun, P. Huei, and S. Jing, Synergistic control of
plug-in vehicle charging and wind power scheduling, IEEE Trans. Power
framework [37], [38]. With several strategic agents, each agent Syst., vol. 28, no. 2, pp. 11131121, May 2013.
faces a bilevel model analog to the one described in the paper, [22] S. I. Vagropoulos, C. K. Simoglou, and A. G. Bakirtzis, Synergistic
and equilibrium constraints couple the individual problems. supply offer and demand bidding strategies for wind producers and
electric vehicle aggregators in day-ahead electricity markets, in Proc.
IREP Symp. Bulk Power Syst. Dyn. Control/IX Optim. Security Control
R EFERENCES Emerging Power Grid (IREP), 2013, pp. 113.
[23] M. A. Ortega-Vazquez, F. Bouffard, and V. Silva, Electric vehicle aggre-
[1] P. H. Andersen, J. A. Mathews, and M. Rask, Integrating private trans- gator/system operator coordination for charging scheduling and services
port into renewable energy policy: The strategy of creating intelligent procurement, IEEE Trans. Power Syst., vol. 28, no. 2, pp. 18061815,
recharging grids for electric vehicles, Energy Policy, vol. 37, no. 7, May 2013.
pp. 24812486, 2009. [24] M. E. Khodayar, L. Wu, and M. Shahidehpour, Hourly coordination of
[2] H. Lund and W. Kempton, Integration of renewable energy into the trans- electric vehicle operation and volatile wind power generation in SCUC,
port and electricity sectors through V2G, Energy Policy, vol. 36, no. 9, IEEE Trans. Smart Grid, vol. 3, no. 3, pp. 12711279, Sep. 2012.
pp. 35783587, 2008. [25] C. Zhou, K. Qian, M. Allan, and W. Zhou, Modeling of the cost of EV
[3] W. Kempton and J. Tomic, Vehicle-to-grid power implementation: From battery wear due to V2G application in power systems, IEEE Trans.
stabilizing the grid to supporting large-scale renewable energy, J. Power Energy Convers., vol. 26, no. 4, pp. 10411050, Dec. 2011.
Sources, vol. 144, no. 1, pp. 280294, 2005. [26] M. A. Ortega-Vazquez, Optimal scheduling of electric vehicle charging
[4] C. K. Ekman, On the synergy between large electric vehicle fleet and and vehicle-to-grid services at household level including battery degrada-
high wind penetrationAn analysis of the Danish case, Renew. Energy, tion and price uncertainty, IET Gener. Transmiss. Distrib., vol. 8, no. 6,
vol. 36, no. 2, pp. 546553, 2011. pp. 10071016, Jun. 2014.
[5] W. Short and P. Denholm, A preliminary assessment of plug-in hybrid [27] S. B. Peterson, J. Apt, and J. F. Whitacre, Lithium-ion battery cell degra-
electric vehicles on wind energy markets, Nat. Renew. Energy Lab., dation resulting from realistic vehicle and vehicle-to-grid utilization, J.
Golden, CO, USA, Rep. TP-620-39729, 2006. Power Sources, vol. 195, no. 8, pp. 23852392, 2010.
[6] M. D. Galus, M. Gonzlez Vay, T. Krause, and G. Andersson, The [28] W. Kempton and J. Tomic, Vehicle-to-grid power fundamentals:
role of electric vehicles in smart grids, Wiley Interdiscip. Rev. Energy Calculating capacity and net revenue, J. Power Sources, vol. 144, no. 1,
Environ., vol. 2, no. 4, pp. 384400, 2013. pp. 268279, 2005.
[7] E. Sortomme and M. A. El-Sharkawi, Optimal charging strategies for [29] M. Gonzlez Vay, L. Baringo, and G. Andersson, Integration of PEVs
unidirectional vehicle-to-grid, IEEE Trans. Smart Grid, vol. 2, no. 1, into power markets: A bidding strategy for a fleet aggregator, in Plug-
pp. 131138, Mar. 2011. in Electric Vehicles in Smart Grid: Management and Control Strategies,
[8] C. Ruiz and A. J. Conejo, Pool strategy of a producer with endoge- S. Rajakaruna, F. Shahnia, and A. Ghosh, Eds. New York, NY, USA:
nous formation of locational marginal prices, IEEE Trans. Power Syst., Springer, 2014.
vol. 24, no. 4, pp. 18551866, Nov. 2009. [30] J. J. Peralta, J. Perez-Ruiz, and S. de la Torre, Unit commitment with
[9] M. Balmer, K. Axhausen, and K. Nagel, Agent-based demand-modeling load uncertainty by joint chance-constrained programming, in Proc.
framework for large-scale microsimulations, Transp. Res. Rec. J. Transp. IEEE Grenoble PowerTech, 2013, pp. 16.
Res. Board, vol. 1985, pp. 125134, 2006. [31] W. van Ackooij, R. A Henrion, A. Mller, and R. A Zorgati, Joint chance
[10] K. Margellos, P. Goulart, and J. Lygeros, On the road between robust constrained programming for hydro reservoir management, Optim. Eng.,
optimization and the scenario approach for chance constrained optimiza- vol. 15, no. 2, pp. 509531, 2014.
tion problems, IEEE Trans. Autom. Control, vol. 59, no. 8, pp. 2258 [32] D. Bertsimas, E. Litvinov, X. A. Sun, Z. Jinye, and Z. Tongxin, Adaptive
2263, Aug. 2013. robust optimization for the security constrained unit commitment prob-
[11] M. Gonzlez Vay and G. Andersson, Optimal bidding strategy of a lem, IEEE Trans. Power Syst., vol. 28, no. 1, pp. 5263, Feb. 2013.
plug-in electric vehicle aggregator in day-ahead electricity markets under [33] P. Pinson, H. Madsen, H. A. Nielsen, G. Papaefthymiou, and B. Klckl,
uncertainty, IEEE Trans. Power Syst., vol. 30, no. 5, pp. 23752385, Sep. From probabilistic forecasts to statistical scenarios of short-term wind
2015. power production, Wind Energy, vol. 12, pp. 5162, 2009.
[12] M. Gonzlez Vay and G. Andersson, Integrating renewable energy [34] J. L. Mathieu, M. Gonzlez Vay, and G. Andersson, Uncertainty in the
forecast uncertainty in smart-charging approaches for plug-in electric flexibility of aggregations of demand response resources, in Proc. 39th
vehicles, in Proc. IEEE Grenoble PowerTech, Grenoble, France, 2013, Annu. Conf. IEEE Ind. Electron. Soc., 2013, pp. 80528057.
pp. 16. [35] B. Nykvist and M. Nilsson, Rapidly falling costs of battery packs for
[13] P. M. R. Almeida, J. A. Peas Lopes, F. J. Soares, and L. Seca, Electric electric vehicles, Nat. Clim. Change, vol. 5, no. 4, pp. 329332, 2015.
vehicles participating in frequency control: Operating islanded systems [36] M. Gonzlez Vay, G. Andersson, and S. Boyd, Decentralized control of
with large penetration of renewable power sources, in Proc. IEEE plug-in electric vehicles under driving uncertainty, in Proc. IEEE PES
Trondheim PowerTech, 2011, pp. 16. Innov. Smart Grid Technol. Conf., Istanbul, Turkey, 2014, pp. 16.
[14] R. J. Bessa, M. A. Matos, F. J. Soares, and J. A. Peas Lopes, Optimized [37] S. Gabriel, A. J. Conejo, J. Fuller, B. Hobbs, and C. Ruiz,
bidding of a EV aggregation agent in the electricity market, IEEE Trans. Complementarity Modeling in Energy Markets. New York, NY, USA:
Smart Grid, vol. 3, no. 1, pp. 443452, Mar. 2012. Springer, 2012, vol. 180.
[15] S. I. Vagropoulos and A. G. Bakirtzis, Optimal bidding strategy for elec- [38] C. Ruiz, A. J. Conejo, and Y. Smeers, Equilibria in an oligopolistic elec-
tric vehicle aggregators in electricity markets, IEEE Trans. Power Syst., tricity pool with stepwise offer curves, IEEE Trans. Power Syst., vol. 27,
vol. 28, no. 4, pp. 40314041, Nov. 2013. no. 2, pp. 752761, May 2012.
[16] R. J. Bessa and M. A. Matos, Optimization models for EV aggrega-
tor participation in a manual reserve market, IEEE Trans. Power Syst.,
vol. 28, no. 3, pp. 30853095, Aug. 2013.
[17] F. K. Tuffner and M. Kintner-Meyer, Using electric vehicles to mitigate Marina Gonzlez Vay (SM11) received the
imbalance requirements associated with an increased penetration of wind Diploma degree in electrical engineering and infor-
generation, in Proc. IEEE Power Energy Soc. Gen. Meeting, 2011, pp. 1 mation technology from the Technische Universitt
8. Mnchen, Munich, Germany, and the Ecole
[18] M. D. Galus and G. Andersson, Balancing renewable energy source Suprieure dElectricit (Suplec), Gif-sur-Yvette,
with vehicle to grid services from a large fleet of plug-in hybrid elec- France, in 2010. She is currently pursuing the Ph.D.
tric vehicles controlled in a metropolitan area distribution network, in degree at Power Systems Laboratory, ETH Zurich,
Proc. CIGRE Int. Symp., 2011. Zurich, Switzerland. Her research focuses on the role
[19] J. A. Peas Lopes, P. Almeida, and F. J. Soares, Using vehicle-to-grid of electric vehicles in smart grids and power markets.
to maximize the integration of intermittent renewable energy resources
GONZLEZ VAY AND ANDERSSON: SELF SCHEDULING OF PLUG-IN ELECTRIC VEHICLE AGGREGATOR 899

Gran Andersson (M86SM91F97) received the


M.S. and Ph.D. degrees from the University of Lund,
Lund, Sweden, in 1975 and 1980, respectively. In
1980, he joined ASEAs, now ABBs, HVDC divi-
sion in Ludvika, Sweden, and in 1986, he was a Full
Professor with Electric Power Systems, KTH (Royal
Institute of Technology), Stockholm, Sweden. Since
2000, he has been a Full Professor in electric power
systems with ETH Zrich (Swiss Federal Institute of
Technology), Zurich, Switzerland, where he also the
Head of the Powers System Laboratory. His research
interests include power systems dynamics and control, power markets, and
future energy systems. He is a Fellow of the Royal Swedish Academy of
Sciences, and of the Royal Swedish Academy of Engineering Sciences. He was
the recipient of the IEEE PES Outstanding Power Educator Award 2007, and
the George Montefiore International Award 2010.

You might also like