Page 1 of 8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
IEEE PES Transactions on Power Systems
Bringing Wind Energy to Market
E.Y. Bitar, R. Rajagopal, P.P. Khargonekar, K. Poolla, P. Varaiya
Abstract—Wind energy is a rapidly growing source of renewable energy. However, the current extra-market approach
to assimilating wind energy into the electric grid will not scale
at deep penetration levels. In our work, we investigate how
an independent wind power producer might bid optimally into
competitive electricity markets for energy. Starting with a simple
model of the uncertainty in the production of power from a wind
turbine farm and a model for the electric energy market, we
derive analytical expressions for the optimal bid amount and
corresponding expected optimal profit. Moreover, as wind is an
inherently variable source of energy, we explore the sensitivity
of optimal expected profit to uncertainty in the underlying wind
process. We also explore the uncertainty in generation. We obtain
analytical expressions for marginal profits from investing in
improved forecasting and local auxiliary generation.
Index Terms—Wind Energy, Smart Grid, Electricity Markets
I. I NTRODUCTION
Global warming, widely regarded as one of the most critical
problems we face, has led to great emphasis on renewable
energy sources such as solar, wind, and geothermal. Many
nations have set ambitious goals for renewable energy penetration. Wind energy is expected to be a major contributor
to the realization these goals. At deep penetration levels, the
significant uncertainty and inherent variability in wind power
pose major challenges in integrating it into the electricity grid.
In this paper, we focus on the scenario in which wind
power producers (WPP) must sell their energy using contract
mechanisms in conventional electricity markets. Our goal is to
formulate and solve problems of optimal contract sizing, value
of sensor information, value of local auxiliary generation,
value of storage, and cost of increased reserves needed to
accommodate the uncertainty in wind power production. We
start with a simple stochastic model for wind power production
and a model for the conventional electricity market for energy.
With these models, we derive explicit formulae for optimal
contract size and the optimal expected profit. Our results
cleanly capture the trade-off between penalty for contract
shortfall and the need to spill some of the wind energy to
increase the probability of meeting the contract. We show
that extra information from meteorological models and data
increases the expected optimal profit. We also make explicit
the relationship between penalty for contract shortfall and
the marginal impact of wind uncertainty on optimal expected
Supported in part by OOF991-KAUST US LIMITED under award number
025478, the UC Discovery Grant ele07-10283 under the IMPACT program,
and NSF under Grant EECS-0925337, and the Florida Energy Systems
Consortium.
Corresponding author: E. Bitar is with the Department of Mechanical
Engineering, U.C. Berkeley ebitar@berkeley.edu
P. Khargonekar is with the Department of Electrical and Computer Engineering, University of Florida ppk@ece.ufl.edu
R. Rajagopal, K. Poolla and P. Varaiya are with the Department of Electrical
Engineering and Computer Science, U.C. Berkeley ramr, poolla, varaiya
[@eecs.berkeley.edu]
profit. We consider the scenario in which the WPP schedules
a capacity reservation to “hedge” against potential shortfalls
corresponding to offered contracts and derive a formula for
optimal contract size. We also explore the role of local
generation and energy storage in managing the operational
and financial risk driven by the uncertainty in generation.
We obtain analytical expressions for marginal profits from
investing in local generation and energy storage. Due to space
constraints, the storage analysis is not included, but can be
found in an extended version [12].
The remainder of this paper is organized as follows: In
Section 2, we provide more detailed background on wind
energy and electricity markets. Our problem formulation is
described in Section 3 and our main results are contained in
Sections 4 through 6. We conduct an empirical study of our
strategies on wind power data obtained from Bonneville Power
Authority in Section 7. Concluding comments and discussion
of current and future research are contained in Section 8.
II. BACKGROUND
A. Wind Energy
Inherent variability of the power output is the most significant difference between wind and traditional power generators.
This variability occurs at various time scales: hourly, daily,
monthly, and annually. Since loads are also uncertain and
variable, the issue of wind power variability does not cause
major problems at low wind energy penetration levels, but
at deep penetration levels it presents major engineering, economic and societal challenges. Recently, National Renewable
Energy Laboratory, has released two major reports [9], [14]
on integration of large amounts of wind power (20-30%)
into the Eastern and Western electric grid interconnections in
North America. These studies show that limitations on the
transmission system, increased need for reserves, impact of
unpredicted large ramps, limited accuracy in wind forecasting,
coordination among and conflicting objectives of independent
power producers, system operators, and regulatory agencies,
are some of the major issues in achieving increased penetration
of wind and solar energy.
Integration of wind power into the power system has been
the subject of many academic and industry studies. Morales
et al [23] formulate and solve a short term optimal trading
strategy problem for a wind power producer. They show how
their problem reduces to a linear programming problem. In
their work, wind uncertainty is dealt with via creation of
scenarios trees. For some earlier work along this approach,
see [2], [21], [24]. Cavallo [7], [8] has studied compressed
air energy storage for utility scale wind farms, and Greenblatt
et al [17] compared gas turbines and compressed air energy
storage (CAES) in the context of wind as part of base-load
electricity generation. Economic viability of CAES in wind
energy systems in Denmark has recently been studied in [20].
IEEE PES Transactions on Power Systems
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
Prediction of wind power generation [18], [26] is vitally
important in wind power integration. Prediction errors decrease with shortening of the prediction horizon and expansion
of the geographic area over which averaging is done. Large
fast ramps in power output cause serious difficulties in the
operation of the power system. Prediction of these ramp events
is challenging and is subject of our current research.
B. Electrical Energy Markets
The deregulation of the electric power industry led to the
development of open markets for electrical energy. The two
dominant modes of trading are bilateral trading and competitive electricity pools. The former involves two parties (a
buyer and a seller) negotiating a price, quantity, and auxiliary
conditions for the physical transfer of energy from the buyer
to the seller at some future time.
In contrast to bilateral trading, a competitive electricity pool
consists of numerous buyers and sellers participating in a
single market cleared by a third party – commonly the ISO.
Essentially, each supplier (consumer) submits an offer (bid) for
energy at a desired price to the ISO. The ISO then combines
these bids and offers to construct aggregate supply and demand
curves, respectively. The aggregate supply curve is constructed
by stacking energy offers in order of increasing offer price. The
aggregate demand curve is similarly constructed by stacking
energy quantities in order of decreasing bid price. The intersection of these curves determines the market clearing price
(MCP). All suppliers that submitted bids at prices below the
MCP are scheduled. Likewise, all consumers that submitted
offers at prices above the MCP get scheduled. All scheduled
parties pay or are paid at the MCP. For a comparison between
bilateral and pool trading, see [31].
For concreteness in our studies, we assume that the wind
power producer (WPP) is participating in a competitive electricity pool, although much of our analysis is portable to the
bilateral trading framework. A common pool trading structure
([22], [23], [19]) consists of two successive ex-ante markets:
a day-ahead (DA) forward market and a real-time (RT) spot
market. The DA market permits suppliers to bid and schedule
energy transactions for the following day. Depending on the
region, the DA market closes for bids and schedules by 10 AM
and clears by 1 PM on the day prior to the operating day. The
schedules cleared in the DA market are financially binding and
are subject to deviation penalties. As the schedules submitted
to the DA market are cleared well in advance of the operating
day, a RT spot market is employed to ensure the balance of
supply and demand in real-time. This is done by allowing
market participants to adjust their DA schedules based on
current (and more accurate) wind and load forecasts. The
RT market is cleared five to 15 minutes before the operating
interval, which is on the order of five minutes.
For those market participants who deviate from their scheduled transactions agreed upon in the ex-ante markets, the
ISO normally employs an ex-post deterministic settlement
mechanism to compute asymmetric imbalance prices. This
asymmetric pricing scheme for penalizing energy deviations
reflects the energy imbalance of the control area as a whole
and the ex-ante clearing prices. For example, if the overall
Page 2 of 8
system imbalance is negative, those power producers with a
positive imbalance with respect to their particular schedules
will receive a more favorable price than those producers who
have negatively deviated from their schedules, and vice-versa.
For a more detailed analysis of electricity market systems in
different regions, we refer the reader to [5], [6], [30], [31].
III. M ODELS FOR W IND P OWER AND M ARKETS
A. Wind Power Model
Wind power w(t) is modeled as a scalar-valued stochastic
process. We normalize w(t) by the nameplate capacity of the
wind power plant, so w(t) ∈ [0, 1]. For a fixed t ∈ R, w(t) is
random variable (RV) whose cumulative distribution function
(CDF) is assumed known and defined as F (w, t) = P(w(t) ≤
w). The corresponding density function is denoted by f (w, t).
In this paper, we will work with marginal distributions
defined on the time interval [t0 , tf ] of width T = tf − t0 .
Of particular importance are the time-averaged density and
distribution defined as
Z
1 tf
f (w, t)dt
(1)
f (w) =
T t0
Z w
Z
1 tf
F (w) =
F (w, t)dt =
f (x)dx
(2)
T t0
0
Also, define F −1 : [0, 1] → [0, 1] as the quantile function
corresponding to the CDF F . More precisely, for β ∈ [0, 1],
the β-quantile of F is given by
F −1 (β) = inf {x ∈ [0, 1] : β ≤ F (x)}
(3)
The quantile function corresponding to the time-averaged CDF
will play a central role in our results.
B. Market Model
The energy market system considered in this paper consists
of a single ex-ante DA forward market with an ex-post
imbalance penalty for scheduled contract deviations. Contracts
offered in the DA market are structured as power levels that
are piecewise constant over contract intervals [typically hour
long]. In the absence of energy storage capabilities for possible
price arbitrage, the decision of how much constant power
to offer over any individual hour-long time interval is independent of the decision for every other time interval. Hence,
the problems decouple with respect to contract intervals and
our analysis focuses on the problem of optimizing a constant
power contract C scheduled to be delivered continuously over
a single time interval [t0 , tf ].
We define p ($/MW-hour) as the clearing price in the
forward market and q ($/MW-hour) as the imbalance penalty
price for [uninstructed] contract shortfalls. The wind power
producer (WPP) is assumed to be a price taker in the forward
market, because the WPP is considered small relative to the
whole market. Moreover, the WPP is assumed to have a zero
marginal cost of production.
Remark 3.1: In this formulation p and q are assumed to be
fixed and known. However, this assumption can be relaxed
to p and q random and time varying without affecting the
tractability of the results as long as they are assumed to be
independent of the wind process w(t). Also, most of the results
Page 3 of 8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
IEEE PES Transactions on Power Systems
to follow can be generalized to the case in which p and q
belong to a class of functions concave in C and convex in the
deviation C − w(t), respectively.
The profit acquired, the energy shortfall, and the amount of
energy spilled by the WPP over the time interval [t0 , tf ] are
defined respectively as Z
tf
Π(C, w)
Σ− (C, w)
Σ+ (C, w)
+
=
=
=
t
Z 0tf
pC − q [C − w(t)] dt
(4)
+
(5)
+
(6)
[C − w(t)] dt
t
Z 0tf
[w(t) − C] dt
t0
where x+ := max{x, 0}. As wind power w(t) is modeled as a
random process, we will be concerned with the expected profit
J(C), the expected energy shortfall S− (C) and the expected
amount of spilled (curtailed) wind energy, S+ (C):
J(C) = E Π(C, w)
(7)
S− (C) = E Σ− (C, w)
S+ (C) = E Σ+ (C, w)
(8)
(9)
Here, the expectation is taken with respect to the random wind
power process w = {w(t) | t0 ≤ t ≤ tf }.
IV. O PTIMAL C ONTRACTS
We begin by defining a profit maximizing contract C ∗ as
C ∗ = arg max J(C).
C≥0
(10)
Theorem 4.1: Define the time-averaged distribution F (w)
as in (2).
(a) An optimal contract C ∗ is given by
p
C ∗ = F −1 (γ) where γ = .
(11)
q
(b) The optimal expected profit, the shortfall, and the
spillage are given by
Z γ
∗
∗
J (C ) = J = qT
F −1 (w)dw
(12)
0
Z γ
∗
∗
C − F −1 (w) dw (13)
S− (C ∗ ) = S−
= T
0
Z 1
−1
∗
F (w) − C ∗ dw. (14)
S+ (C ∗ ) = S+
= T
γ
Proof: Consider item (a). It is straight forward to prove
concavity of J(C) in C on [0, 1] by using the properties: (1)
x+ is sub-additive and (2) x+ is homogeneous of degree one.
Now, notice that J(C) can be rewritten as
Z 1 Z tf
[C − w]+ f (w, t)dt dw
J(C) = pCT − q
0
t0
Z
1
= pCT − qT
C
= pCT − qT
Z
0
0
[C − w]+ f (w)dw
(C − w)f (w)dw.
Clearly J(C) is also continuous in C on [0, 1] for any
probability density function f (w). It follows that the set of all
maxima, denoted by C ⊆ [0, 1], is convex. Hence, any C0 ∈ C
is a global maximum. Moreover, C0 is a global maximum if
and only if zero is contained in the subdifferential set of J at
C0 . For technical simplicity in the proof, we assume that f (w)
is continuous on [0, 1]. It follows that J(C) is differentiable
in C on [0, 1]. Hence, the set of all global maxima is given by
C = {C ∈ [0, 1] : dJ/dC = 0}. Application of the Leibniz
integral rule yields
Z C
dJ
= pT − qT
f (w)dw = pT − qT F (C).
dC
0
Setting dJ/dC = 0 yields the desired optimality condition
γ = F (C).
Now consider item (b). The result is easily proven by
considering the change of variables θ = F (w). We prove only
the result for J ∗ as expressions for the remaining quantities,
∗
∗
S−
and S+
are established analogously. Using a change of
variables, we have
Z C∗
∗
∗
(C ∗ − w)f (w)dw
J = pC T − qT
0
Z γ
= pC ∗ T − qT
(C ∗ − F −1 (θ))dθ
0
Z γ
∗
F −1 (θ)dθ,
= C T (p − γq) +qT
| {z }
0
=0
which gives us the desired result.
Remark 4.2: (Newsboy) The structure of our expected profit
criterion and optimal policy are closely related to the classical
Newsboy inventory problem in economics [25].
Remark 4.3: (Non-uniqueness of C ∗ ) Clearly, any contract
C that solves γ = F (C) is profit maximizing with respect to
problem (10). Because the CDF F is only guaranteed to be
monotone non-decreasing on it’s domain [0, 1], it may have
intervals in its domain on which it is constant, which allows
for non-uniqueness of the optimizer C ∗ . Hence, it is straight
forward to see that C ∗ is unique if and only if the set
Γ(F, γ) := {x ∈ [0, 1] : γ = F (x)}
is a singleton. As stated in Theorem 4.1-(a), a particular choice
for an optimal contract is C ∗ = F −1 (γ) – the γ th - quantile of
F , as specified in equation (3). Although the optimal expected
profit J ∗ is independent of the choice of C ∗ ∈ Γ(F, γ), it is
straightforward to see that C ∗ = F −1 (γ) is the minimizer of
∗
the expected optimal shortfall S−
among all contracts C ∈
∗
Γ(F, γ). The opposite is true for S+
. The effect of alternative
∗
∗
and S−
is quantified as
choices of C ∗ from Γ(F, γ) on S+
−1
∗
−1
∗
(15)
S− (C ) = S− F (γ) + γ C − F (γ)
∗
−1
∗
−1
S+ (C ) = S+ F (γ) − (1 − γ) C − F (γ)
(16)
for C ∗ ∈ Γ(F, γ).
Remark 4.4: (Graphical Interpretation) Theorem 4.1-(b)
provides explicit characterizations of the optimal expected
∗
∗
profit J ∗ , energy shortfall S−
, and energy spilled S+
. These
three quantities can be graphically represented as areas
bounded by the time-averaged CDF F (w) as illustrated in
Figure 1 for γ = 0.5.
J ∗ = qT A1 ,
∗
S−
= T A2 ,
∗
S+
= T A3
IEEE PES Transactions on Power Systems
From Figure (1), it is apparent that a reduction of “statistical
dispersion” in the time-averaged distribution F (w) will result
in an increase in optimal expected profit (A1 ) and a decrease
in the optimal expected energy shortfall and spillage (A2 , A3 )
– all of which are favorable consequences.
1
A3
0.8
F (w)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
0.6
γ = 0.5
0.4
A1
0.2
0
0
A2
C∗
0.2
0.4
0.6
0.8
w (MW generation/capacity)
1
∗,
Fig. 1. Graphical interpretation of (A1 ) optimal profit J ∗ , (A2 ) deficit S−
∗ where γ = 0.5.
and (A3 ) spillage S+
Remark 4.5: (Price Elasticity of Supply) Under certain assumptions, the quantile rule (11) in Theorem 4.1 can be
interpreted as the supply curve for the WPP. Of primary
importance is the assumption that the WPP is a price taker
in the DA forward market, ensuring that it weild no influence
over the market price. This is reasonable given the relatively
low penetration of wind energy in existing markets. For a fixed
deviation penalty price q, one can interpret the optimal quantile
rule (11) as indicating the amount of energy that the WPP is
willing to supply at a price p. Specifically, the supply curve
is given by
p
−1
C(p) = F
q
The total energy that the WPP is willing to supply at price p
over the time interval [t0 , tf ] is then T C. With this explicit
characterization of the WPP’s supply curve, the price elasticity
of supply, EC , can be readily derived as
EC :=
dF −1 (γ)
γ
γ
d ln C(p)
= −1
=
.
d ln p
F (γ)
dγ
Cf (C)
Remark 4.6: (Role of γ, VaR) It follows from the γ-quantile
structure of the optimal solution that the optimal contract C ∗
is chosen to be the largest contract C such that the probability
of a shortfall occurrence – with respect to the time averaged
distribution F (w) – is less than or equal to γ. In the finance
literature, C ∗ is referred to as the γ-Value-at-Risk (VaR),
where 1 − γ is interpreted as a confidence level.
Clearly then, the price-penalty ratio γ = p/q plays a critical
role in implicitly controlling the probability of shortfall with
respect to optimal offered contracts C ∗ = F −1 (γ). Consider
the scenario in which the ISO has direct control over the
shortfall deviation penalty price q. As the penalty price q
becomes more harsh, (i.e., larger), the price ratio γ decreases
– resulting in smaller offered contracts C ∗ . This follows from
the fact that the quantile function F −1 (γ) is non-decreasing
in γ (non-increasing in q). Consequently, the probability of
shortfall F (C ∗ , t) with respect to the optimal contract C ∗ , is
non-increasing in q.
Page 4 of 8
∗
Remark 4.7: (Spillage) The expected optimal shortfall S−
can be further interpreted as the expected amount of energy
supplied by the ISO to balance the shortfalls in the WPP’s
contractual obligation. A straightforward corollary of Theorem
∗
∗
4.1 is that the expected optimal shortfall S−
and spillage S+
are monotonically non-decreasing and non-increasing in γ,
respectively. This makes explicit the claim that some wind
energy must be spilled in order to reduce the amount of operational reserve capacity needed to hedge against uncertainty
in the wind power.
Remark 4.8: We are able to derive additional insight into
the structure of the optimal solution by considering an equivalent reformulation of the optimal expected profit J ∗ (12)
in terms of conditional expectation. Straightforward manipulations yield J ∗ = pT C̃, where C̃ = E [ W | W ≤ C ∗ ]
and expectation is taken with respect to the time averaged
distribution F (w) conditional on the shortfall event {W ≤
C ∗ }. In the finance literature, the quantity C̃ is referred to
as the conditional Value-at-Risk (CVaR) [27] – the mean of
the γ probability tail. This reformulation reveals that optimal
expected profit J ∗ is equal to the revenue associated with C̃.
∗
Moreover, it follows that the expected optimal shortfall S−
∗
−1
is proportional to the gap between the VaR, C = F (γ),
and the CVaR, C̃.
∗
= γT C ∗ − C̃
S−
V. ROLE OF I NFORMATION
Intuitively, an increase in uncertainty in future wind power
output will increase contract sensitivity to the price-penalty
ratio γ = p/q. Hence, it is of vital importance to understand
the effect of information [such as available implicitly through
forecasts] on expected optimal profit. Consider a simple simple
scenario in which the WPP observes a random variable Y that
is correlated to the wind process w(t). The random variable
Y can be interpreted as an observation of a meteorological
variable relevant to the wind. Using the results in theorem
4.1, it is natural to define the optimal expected profit J ∗ (y)
conditional on information Y = y as
Z γ
J ∗ (y) = qT
F −1 (w|y)dw.
0
1
T
R tf
F (w, t|y)dt and F (w, t|y) is the
where F (w|y) :=
t0
CDF of w(t) conditioned on the realization Y = y.
Theorem 5.1:
E [ J ∗ (Y ) ] ≥ J ∗
∗
Proof: Define C (y) as a profit maximizing contract
conditional on the observation Y = y. More precisely,
C ∗ (y) = arg max E[Π(C, w) | Y = y],
C∈[0,1]
where expectation is taken with respect to the time-averaged
conditional distribution. The following inequality holds for all
C0 ∈ [0, 1] by optimality of C ∗ (y).
Z 1
J ∗ (y) = pC ∗ (y) − q
[C ∗ (y) − w]+ f (w|y)dw
0
Z 1
≥ pC0 − q
[C0 − w]+ f (w|y)dw
0
Page 5 of 8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
IEEE PES Transactions on Power Systems
Taking expectation with respect to y of both sides of the
inequality yields
Z Z 1
E[J ∗ (Y )] ≥ pC0 − q
[C0 − w]+ f (w|y)dwf (y)dy
0
y
= pC0 − q
Z
1
0
[C0 − w]+ f (w)dw.
The equality follows from a straightforward application of
Bayes rule. Finally, maximizing the right hand side with
respect to C0 yields the desired result.
Remark 5.2: It follows from Theorem 5.1 that information
helps in the metric of expected profit. Moreover, figure 1 offers
some intuition as to how a reduction in “statistical dispersion”
of the CDF F results in increased expected optimal profit.
A. Quantifying the effect of uncertainty
It is of interest to quantify the marginal improvement of
expected optimal profit with respect to information increase in
various metrics of dispersion. In practice, there are numerous
deviation measures of dispersion of probability distributions
(e.g. standard deviation, mean absolute deviation). In [27], the
authors take an axiomatic approach to construct a class of deviation measures for which there is a one-to-one correspondence
with a well known class of functionals known as expectationbounded risk measures. We refer the reader to [1], [27] for a
detailed exposition on topic. For our purposes, it suffices to
Z
realize that
1 γ −1
Dγ (X) = E[X] −
F (x)dx
(17)
γ 0
is a valid deviation measure [27] for all square-integrable
random variables X with CDF F (x) and γ ∈ (0, 1). It is
sometimes referred to as the conditional value-at-risk (CVaR)
deviation measure.
This particular choice of the CVaR deviation measure Dγ
is special in that it permits the analytical computation of the
marginal improvement of optimal expect profit J ∗ with respect
to the wind variability, as measured by Dγ . Simple algebraic
manipulation of the formula for optimal expected profit (12)
reveals J ∗ to be an affine function in Dγ (W ), where W is
distributed according to the time averaged distribution F (w).
J ∗ = pT ( E[W ] − Dγ (W ) )
(18)
This result quantifies the increase in expected profit that results
from a reduction in Dγ (W ) using sensors and forecasts.
Further, it makes explicit the joint sensitivity of optimal
expected profit J ∗ to uncertainty and prices.
Remark 5.3: (Role of γ) As we discovered earlier, the pricepenalty ratio, γ = p/q, plays a role in controlling the probability of shortfall with respect to optimal bids C ∗ = F −1 (γ). In a
related capacity, the price-penalty ratio γ also acts to discount
the impact of uncertainty in the underlying wind process, w(t),
on optimal expected profit J ∗ . This assertion is made rigorous
by the fact that Dγ (W ) is monotone non-increasing in γ for
a fixed W . Its limiting values are given by
lim Dγ (W ) =
E[W ] − inf W = E[W ]
lim Dγ (W ) =
0
γ→0
γ→1
Remark 5.4: (Effect of Uncertainty) This interpretation of
optimal expected profit agrees with intuition. As the “uncertainty” in the random wind process decreases, we have
that Dγ (W ) decreases monotonically to zero resulting in J ∗
approaching pT E[W ]. In the limit where the uncertainty goes
to zero (i.e., Dγ (W ) → 0) we have that
Z tf
w(t)dt
lim J ∗ = p
Dγ (W )→0
t0
where {w(t) | t ∈ [t0 , tf ]} is known for Dγ (W ) = 0.
Example 5.5: (Uniform Distribution) It is informative to
consider the case in which F (w) is taken be a uniform
distribution having support on a subset of [0, 1]. Under this
assumption, it is straightforward to compute the optimal expected profit as
√
J ∗ = pT E[W ] − σ 3(1 − γ) ,
where σ is the standard deviation of W – the most commonly
used measure of statistical dispersion. The marginal expected
profit with respect to wind uncertainty, as measured by σ, is
√
dJ ∗
= −pT 3(1 − γ).
dσ
A direct consequence is that the expected profit’s sensitivity to
uncertainty, σ, increases as the penalty price q becomes more
harsh – or equivalently, as γ → 0.
VI. ROLE OF R ESERVE M ARGINS
G ENERATION
AND
L OCAL
A. Reserve Margins
In order to maintain reliable operation of the electric grid,
the ISO is responsible for procuring ancillary services (AS) to
balance potential deviations between generation and load. The
various underlying phenomena responsible for these deviations
result in system imbalances with varying degrees of uncertainty on differing time scales. In order to absorb this variability on the different time scales, multiple ancillary services
must be procured. Broadly, these services consist of regulation,
load-following, reserve (spinning and non-spinning), voltage
control, and reactive power compensation.
Based on the scheduled energy, the ISO first determines the
total reserve requirement for the entire control area needed to
satisfy pre-specified reliability criteria. The ISO then assigns to
each participating load serving entity (LSE) a share of the total
reserve requirement based roughly on its scheduled demand,
because of the uncertainty in load [16]. Each LSE has the
option to procure all or a portion of its reserve requirement
through bilateral contracts or forward markets. The remaining
portion of the reserve requirement not provided by the LSE
is procured by the ISO through ancillary service markets. A
detailed exposition on ancillary services can be found in [31].
Wind power is inherently difficult to forecast. Moreover,
it exhibits variability on multiple time scales ranging from
single-minute to hourly. It follows then that regulation, loadfollowing, and reserve services will be necessary to compensate imbalances resulting from fluctuations in wind[13]. To
simplify our analysis, we will lump all of these ancillary
services into a single service that we refer to as “reserve
IEEE PES Transactions on Power Systems
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
margin”. Under the current low capacity penetration levels of
wind power (∼ 1%), the added variability of wind is largely
absorbed by existing reserve margins used to cover fluctuations
in the load. As the capacity penetration of wind increases,
its affect on operating reserve margins will become more
pronounced [13], [14]. Moreover, it will become economically
infeasible to continue the socialization of the added reserve
costs, stemming from wind variability, among participating
LSEs. Hence, it is likely that the wind power producer (WPP)
will have to bear the added cost of reserve margins [10], [34].
Consider now a departure from the current deterministic
practice of scheduling reserve margins. Assume that the WPP
is responsible for procuring enough reserve power such that
the loss of load probability (LOLP), with respect to an offered
contract C, is less than or equal to ǫ. More precisely, given
an offered contract C on interval [t0 , tf ], the WPP procures
reserve capacity R(C, ǫ) satisfying
R(C, ǫ) = min R s.t. P R ≤ [C − w(t)]+ ≤ ǫ
R≥0
for all t ∈ [t0 , tf ]. It is straightforward to show that the reserve
capacity R(C, ǫ) is given by
+
R(C, ǫ) = [C − δ(ǫ)] ,
(19)
where δ(ǫ) := mint F −1 (ǫ, t).
Remark 6.1: (Interpretation of δ(ǫ)) The probabilistic
quantity δ(ǫ) can be interpreted as the highest power level
that the WPP can supply constantly over the time interval
[t0 , tf ] with probability greater than or equal to 1 − ǫ. Hence,
in order to satisfy the 1 − ǫ reliability criteria with respect to
an offered contract C, the WPP will procure enough reserve
power capacity R(C, ǫ) to cover the additional risk introduced
by contracts larger than δ(ǫ). For contracts C less than δ(ǫ),
the ISO schedules no reserve power, as indicated by (19).
Ex-ante, the WPP must make a capacity payment for the
operating reserve at the capacity price qc . Ex-post, if the
WPP under produces with respect to the offered contract C, it
must make an energy payment. For shortfalls less than R(C, ǫ),
the WPP pays at the energy price qe . For shortfalls larger
than R(C, ǫ), the WPP pays at the imbalance energy penalty
price q. All of the prices are in units of ($/MW-hour). This
augmented penalty mechanism is captured by the following
penalty function φ : R × R+ → R+ .
qx − (q − qe )R x ∈ (R, ∞)
φ(x, R) =
(20)
q x
x ∈ [0, R]
e
0
x ∈ (−∞, 0)
It follows that the fiscal cost and benefit of reserve capacity
to the WPP can be explicitly accounted for in the following
expected profit criterion
Z tf
pC − qc R(C, ǫ) − φ (C − w(t), R(C, ǫ)) dt
JR (C) = E
t0
(21)
Fiscal benefit is derived from the assumption that qc ≤ p.
Moreover, it is assumed that qe ≤ q, guaranteeing convexity
∗
as
of (20) in x. Define a profit maximizing contract CR
Page 6 of 8
∗
CR
= arg max JR (C)
(22)
C≥0
Theorem 6.2: Define the time-averaged distribution F (w)
as in (2).
∗
(a) An optimal contract CR
is given
p − qc
∗
.
(23)
CR
= F −1 (γR ) where γ =
qe
(b) The optimal expected profit is given by
∗
JR (CR
) = qc T δ(ǫ) [ 1 + F (δ(ǫ)) ]
Z
Z γR
−1
+ qe T
F (w)dw − qT
F (δ(ǫ))
δ(ǫ)
F (w)dw
(24)
0
Proof: The proof technique parallels that of Thm. 4.1.
Remark 6.3: It is interesting to note that the optimal con∗
tract size CR
offered by the WPP does not depend on the
LOLP ǫ and the imbalance penalty price q. However, the
expected profit certainly does depend on these parameters.
B. Local Generation
As ancillary service markets have been known to exhibit
price volatility [29], it may be advantageous for the WPP to
circumvent the AS markets and procure its reserve margin
from a small fast-acting generator co-located with its wind
farm. Consider the scenario in which the WPP has at its
disposal a co-located generator of power capacity L and
operational cost qL ($/MW-hour). Hence, the local plant can be
used to mitigate financial risk by covering contract shortfalls
up to a limit L at a reduced energy cost qL < q.
It is straightforward to capture this scenario by reinterpreting the penalty function (20) and expected profit criterion (21)
through an alternative choice of parameters.
qc = 0,
qe = qL ,
R(C, ǫ) = L.
Because of the significant capital costs associated with the
investment in local generation, it’s important to quantify the
marginal improvement in profit resulting from the investment
in a generator with small power capacity L. The following
theorem distills this notion.
Theorem 6.4: Define the time-averaged distribution F (w)
as in (2).
(a) An optimal contract CL∗ is given by any solution C of
p = qL F (C) + (q − qL )F (C − L).
(b) The marginal expected optimal profit with respect to
power capacity L is given by
qL
dJ ∗
pT
= 1−
dL L=0
q
Proof: The proof for part (a) follows from direct application of the proof technique for theorem 4.1-(a).
Part (b) is proven as follows. It is straightforward to show
that the expected profit is given by
Z C−L
J(C, L) = pCT − T
[q(C − w) − (q − qL )L] f (w)dw
0
Z C
− T
qL (C − w)f (w)dw
C−L
Page 7 of 8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
IEEE PES Transactions on Power Systems
for any choice of C and L. Taking the derivative with respect
to L yields
dJ(C, L)
= (q − qL )F (C − L)T.
dL
Taking the limit as L goes to zero and substituting for the first
order optimality condition (a) yields the desired result.
Remark 6.5: For operational cost qL ≤ p, the optimal bid
is given by CL∗ = F −1 (γ) + L. For qL ≥ q, we have CL∗ =
F −1 (γ). Moreover, for L small, we have CL∗ ≈ F −1 (γ) –
recovering the optimal policy in Theorem 4.1.
VII. E MPIRICAL S TUDIES
Using a wind power time series data set provided by the
Bonneville Power Administration (BPA), we are in a position
to illustrate the utility and impact of the theory developed in
this paper.
A. Data Description
The data set consists of a time series of measured wind
power aggregated over the 14 wind power generation sites in
the BPA control area [3]. The wind power is sampled every 5
minutes and covers the 2008 and 2009 calendar years. Accompanying the measured wind power is a time series of rolling
one hour-ahead forecasts sampled at the same frequency. To
account for additional wind power capacity coming online at
various points in time over the 2-year horizon, all of the data
are normalized by the aggregate nameplate power capacity of
the wind farms.
B. Empirical Probability Model
As stated earlier, wind power is modeled as a continuous
time stochastic process whose marginal cumulative distribution
is denoted by F (w, t). While the identification of stochastic
models that accurately capture the statistical variability in wind
power is of critical importance, this is not the focus of our
paper. We will make some simplifying assumptions on the
underlying physical wind process to facilitate our analysis.
A1: The wind process w(t) is assumed to be first-order
cyclostationary in the strict sense with period T0 = 24 hours
– i.e F (w, t) = F (w, t + T0 ) for all t [32], [11]. Thus, we are
ignoring the effect of seasonal variability.
A2: For a fixed time τ , the discrete time stochastic process
{w(τ + nT0 ) | n ∈ N} is independent in time (n).
Fix a time τ ∈ [0, T0 ] and consider a finite length sample
realization of the discrete time process zτ (n) := w(τ + nT0 )
for n = 1, · · · , N . Using this data set, we take the empirical
distribution F̂N (w, τ ) as an approximation of the underlying
distribution F (w, τ ):
N
1 X
F̂N (w, τ ) =
1 {zτ (n) ≤ w}
N i=n
(25)
Invoking the strong law of large numbers under the working
assumptions, it can be shown that the F̂N (w, τ ) is consistent
with respect to F (w, τ ) [4]. Figure 2 (a) depicts nine representative marginal empirical distributions identified from the BPA
data set described earlier. Note that the times corresponding to
the nine distributions are equally spaced throughout the day
to provide a representative sample. Figure 2 (b) depicts the
trajectory of the empirical median F̂N−1 (0.5, t) and its corresponding interquartile range [F̂N−1 (0.25, t), F̂N−1 (0.75, t)].
C. Optimal Contracts in Conventional Markets
Using empirical wind power distributions identified from
the BPA wind power data set, we are now in a position
to compute and appraise optimal day-ahead (DA) contracts
offered by a representative Oregon wind power producer
(WPP) participating in the idealized market system described
in Section IV. We are also able to examine the effect of γ on
∗
∗
J ∗ , S−
, and S+
using this particular characterization of wind
uncertainty. The following empirical studies assume a contract
24
structure {[ti−1 , ti ), Ci }i=1 , where [ti−1 , ti ) is of length one
hour for all i.
Remark 7.1: (Optimal DA Contracts) Figure 2 (c) depicts
∗
) for various price ratios γ =
optimal contracts (C1∗ , · · · C24
0.3, 0.4, · · · , 0.9. As expected, as the price-penalty ratio γ =
p/q decreases, the optimal contract C ∗ decreases. From Figure
2 (c), it is evident that WPPs will tend to offer larger contracts
during morning/night periods when wind speed is typically
higher than during mid-day (as indicated by Figure 2 (b)).
Remark 7.2: (Profit, Shortfall, and Spillage) Figures 3 (a)
and (b) demonstrate the effect of the price-penalty ratio γ
on the optimal expected profit, energy shortfall, and energy
∗
∗
spillage. The units of S−
and S+
are (MW-hour)/(nameplate
capacity), while the units of J ∗ are in $/(q · nameplate
capacity). When γ = 1, the WPP sells all of its energy
production at price p = q. In this situation, the expected
profit per hour (see Figure 3 at γ = 1) of approximately 6.4
24
equals the ratio of average production to nameplate capacity.
This number is consistent with typical values of the wind
∗
∗
production capacity factor. The spillage S+
and shortfall S−
are relatively insensitive to variations in γ [for γ ∈ [0, 0.1]]
because the marginal empirical distributions are steep here.
D. Local Generation
We now consider the optimal contract sizing formulation in
section VI. Figure 3 (c) depicts the marginal expected optimal
profit with respect to local generation power capacity L. As
q → qL , the marginal value of local generation diminishes.
VIII. C ONCLUSION
In this paper we have formulated and solved a variety
of problems on optimal contract sizing for a wind power
producer operating in conventional electricity markets. Our
results have the merit of providing key insights into the tradeoffs between a variety of factors such as penalty for shortfall,
cost of reserves, value of storage and local generation, etc. In
our current and future work, we will investigate a number of
intimately connected research directions: improved forecasting
of wind power, optimization of reserve margins, making wind
power dispatchable, network aspects of renewable energy
grid integration, and new market structures for facilitating
integration of renewable sources. We are also studying the
important case of markets with recourse where the producer
has opportunities to adjust bids in successive stages. We are
also developing large scale computational simulations which
can be used to test the behavior of of simplified analytically
tractable models and suggest new avenues for research applicable to real-world grid-scale problems.
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
Page 8 of 8
F̂N−1 (.25, t)
0.8 γ = 0.9
0.7
F̂N−1 (.50, t)
0.6
F̂N−1 (.75, t)
0.5
C ∗ (t)
1
0.4
0.3
0.2
0
0
0.2
0.4
0.6
0.8
w (MW generation/capacity)
1
0.1
0
0
5
10
15
t (hours)
0 γ = 0.3
0
5
20
10
15
t (hours)
20
−1
Fig. 2. (a) Empirical CDFs F̂N (w, τ ) for nine equally spaced times throughout the day, (b) Trajectory of the empirical median F̂N
(.5, t) and its
−1
−1
corresponding interquartile range [F̂N
(.25, t), F̂N
(.75, t)], (c) Optimal contracts offered in the DA market for various values of γ = 0.3, 0.4, · · · , 0.9.
0.5
8
0.8
∗
S+
J∗
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
F̂N (w, t)
IEEE PES Transactions on Power Systems
7
0.7
6
0.6
5
0.5
4
0.4
3
0.3
2
0.2
1
0.1
0.4
∗
S−
0.3
qL = p
0.2
0.1
0
0
0.2
0.4
γ
0.6
0.8
1
0
0
0.2
0.4
γ
0.6
0.8
1
qL = q
0
0.2
0.4
0.6
0.8
L
Fig. 3. (a) Optimal expected profit J ∗ as a function of γ, (b) Optimal expected energy shortfall and spillage for the 12th hour interval, as a function of γ,
(c) Marginal expected optimal profit with respect to power capacity L of a local generation plant for various operational costs qL ∈ [p, q].
R EFERENCES
[1] P. Artzner, F. Delbaen, J.-M. Eber, D. Heath, “Coherent Measures of
Risk,” Mathematical Finance 9 (1999), 203-227.
[2] G. N. Bathurst, J.Weatherill, and G. Strbac, “Trading wind generation
in short term energy markets,” IEEE Trans. Power Syst., vol. 17, pp.
782-789, 2002.
[3] http://www.transmission.bpa.gov/business/operations/Wind/default.aspx
[4] P.J. Bickel, K.A. Doksum, “Mathematicel Statistics, Vol 1: Basic ideas
and selected topics,” Prentice Hall, New Jersey, Second Edition, 2001.
[5] CAISO
Market
Scheduling
and
Bidding
Guidelines,
Jul.
2004.
[Online].
Available:
http://www.caiso.com/docs/2004/01/16/200401160832403633.pdf.
[6] CAISO Market Simulation Guidebook, Jul. 25, 2008. [Online]. Available: http://www.caiso.com/18d3/18d3d1c85d730.pdf.
[7] A. Cavallo, “High capacity factor wind energy systems,” J Sol Energy
Eng 1995;117(5):13743.
[8] A. Cavallo, “Controllable and affordable utility-scale electricity from
intermittent wind resources and compressed air energy storage (CAES),”
Energy, 2007;32(2):1207.
[9] EnerNex Corp., Eastern Wind Integration and Transmission Study,
National Renewable Energy Laboratory, Report NREL/SR-550-47078,
January 2010.
[10] M. Del Franco, “Wind energy under fire within ERCOT,”
http://www.windaction.org, April 4, 2010.
[11] W.A. Gardner, A. Napolitano, and L. Paura, “Cyclostationarity: Half a
century of research,” Signal Processing (Elsevier) (2006) 639–697.
[12] http://jagger.me.berkeley.edu/ ebitar
[13] GE Energy, “Analysis of Wind Generation Impact on ERCOT Ancillary
Services Requirements”
[14] GE Energy, Western Wind and Solar Integration Study, National Renewable Energy Laboratory, Report NREL/SR-550-47434, May 2010.
[15] Global Wind 2009 Report, Global Wind Energy Council, Brussels,
Belgium, 2010.
[16] T. Gomez et. al., “Ancillary Services Markets in California,” LBNL
Report-43986, 1999.
[17] J. B. Greenblatt, S. Succar, D. C. Denkenberger, R. H. Williams, and
R. H. Socolow, “Baseload wind energy: modeling the competition between gas turbines and compressed air energy storage for supplemental
generation,” Energy Policy, vol. 35, pp. 1474-1492, Mar 2007.
[18] A. Kusiak and W. Y. Li, “Short-term Prediction of Wind power with a
Clustering Approach,” Renewable Energy, vol. 35, pp. 2362-2369, 2010.
[19] M. Korpaas, A. T. Holen, and R. Hildrum, “Operation and sizing of
energy storage for wind power plants in a market system,” International
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
Journal of Electrical Power and Energy Systems, vol. 25, no. 8, pp.
599-606, Oct. 2003.
H. Lund and G. Salgi, “The role of compressed air energy storage
(CAES) in future sustainable energy systems,” Energy Conversion and
Management 50 (2009) 11721179.
J. Matevosyan and L. Sajder, “Minimization of imbalance cost trading
wind power on the short-term power market,” IEEE Trans. Power Syst.,
vol. 21, pp. 1396-1404, 2006.
Y. Makarov, C. Loutan, J. Ma, and P. de Mello, “Operational impacts
of wind generation on California power systems,” IEEE Transactions on
Power Systems 2009; 24: 1039-1050.
J.M. Morales, A.J. Conejo, J. Perez-Ruiz, “Short-term trading for a wind
power producer,” IEEE Trans Power Syst; in press.
P. Pinson, C. Chevallier, and G. N. Kariniotakis, “Trading wind generation from short-term probabilistic forecasts of wind power,” IEEE Trans.
Power Syst., vol. 22, pp. 1148-1156, 2007.
N. Petruzzi, M. Dada, “Pricing and the newsvendor problem: a review
with extensions,” Operations Research. 47.183-94, 1999.
P. Pinson, H. A. Nielsen,H. Madsen, and K. Kariniotakis, “Skill Forecasting from Ensemble Predictions of Wind Power,” Applied Energy,
vol. 86, pp. 1326-1334, 2009.
R. T. Rockafellar, S. Uryasev, and M. Zarabankin, “Deviation Measures
in Risk Analysis and Optimization,” Research Report 2002-7, Department of Industrial and Systems Engineering, University of Florida,
Gainesville, FL, 2002.
S. Sethi, H. Yan, J. Yan, and H. Zhang, “An Analysis of Staged
Purchases in Deregulated Time-Sequential Electricity Markets,” Journal
of Industrial and Management Optimization 2005, 1, 443463.
A.S. Siddiqui, C. Marnay, and M. Khavkin, “Excessive Price Volatility
in the California Ancillary Services Markets: Causes, Effects, and
Solutions,” LBNL Report-46083, May 2000.
“Market Operator of the Electricity Market of the Iberian Peninsula,”
OMEL 2008. [Online]. Available: http://www.omel.es.
Strbac, “Power System Economics”
G.C. Thomann, M.J. Barfield, “The time variation of wind speeds and
windfarm output in kansas,” IEEE Transactions on Energy Conversion
1988.
P.P. Varaiya, F. Wu, J.W. Bialek, “Smart Operation of Smart Grid: Risklimiting Dispatch,” submitted to Proceedings of the IEEE, 2010.
“Pilot Project Conducted by Iberdrola Renewables, BPA and Constellation Energy Collaboration Seeks to Put More Clean Energy on the
Grid,” http://www.windtoday.net, Sept. 22, 2010.