This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON POWER SYSTEMS 1

A Frequency-Constrained Stochastic
Economic Dispatch Model
Yen-Yu Lee, Member, IEEE, and Ross Baldick, Fellow, IEEE

Abstract—In this paper, we formulate the economic dispatch small system inertia [1], [2]. In such systems, the levels of in-
problem as a two-stage stochastic convex program for operational ertia and spinning reserve requirements are set to ensure that the
decision making under uncertainty. Post-contingency minimum
frequency deviations are limited subsequent to the loss of the
frequency constraints are formulated as convex sets using a
simplified model of frequency dynamics. The L-shaped method largest unit [1], [3]. With increasing levels of wind power pen-
is used to decompose the large stochastic program into smaller etration, frequency excursions might also be problematic even
sub-problems. Large-scale system examples are provided to in large-scale isolated systems such as the Electric Reliability
demonstrate the computational efficiency of the proposed model Council of Texas (ERCOT) system, since wind resources pro-
formulation. Numerical results show the value of the proposed
model and indicate that the proposed model is especially useful vide only limited inertia and frequency control. Reference [4]
under a higher level of wind penetration. also reports that the quality of frequency control in the U.S. has
been declining in recent years and suggests that careful planning
Index Terms—Economic dispatch, operating reserves, power
system security, primary frequency control, stochastic program- of primary frequency control is necessary.
ming, transmission constraints, wind integration. Several models reported in the literature attempt to dispatch
generation and allocate reserves under uncertainty. Takriti
and Birge [5] describe a multi-stage stochastic program to
I. INTRODUCTION solve a unit commitment problem with uncertain demand, but
without transmission constraints. Bouffard et al. [6] propose

M AINTAINING power system reliability has always

been a challenging task due to the possible failure of
system components and uncertainty in demand. The integration
a stochastic market-clearing model considering transmission
constraints. Their model deals with equipment failures and uses
the expected value of unserved energy as the security metric.
of wind poses more difficulties for system operation due to its Wu et al. [7] propose a long-term security-constrained sto-
uncertain and variable nature. In most power systems, system chastic unit commitment model considering fuel and emission
reliability is maintained through procuring operating reserves constraints. Ruiz et al. [8] propose a two-stage stochastic unit
from generation resources, with the reserve requirements commitment model that incorporates reserve requirements.
usually set to ensure that, for example, occurrence of any Hedman et al. [9] propose a transmission switching and unit
contingency in a pre-defined list does not cause any involuntary commitment model to ensure security. Kimball et al.
load shedding. In other words, the scheduling of reserves in [10] propose a stochastic optimal power flow model and solve it
most power systems today is a deterministic process. by Benders’ decomposition. They demonstrate that a stochastic
The current approach of imposing fixed reserve requirements model can indeed reduce the expected operation costs. Saric
in an economic dispatch model has several disadvantages. With et al. [11] formulate the single period market clearing process
large-scale integration of wind power, the appropriate reserve using a two-stage stochastic program and show that the model
requirements are no longer obvious. Also, the requirements usu- is computationally tractable using an interior-point algorithm
ally ignore possible transmission congestion when reserves are together with Benders’ decomposition. Note that none of the
deployed. models reported in these references represent the response of
In addition to post-contingency transmission constraints, an- system frequency.
other important issue in reserve scheduling is the capability of Several papers in the literature attempt to incorporate
primary frequency control, the autonomous response by genera- frequency response into the dispatch problem. Somuah and
tion resources to the deviation of frequency from nominal. This Schweppe [12] develop an economic dispatch model with
issue is more significant in small isolated power systems, such a constraint on the maximum frequency deviation given a
as the systems in Ireland and New Zealand, because of their specified disturbance. The frequency constraint incorporates
turbine dynamics, which makes the problem non-convex.1
Manuscript received February 13, 2012; revised July 04, 2012 and October Restrepo and Galiana [13] develop a unit commitment model
02, 2012; accepted November 11, 2012. This work was supported by the Electric with requirements for primary frequency control. They char-
Reliability Council of Texas. Paper no. TPWRS-00132-2012.
The authors are with the Department of Electrical and Computer Engi- acterize the amount of primary frequency control a unit could
neering, The University of Texas at Austin, Austin, TX 78712 USA (e-mail: provide using the steady-state frequency deviation, without
[email protected]; [email protected]).
Color versions of one or more of the figures in this paper are available online 1Although the authors of [12] claim the frequency deviation function is
at http://ieeexplore.ieee.org. convex, no rigorous proof is provided. In fact, the co-existence of the fifth and
Digital Object Identifier 10.1109/TPWRS.2012.2236108 sixth term in (13) in [12] makes the function non-convex.

0885-8950/$31.00 © 2012 IEEE

 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON POWER SYSTEMS

consideration of under-frequency load shedding (UFLS); that

is, they impose frequency constraints as hard constraints,
without allowing any constraint violation. Doherty et al. [14]
develop a market-clearing model with linear frequency control
constraints. They simulated a large number of events using
a simplified dynamic model to construct a feasible region in
terms of system kinetic energy, reserve procurement, and size
of contingency. Ruiz and Sauer [15] describe a simplified
dynamic model to estimate the amount of UFLS and construct
a demand curve for reserves. The models in both [14] and Fig. 1. Timeline of post-contingency operation represented by the second-stage
[15] aggregate the primary frequency control of each unit into problem.
a single quantity. That is, each unit’s impact on the system
frequency, depending on its response speed and operating point,
is not well differentiated. In SED, a contingency scenario consists of one or more
This paper proposes a new model, the stochastic economic generator failures, together with significant demand forecast or
dispatch model (SED), to find the optimal schedules for energy wind forecast errors. Note that in the second-stage, all unknown
and reserves that minimize expected operation costs (including information in the first-stage is revealed. That is, the avail-
penalties for load shedding). Our main contributions are sum- ability of each generator and the exact value of demand and
marized as follows: wind power output are known. We assume that any generation
• We propose the idea of incorporating minimum frequency outages occur at the beginning of the operational time frame.
constraints in a stochastic optimization framework, so that The operational time frame is divided into two non-overlapping
primary frequency control and geographical allocation of periods, as shown in Fig. 1. These periods characterize the
reserves can be co-optimized to minimize the expected op- typical process of restoring the system back to a normal state
eration costs. after the disturbance:
• We propose a new formulation for frequency constraints Frequency stabilizing period (FSP): Over a duration of
and a method for estimating the unserved demand resulting seconds, system frequency is stabilized by primary frequency
from UFLS. To improve the accuracy of the frequency con- responses of generators together with UFLS, if necessary.
Transmission constraints are ignored in this period since the
straints, we characterize each individual generator’s im-
main operational goal here is to stabilize frequency; also, a
pact on the system frequency based on its maximum ramp
short period of violation of transmission constraints usually
rate and “headroom,” i.e., available capacity left for in-
does not threat system reliability. The FSP is analogous to the
creasing output. Moreover, the proposed formulation is
“emergency state” in [17], where the control objective is to
convex and computationally efficient.
relieve system distress and prevent further degradation.
• We apply the L-shaped decomposition method [16] and
Re-dispatch period (RDP): After the FSP, the system operator
provide numerical results based on large-scale systems.
re-dispatches generation to restore the system security. Involun-
Computational efficiency of the formulation is demon- tary load shedding is applied if generators alone are insufficient
strated. We also show that the proposed model improves to restore system security. We suppose that spinning reserve
the dispatch in terms of expected operation costs, com- and fast-start reserves are ready to be deployed in this period.2
pared to deterministic methods. At the end of the RDP, non-spinning reserves can be brought
The rest of this paper is organized as follows: Section II in- online to meet the minimum spinning reserve requirement. The
troduces the formulation of SED. Section III describes the new typical duration of the RDP is from 15 to 30 minutes [18]. This
method of estimating UFLS. Section IV discusses the decompo- period is analogous to the “restorative state” in [17], where
sition algorithm. Section V provides numerical examples, and the control objective is the safe transition from partial to full
Section VI concludes. satisfaction of all security constraints.
Unlike other stochastic dispatch models in the literature [10],
II. STOCHASTIC ECONOMIC DISPATCH MODEL [11], which use the post-contingency “steady-state” equilibrium
to characterize the post-contingency operation, SED represents
A. Introduction the process of disturbance recovery by incorporating frequency
dynamics after the disturbance. Fig. 1 shows the timeline for
The proposed model, the stochastic economic dispatch model the post-contingency operation described above. It illustrates
(SED), is formulated as a two-stage stochastic convex program. the timing of enforcing different operational constraints and the
The first-stage decisions represent pre-contingency generation availability of each type of reserve.
dispatch and procurements of reserves. After a specific realiza-
tion of system conditions is revealed, the second-stage decisions B. Notation
of generation re-dispatch are then made to optimally respond to
This section summarizes the notation used in the develop-
this specific scenario. The objective of SED is to minimize the
ment of the model. To simplify the presentation of the model
expected operation costs, subject to both pre- and post-contin-
gency operation constraints over an operational time frame. 2Spinning reserves are usually available prior to the beginning of the RDP.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LEE AND BALDICK: A FREQUENCY-CONSTRAINED STOCHASTIC ECONOMIC DISPATCH MODEL 3

but without loss of generality, we assume that each bus has ex- fraction of UFLS deployed at bus ;
actly one generator and one load.
maximum wind power output at bus [MW].
Indices and Sets:
First-stage decision variables:
buses;
generation dispatch of generator at bus [MW];
subset of buses with wind power generation;
spinning reserve from generator at bus [MW];
subset of buses with generators equipped with fast-start reserve from generator at bus [MW];
governors;
non-spinning reserve from generator at bus
subset of buses with committed generators3; [MW].
subset of buses with un-committed generators; Second-stage decision variables, under scenario :
subset of buses with available generators in generation re-dispatch at bus during the RDP
scenario ; [MW];
subset of buses with generation outage in scenario load shedding triggered in the FSP [MW];
;
load shedding triggered in the RDP at bus [MW];
transmission security constraints;
restored spinning reserve from generator at bus
contingency scenarios. [MW].
Random parameters, under scenario : Functions:
load at bus at the end of the RDP [MW]; energy cost incurred in operational time frame [$];
maximum wind power output at bus at the end reserve cost incurred in operational time frame
of the RDP [MW]; [$];
availability of generator at bus : 1 for normal second-stage post-contingency cost in operational
operation, 0 for outage. time frame [$] (see Sections II-C and III-B);
Parameters: approximated UFLS under scenario [MW];
probability of contingency scenario ; kinetic energy shortfall to meet the frequency
threshold [MWs].
probability that no contingencies occur;
length of frequency stabilizing period [s]; C. Stochastic Economic Dispatch Model
duration of involuntary load shedding until Let be the vector collecting together the dispatch
restored [h]; variables . Similarly, is the collection of all
reserves variables, , and . Function , the
load forecast at bus [MW];
post-contingency cost, will be introduced in Sections II-D and
limit of transmission constraint [MW]; III. SED is formulated as
shift factors of injection at bus on constraint ;
(1a)
commitment status of generator at bus : 1 is
online, 0 is not; (1b)
for generator at bus : 1 is fast start capable, 0 is
not; (1c)
inertia constant4 for generator at bus [s];
(1d)
maximum output of generator at bus [MW];
(1e)
minimum output of generator at bus [MW];
(1f)
spinning reserve limit of generator at bus [MW];
(1g)
upper ramp rate limit of generator at bus [MW/s]; (1h)
penalty cost for unserved load [$/MWh]; (1i)
post-contingency spinning reserves requirement
The objective of SED (1a) is to minimize the expected operation
[MW];
costs. The pre-contingency operation costs consist of energy
3Sets and do not include wind generation. 4 stored kinetic energy at nominal frequency machine MVA rating
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON POWER SYSTEMS

production costs and reserve procurement costs. Note that en- (2d)
ergy production costs are weighted by , the probability of no
occurrence of contingency. The post-contingency costs , the (2e)
key innovation of SED, includes the penalty costs for unserved
(2f)
load and re-dispatch energy costs. We assume that function
is convex quadratic and that is linear.5 Constraint (1b) is (2g)
the power balance constraint that ensures supply meets demand. (2h)
Constraints (1c) are transmission constraints that limit the flow (2i)
on each transmission line, where represents the amount of
power flowing on line given a unit injection at bus and with-
(2j)
drawal at the reference bus. We use DC power flow here, where
can be derived from linearization about the real-time op- (2k)
erating point [19]. Note that transmission security con- (2l)
straints are imposed to ensure that there will be no line-flow vio-
lations if any single line is removed from the system. Constraints Note that we ignore the start-up and min-load costs of deploying
(1d) specify the upper limits of wind power output, where the fast-start and non-spinning reserves. The parameter represents
values are typically obtained from forecast. Constraints (1e) the penalty cost of involuntary load shedding. Constraint (2b)
and (1f) place lower and upper limits, respectively, on power places a lower limit on the unserved load due to UFLS. After
generation for each generator. Constraints (1g) are maximum the FSP, the transmission constraints (2d) are enforced to pre-
procurable spinning reserve constraints, which are typically de- vent over-loaded lines and restore transmission security. Note
termined by the ramp rate limit. Constraints (1h) specify that that the portion of load that is disconnected in the FSP is con-
only fast-start capable generators can provide fast-start reserves, sidered in constraints (2c) and (2d). Constraint (2e) place upper
while constraints (1i) indicate that offline capacity can provide limits, which are random parameters, on wind power generation.
either fast-start or non-spinning reserve, but not both. Note that Constraints (2f)–(2i) ensure that spinning reserve is restored to
the primary frequency control at bus is implicitly determined the minimum level to meet the reliability standard ([18], R6),
by the generation dispatch , which decides the amount of assuming that non-spinning reserves are on-line at the end of
available capacity left for increasing output in the contingency the RDP. Constraints (2j) and (2k) enforce limits on power gen-
state. We do not model regulation reserves here, but they can eration for each generator in the RDP. Note that we do not en-
also be incorporated. force minimum generation limit on fast-start generators, since
their commitment decisions in the second-stage are not explic-
D. Second-Stage Problem itly modeled. In addition, we assume that non-spinning reserves
are only deployable at the end of the RDP for restoring re-
The second-stage problem models post-contingency system
serves. That is, they cannot provide energy during the RDP.6
operations. The amount of unserved demand due to UFLS in
Constraints (2l) limit load shedding at each bus to the total load
the FSP is estimated and penalized. Directly after the FSP, the
at that bus.
system operator re-dispatches generation in order to restore
The key innovation of SED is to formulate as a convex
system security. The objective in the RDP is to minimize the
function. According to [22, proposition 2.1 and 2.3], is a
generation costs and the penalty costs of involuntary load
convex function in if is a convex function. The
shedding. In practice, the system operator might run a sequence
convexity of then ensures that SED is a convex optimization
of transmission-constrained economic dispatches as load varies
problem.
with time in the RDP. Because of computational limitation,
only a single re-dispatch is modeled here, which is based on III. ESTIMATING UNSERVED LOAD DUE TO UFLS
the assumed known load and wind realization at the end of the
RDP. Let and be the vector collecting the This section describes a method to estimate the amount of
dispatch variables and , respectively. Function , under-frequency load shedding (UFLS) that must be applied,
which is developed in Section III, represents the amount of given minimum frequency requirements and incorporates that
unserved load resulting from UFLS. The second-stage problem estimate into the second-stage problem (2). The estimated in-
is formulated as follows: voluntary load shedding is formulated as a convex func-
tion of the generation dispatch vector . Note that convexity
is a crucial property for applying decomposition methods and
ensuring global optimality. We also provide analytical expres-
(2a) sions for evaluating the subgradients of , in order to ef-
ficiently generate linear approximations for the decomposition
(2b) algorithm.
Directly after a disturbance, the frequency of the system de-
(2c)
clines and generators with governors respond by increasing me-
chanical power output in order to stabilize system frequency.
5Typical generation costs can be well approximated by convex quadratic
functions ([19], 2), while reserve costs are modeled as linear functions in 6We therefore do not include the start-up process of non-spinning reserves
most electricity markets ([20], 4.4.7) ([21], 4.2). and thus ignore their minimum generation limits.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LEE AND BALDICK: A FREQUENCY-CONSTRAINED STOCHASTIC ECONOMIC DISPATCH MODEL 5

If primary frequency responses from generators alone cannot is the time when generator reaches its maximum output limit
maintain the frequency above a threshold, UFLS will be trig- . For the convenience of analysis, we assume that total gener-
gered to protect the system from further damage. We assume ation mechanical output will never exceed the load.8 Now, as-
that following a disturbance, all generators with headroom ramp sume that total load is constant during the FSP and that pre-con-
their mechanical power production at the maximum ramp rate tingency dispatch satisfies power balance constraint (1b). The
limits. This is a standard assumption in simplified frequency dy- increased mechanical power output from all available genera-
namic models [12], [15]. tors at time would be a function of :

A. Evaluating System Frequency

We can estimate system frequency at an instant in time by (7)
evaluating the stored kinetic energy after the disturbance. Recall
that the stored kinetic energy of rotating mass is

where is either
• the time when mechanical power output matches the load,
where is the system inertia and is the system frequency if this occurs within the duration of the FSP ; or
in Hz. Since is non-negative, we can express frequency as • , if mechanical power output fails to meet the load
a function of : within .
That is
(3)

Because is a monotonic function of , enforcing the minimum

frequency requirement would be equivalent to enforcing a lower
bound on stored kinetic energy. Assuming that the system is at (8)
the nominal frequency before the disturbance and no UFLS is
applied, we obtain the frequency constraints Given the mechanical power output function can be
easily computed. Note that without any UFLS, the system fre-
(4)
quency reaches its nadir at .
where The change in stored kinetic energy at time is the integral of
• is the change of the stored kinetic energy at time power mismatch between generation and load up to time , that
from the pre-contingency state, given pre-contingency is
dispatch ;
• , the post-contingency system kinetic en- (9)
ergy at nominal frequency,7 that is, the system kinetic en-
ergy immediately after the disturbance but before the fre-
Let be the set of generators having binding ca-
quency drops; and
pacity constraints before , that is
• , the system kinetic energy at the
low frequency limit .
The quantity can be estimated by the energy
imbalance between load and generation up to time due
and let be the complement set of .
to the disturbances. Under a contingency scenario , let
Combining (5) with (9), we obtain
be the set of buses with generation outage
and let be the set of buses with available
generators (in normal operation) having governors. Suppose
that disturbance occurs at . According to our assumption
on ramping of available generators, the increase in mechan-
ical power output after the distrubance from generator at bus
is
(5)

where is the ramp rate limit and

(10)
(6)

7We assume that the machine MVA rating is equal to the maximum power 8Although, in practice, generation mechanical output will exceed the load to
output . We can also use estimates of system inertia constant using historical restore frequency after the frequency nadir, this assumption has no impact on
data [23], which includes inertia of load. the estimation of amount of unserved energy in the adopted UFLS model.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON POWER SYSTEMS

Note that is negative and non-increasing in . Also, it or evaluated during the process of solving SED, as will be dis-
reaches the minimum at time , that is cussed in Section IV.
Proposition 1: is a convex function in .
Proof: In (5), since is non-decreasing, can
(11) be expressed as a pointwise minimum function:
Therefore, enforcing minimum frequency constraints at time
ensures that (4) is satisfied, that is
This function is concave in . Therefore,
(12)
is also concave in . By the same reasoning of pointwise min-
imum, is also concave in . By (7), we know that
B. Estimating the Amount of UFLS
In the case that governor responses alone cannot suffice to sat-
isfy the frequency constraint (12), UFLS needs to be deployed
to maintain system frequency above the threshold . One ap- Therefore, can also be expressed as
proach to approximate the amount of UFLS [15] uses the short-
fall in stored kinetic energy to meet the frequency threshold, that
is
(13)
Since for any is a concave function in ,
Under this scheme, load shedding ends immediately when the
the integral over of is also concave. The maximum
mechanical power output meets the load. However, the load
operator in (13) implies that is a convex function in .
shed will remain disconnected from several minutes to several
Since in (16) is non-negative, is therefore convex.
hours in practice. Therefore, we must estimate the amount of
Some governors do not respond to decline in system fre-
power shed that maintains the system frequency above the limit,
quency immediately, but with some response delay. This ini-
and model the duration of the power shed.
tial response delay can be incorporated in our model. See [24,
In practice, there may be multiple frequency thresholds trig-
4.3.2] for the detailed discussion.
gering different sets of under-frequency relays. There are also
As we will discuss in Section IV, the L-shaped method (al-
types of UFLS that are triggered when the frequency remains
gorithm) is applied to solve SED. The method requires repeated
below a certain threshold for a specified time interval, or trig-
evaluation of the subgradient of . Appendix B develops an
gered when the rate of change in frequency is sufficiently large.
analytical expression for that subgradient based on one partic-
We therefore define the “theoretically optimal UFLS,” ,
ular directional derivative. This analytical expression can sig-
to approximate the actual UFLS. Note that we assume a single
nificantly improve computational efficiency.
frequency threshold . This is a widely used assumption for
estimating under-frequency load shedding [15]. We assume that IV. DECOMPOSITION ALGORITHM
UFLS is triggered at , the time when the system frequency
reaches the low frequency limit. That is For large-scale stochastic program, decomposing the orig-
inal problem into smaller sub-problems can significantly im-
(14) prove computational efficiency. Slyke and Wets [16] develop
the L-shaped method to decompose a stochastic linear program
The theoretically optimal UFLS is defined as the power imbal- by exploiting its special structure; see also Benders [25]. Birge
ance between load and mechanical power output at : and Tang generalize this method to stochastic convex programs
[26]. Since is convex, the optimal value of the second-
(15)
stage problem is a convex functions of [22]. There-
fore, the L-shaped method can be applied to solve SED. The
where represents the minimum amount of load shedding main idea of the L-shaped method is to represent sub-problems
needed to ensure the frequency stays above the low frequency with their outer linearizations in (1), so that can be
limit. Although can be computed efficiently given a dispatch solved independently for each scenario. The decomposition al-
, the function is non-convex. However, from numerical gorithm is described below as Algorithm 1.9
experiments, we find that the mapping between and is
almost linear for scenarios having similar sizes of disturbance. Algorithm 1 Solving SED Model
Since is convex (shown in Proposition 1 below), we use
to approximate and ensure a convex approximation of the Input: Network data, optimality tolerance .
problem, that is Output: -optimal and the associated cost .
(16)
.
where is the approximated power shed. We discuss the ratio- 9Note that we assume the second-stage problem (2) is feasible for every fea-
nale of such approximation in Appendix A. Non-negative coef- sible first-stage decisions. This can be easily achieved by adding a artificial vari-
ficients and can be either obtained by ex ante simulation, able, with a large objective function penalty, to constraint (2i).
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LEE AND BALDICK: A FREQUENCY-CONSTRAINED STOCHASTIC ECONOMIC DISPATCH MODEL 7

loop and that yield good convex approximation of the function

. Note that we only update and once in the algorithm.
Solve the relaxed master program: Therefore, after the update, the algorithm becomes a standard
decomposition algorithm for a convex non-linear program, for
(17a)
which convergence is guaranteed [27], [28].
In the L-shaped method, optimality cuts (17b) should form
lower bounds for . However, updating the coefficients for
scenario changes the function . As a result, existing opti-
mality cuts associated with might no longer yield valid lower
bounds on and thus should be removed. However, under
(17b) certain conditions, as described in Appendix C, the existing op-
timality cuts can be preserved. Some computational effort can
Let be the optimal solution of the relaxed master be saved by keeping the existing cuts since they still provide
program and let be the corresponding optimal value. useful information about the updated second-stage problem. In
For all , compute and using (13) the context of solving successive SEDs for consecutive dis-
and (16), respectively. Form the right-hand sides of (2) patch intervals, the updated coefficients can serve as good initial
and solve . Then compute the associated cost values for solving the SED for the subsequent dispatch interval.
using (1a). To simplify the presentation of the L-shaped method, we do not
if then include this procedure of updating and in Algorithm 1.
Since the cost of involuntary load shedding is dominant in
(2a) [29], we may ignore the difference in pre- and post-con-
tingency energy-production costs and the second-stage energy
costs are aggregated with the first-stage energy costs. The ap-
end if proximated objective function of SED (1a) becomes
if then
(18)
Stop: is the -optimal solution.
end if This approximation simplifies the second-stage problems (2)
For each , compute using (31) in Appendix B. from quadratic programs into linear programs, which signifi-
Augment the set of optimality cuts in (17b) with cut cantly reduces the computational burden for large problems.
coefficients:
V. NUMERICAL EXAMPLES
The ERCOT network and load data for summer 2011 together
with generic costs10 and pre-defined contingency scenarios are
where , and are the optimal dual variables
used to compare SED and conventional economic dispatch. The
of constraints (2b), (2f), and (2j), respectively. The cut
minimum frequency is set as 59.5 Hz,11 with the nominal fre-
intercept is
quency of 60 Hz. The total load in the system is 59.60 GW, on-
line generation capacity (without wind capacity) is 59.68 GW,
and the forecasted wind output is 4431 MW. The penalties are
set $30 000/MWh with load-shedding duration of one half
hour.12 In this example, we focus on the benefits of modeling
frequency control and geographical allocation of spinning re-
serves and thus do not include fast-start and non-spinning re-
end loop serves. Also, besides 7186 base-case transmission constraints,
we add 10 binding and near-binding transmission security
We need to set the coefficients and in (16) to some constraints, obtained from solving security-constrained optimal
initial values before the algorithm starts. A reasonable initial power flow (SCOPF) in PowerWorld with the given power flow
value for is between 0 to 1, where the unit are , since and generic cost data.
the value of is usually smaller than the value of according We use historical wind data in the ERCOT during 2008 and
to numerical experiments. A good initial guess for is 0, be- 2009 to estimate the probabilities and sizes of significant wind
cause there would be no load-shedding with of 0. During the 10For instance, for a 600-MW coal-fired generation unit, the marginal cost
first several iterations, in addition to evaluating , we compute curve, represented as a linear function, ranges from $24–30/MWh. As another
using (15) to obtain best estimates for and by linear example, for a 400-MW gas-fired steam turbine, the marginal cost curve, also
represented as a linear function, ranges from $63–71/MWh.
regression. From numerical experiments, we observe that the 11The highest frequency thresholds in North America range from 59.3 to 59.7
correlation between and tends to be strong under similar Hz [4].
sizes of contingency (see Fig. 7 in Appendix A); thus a moderate 12The parameter roughly corresponds to the reliability standard “one day in
number of samples of and is likely to provide values for ten years” [30].
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON POWER SYSTEMS

TABLE I to , depending on types of generators.13 Each con-

ESTIMATED CONDITIONAL PROBABILITY OF SIGNIFICANT WIND RAMP EVENT tingency scenario consists of a single generation contingency
or a realization of significant wind ramp, or both. We assume
that wind events and generation outages are independent; thus
the computation of is straightforward given the probabil-
ities of generation contingency and wind ramp. Note that the
largest generation contingency is the loss of two co-located nu-
clear units, totaling 2443 MW of capacity. To make the number
of scenarios manageable, only generation contingencies greater
than 900 MW in capacity are considered. Furthermore, contin-
gency scenarios are only incorporated if the associated probabil-
ities are greater than . Once the set of scenarios is deter-
mined, we approximate , the probability that no contingency
occurs, as . There are a total of 33 contingency
scenarios after applying this scenario reduction rule.
1) Impact of System Inertia: We use different system inertia
constants, ranging from 3–5 s, to simulate the impact of inertia
on system operations. In particular, the smaller system inertia
Fig. 2. Expected cost in the ERCOT example versus inertia. corresponds to the case with high levels of wind power integra-
tion. In addition to SED, we solve two different type of models
for comparison:
• the ED models with spinning reserve requirements of 2443
MW (size of the largest generation contingency, labelled
“ED-2443 MW” in Figs. 2 and 3), 2650 MW (“ED-2650
MW”), 2850 MW (“ED-2850 MW”); and
• an SED model without enforcing UFLS constraint (2b)
(“SED w/o UFLS”).
We use Monte Carlo simulation with a sample size of 10 000
to evaluate the expected operation costs. Due to computational
limitations, we again use (18) to approximate the expected op-
eration costs. To further increase computational efficiency, we
only solve the second-stage problems for contingency scenarios
Fig. 3. Expected UFLS in the ERCOT example versus inertia.
with size greater than 600 MW, because contingencies smaller
than 600 MW are unlikely to cause significant involuntary load
ramp events in the operational time frame of a half hour. (Note shedding. Fig. 2 shows that SED results in the lowest expected
that in real-time operation, we can utilize a wind ramp fore- costs for each inertia constant. In particular, the cost of SED
casting tool to obtain more accurate estimates of the likelihood with the inertia constant of 3s are about 1.15% lower than “ED-
of large wind events [31].) Given a sequence of wind power 2443 MW,” and 1.0% lower than “SED w/o UFLS.” Because
output data sampled with fixed time interval , the SED model without enforcing UFLS constraints results in
and assume samples are taken in 30 min. Define set similar costs compared to the ED models, the cost savings are
and the set . apparently due to the lower expected UFLS by incorporating
The conditional probability of a wind ramp event is estimated UFLS constraints. Fig. 3 shows that SED significantly improves
as system frequency response for low values of system inertia.
These results suggests that SED, which incorporates the effects
ramp size wind power output of UFLS, becomes more useful as the levels of wind power pen-
etration increases.
where the size of the ramp is represented by the mean of sam- 2) Impact of Increased Level of Wind Integration: We sim-
ples in . Table I summarizes the derived wind ramp sizes and ulate the impact of expansion in installed wind capacity by in-
probabilities conditioned on the forecasted wind output. Note creasing the size of wind ramps by 25%: the largest wind ramp
that the decreased wind output is distributed to each wind farm, is now 2848 MW. This corresponds to the estimate reported in
which is treated as a generator, proportional to its forecasted [33], which predicts that, in the future, the largest wind ramp-
output. There are 87 online wind farms in the power flow data. down will exceed 2800 MW with 15 000 MW of installed wind
We assume that the time of failure of a generator is a random capacity in the ERCOT region. We scale the probabilities of
variable with exponential distribution [6], [8], where the mean wind ramps in Table I to simulate the system under different risk
time to failure of each generator is estimated using the gener- of significant wind events. Assume the system inertia constant
ating availability data system (GADS) provided by NERC [32]. 13For example, the estimated outage probabilities are for diesel
The probability of a single generation contingency during the generators, for coal plants, for combined cycle plants,
operational time frame of one half hour ranges from and for nuclear plants.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LEE AND BALDICK: A FREQUENCY-CONSTRAINED STOCHASTIC ECONOMIC DISPATCH MODEL 9

Fig. 6. Spinning reserves procured in the ERCOT example.

Fig. 4. Expected cost in the ERCOT example with increased wind penetration.

Fig. 7. versus for the ERCOT System.

Fig. 5. Expected load shedding during RDP in the ERCOT example.

includes the time for loading network data and creating opti-
is [4]. Three ED models with different reserve requirements mization models, which takes 132 s on average. In the case of
are solved: 2443 MW (size of the largest generation contin- solving the model repeatedly with different system conditions,
gency, labelled “ED-2443 MW” in Figs. 4–6), 2650 MW (“ED- the created optimization models can be re-used after modifica-
2650 MW”), and 2848 MW (size of the largest wind ramp, la- tion. Also, since the subproblem for each scenario (2) can be
belled “ED-2848 MW”). Fig. 4 shows that the cost savings using solved independently, parallel computing could be used for ap-
stochastic approach increase with higher risk of wind events. plying SED in real-time operation. For example, if sub-prob-
Again, we use Monte Carlo simulation with a sample size of lems are distributed to five computing resources, we may be able
10 000 to evaluate the expected operation costs. The cost sav- to solve this problem instance within one minute.
ings range from 0.64% to 4.13%, compared to “ED-2443 MW”.
On the other hand, the cost savings range from 0.83% to 2.08%, VI. CONCLUSION
compared to “ED-2848 MW”. Interestingly, SED produces a In this paper, a new power system operation model is pro-
much more secure dispatch under high risk of wind ramps in posed. This model, formulated as a two-stage stochastic convex
terms of expected load shedding during the RDP, as shown in program, searches for the optimal energy and reserve schedules
Fig. 5. Note that the total amount of spinning reserves of the considering important operational constraints. In particular, we
SED solution is close to 2848 MW when the probability of the propose a convex formulation for frequency constraints which
largest wind ramp exceeds , as shown in Fig. 6. How- incorporate under-frequency load shedding. Numerical results
ever, the expected load shedding during the RDP is much lower imply that SED generates energy and reserves schedules that
using stochastic solutions. This result is apparently due to the significantly reduce the expected operation costs, compared to
better geographical allocation of reserves of the stochastic so- the deterministic reserve requirements. The results also suggest
lutions. This numerical experiment suggests that geographical that the cost saving using SED becomes more significant under
allocation of reserves should be incorporated to enhance reli- higher levels of wind integration.
ability. The results also implies that the stochastic approach is In future work, we plan to incorporate transmission line
more useful as the penetration levels of wind increase. As sto- outages into second-stage problems. We currently enforce
chastic variability increases for supply, deterministic solutions traditional transmission security constraints in the
compare more and more unfavorably with stochastic solutions. first-stage. These constraints ensure that, even without any gen-
3) Computational Costs: With an optimality tolerance of eration re-dispatch, no transmission limits will be violated if any
0.05%, the average CPU time for solving SED is 278.8 s,14 with single line is removed from the system. It may be beneficial to
27 outer iterations on average. Note that the reported CPU time incorporate the transmission line outages into the second-stage
14We used a laptop computer on Windows 7 with 2.0-GHz Intel Core 2 Dual problems so that the flexibility of re-dispatching generators
CPU and 3 GB of RAM. can be exploited. This flexibility may substantially lower the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON POWER SYSTEMS

pre-contingency costs. However, the typically large number Fig. 7 shows the scatter plot of and , obtained by solving
of possible line outages would pose significant computational SED with the same data set in Section V-A (with inertia con-
challenges and thus only a limited number of contingency stant of 3.5 s). We collect contingencies with similar sizes into
scenarios can be incorporated. In principle, we should choose a group, and each type of marker in the figure represents such
contingency scenarios based on their probability-weighted a group. Interestingly, the figures shows that is roughly pro-
costs. The development of such a method would be an impor- portional to , which is consistent with (25). This figure also
tant and interesting research topic for enhancing the efficiency indicates that is approximately linear in for each group
of SED. of contingencies. This finding suggests that we can reasonably
approximate using for ensuring the convexity of SED, as
described by the (16) in Section III.
APPENDIX A
RATIONALE OF CONVEX APPROXIMATION OF UFLS
We use a single generator model to explain why (16) forms a APPENDIX B
reasonable approximation to . Assume that this generator DERIVATION OF SUBGRADIENT OF
has a pre-contingency dispatch , a maximum output limit , We start with deriving a subgradient of , which is the key
and a maximum ramp rate . Let the size of disturbance be , part of the whole derivation. By (10) and (11), we have
where . We assume the generator’s headroom
is greater than the size of disturbance, that is, . We
also assume ; therefore, . We also suppose that
UFLS is triggered. By (10), we have

(19)
Also, by (11), we obtain Assume that sets and remain unchanged under
a small perturbation of . By partially differentiating
(20) with respect to , we obtain

Substituting (20) into (13) gives

(21)
(26)
Moreover, according to (15), is the power imbalance at time
:
(22) By (7) and (8), we know that

Substitute (19) into (14). Then satisfies the following equation:

By solving the above equation, we obtain15 under the case that mechanical power output meets the load.
Otherwise, according to (8), which implies .
(23) Therefore

Substituting (21) into the above equation gives

(24)
Together with (6), (26) is reduced to
Combining with (22), we obtain
(27)
(25)
Similarly, for is
For a given scenario , if the associated kinetic energy shortfall
concentrates in a particular range, then (25) can be well-
approximated by the linear function (16).
15 is the smaller root since UFLS is triggered once the frequency reaches the
low limit. (28)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LEE AND BALDICK: A FREQUENCY-CONSTRAINED STOCHASTIC ECONOMIC DISPATCH MODEL 11

and for is ACKNOWLEDGMENT

The authors would like to thank ERCOT for providing wind
data and D. Lee of the University of Texas at Austin for com-
piling the historical wind power data. The authors also would
(29) like to thank Dr. D. Morton of the University of Texas at Austin,
Dr. K. Wood of the Naval Postgraduate School, and three anony-
Let . Then (13) can be re-written mous reviewers for their invaluable comments.
as

(30) REFERENCES
[1] Transpower, Reserve Management Tool, 2004, accessed Sep. 21, 2010.
In the case that , we have that [Online]. Available: http://www.electricitycommission.govt.nz/opdev/
servprovinfo/servprovpdfs/spc-system-operator-rmt-specification.pdf.
[2] A. Tuohy, E. Denny, P. Meibom, R. Barth, and M. O’Malley, “Oper-
ating the Irish power system with increased levels of wind power,” in
Proc. IEEE Power and Energy Society General Meeting—Conversion
On the other hand, if , which implies and Delivery of Electrical Energy in the 21st Century, Jul. 2008, pp.
. With the partial derivatives in (27)–(29) and def- 1–4.
[3] Eirgrid, Operating Reserve Requirements, 2010, accessed Feb. 11,
inition of (16), we have , a subgradient of , as 2012. [Online]. Available: http://www.eirgrid.com/aboutus/publica-
follows: tions.
[4] Lawrence Berkeley National Laboratory, Use of Frequency Response
Metrics to Assess the Planning and Operating Requirements for Re-
liable Integration of Variable Renewable Generation, 2010, accessed
Feb. 14, 2012. [Online]. Available: http://www.ferc.gov.
(31) [5] S. Takriti, J. Birge, and E. Long, “A stochastic model for the unit
commitment problem,” IEEE Trans. Power Syst., vol. 11, no. 3, pp.
1497–1508, Aug. 1996.
[6] F. Bouffard, F. Galiana, and A. Conejo, “Market-clearing with sto-
Note that might be non-differentiable at some , which chastic security—Part I: Formulation,” IEEE Trans. Power Syst., vol.
20, no. 4, pp. 1818–1826, Nov. 2005.
occurs when some generators have just binding capacity con- [7] L. Wu, M. Shahidehpour, and T. Li, “Stochastic security-constrained
straints at . In this case, any directional derivative can serve as unit commitment,” IEEE Trans. Power Syst., vol. 22, no. 2, pp.
a subgradient. One approach to evaluate a particular directional 800–811, May 2007.
[8] P. Ruiz, C. Philbrick, E. Zak, K. Cheung, and P. Sauer, “Uncertainty
derivative is applying a small change in and re-computing the management in the unit commitment problem,” IEEE Trans. Power
corresponding , and . The directional deriva- Syst., vol. 24, no. 2, pp. 642–651, May 2009.
tive is then calculated using (27)–(29). Note that (27)–(29) also [9] K. Hedman, M. Ferris, R. O’Neill, E. Fisher, and S. Oren, “Co-opti-
mization of generation unit commitment and transmission switching
provide a subgradient of in the case of governor delay. with n-1 reliability,” IEEE Trans. Power Syst., vol. 25, no. 2, pp.
The derivation is similar and therefore omitted. 1052–1063, May 2010.
[10] L. Kimball, K. Clements, and P. Davis, “Stochastic OPF via Bender’s
method,” in Proc. 2001 IEEE Porto Power Tech, 2001, vol. 3, 4 pp.
[11] A. Saric, F. Murphy, A. Soyster, and A. Stankovic, “Two-stage sto-
chastic programming model for market clearing with contingencies,”
APPENDIX C IEEE Trans. Power Syst., vol. 24, no. 3, pp. 1266–1278, Aug. 2009.
CONDITIONS FOR PRESERVING OPTIMALITY CUTS [12] C. B. Somuah and F. C. Schweppe, “Economic dispatch reserve
allocation,” IEEE Trans. Power App. Syst., vol. PAS-100, no. 5, pp.
For scenario with any first-stage decisions , let and 2635–2642, May 1981.
be initial values of and in (16), and let be [13] J. F. Restrepo and F. D. Galiana, “Unit commitment with primary fre-
quency regulation constraints,” IEEE Trans. Power Syst., vol. 20, no.
the associated second-stage problem. Similarly, define and 4, pp. 1836–1842, Nov. 2005.
as the updated values and let be the associated [14] R. Doherty, G. Lalor, and M. O’Malley, “Frequency control in com-
second-stage problem. If petitive electricity market dispatch,” IEEE Trans. Power Syst., vol. 20,
no. 3, pp. 1588–1596, Aug. 2005.
[15] P. Ruiz and P. Sauer, “Spinning contingency reserve: Economic value
(32) and demand functions,” IEEE Trans. Power Syst., vol. 23, no. 3, pp.
1071–1078, Aug. 2008.
then the right-hand side of (2b) of is greater than or [16] R. M. V. Slyke and R. Wets, “L-shaped linear programs with applica-
equal to that of . This implies has a smaller tions to optimal control and stochastic programming,” SIAM J. Appl.
Math., vol. 17, no. 4, pp. 638–663, 1969.
feasible region and larger optimal value, that is [17] T. Dy Liacco, “The adaptive reliability control system,” IEEE Trans.
Power App. Syst., vol. PAS-86, no. 5, pp. 517–531, May 1967.
[18] Disturbance Control Performance, Standard BAL-002-0, North Amer-
ican Electric Reliability Corporation, 2005, accessed Feb. 14, 2012.
Therefore, the existing optimality cuts, which are lower bounds [Online]. Available: http://www.nerc.com/files/BAL-002-0.pdf.
[19] A. J. Wood and B. F. Wollenberg, Power Generation, Operation, and
of , are still valid lower bounds of the updated second- Control, 2nd ed. New York: Wiley, 1996.
stage problem and therefore can be preserved. The con- [20] Electric Reliability Council of Texas, ERCOT Protocols, 2011, ac-
dition (32) also suggests that we can start with relatively smaller cessed Apr. 22, 2012. [Online]. Available: http://www.ercot.com.
[21] Midwest Independent System Operator, MISO Business Practices
initial values, in order to avoid removing the optimality cuts Manual: Energy and Operating Reserve Markets, 2010, accessed Feb.
after updating the coefficients. 14, 2012. [Online]. Available: http://www.midwestiso.org.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON POWER SYSTEMS

[22] A. V. Fiacco and J. Kyparisis, “Convexity and concavity properties of [32] North America Electric Reliability Corporation, Generating Avail-
the optimal value function in parametric nonlinear programming,” J. ability Data System (GADS), 2010, accessed Feb. 14, 2012. [Online].
Optimiz. Theory Appl., vol. 48, pp. 95–126, 1986. Available: http://www.nerc.com.
[23] T. Inoue, H. Taniguchi, Y. Ikeguchi, and K. Yoshida, “Monitoring [33] AWS Truewind, Analysis of West Texas Wind Plant Ramp-Up and
the first frequency derivative to improve adaptive underfrequency Ramp-Down Event, 2008, accessed Feb. 14, 2012. [Online]. Avail-
Load-Shedding schemes,” IEEE Trans. Power Syst., vol. 12, no. 1, pp. able: http://interchange.puc.state.tx.us/WebApp/Interchange/Docu-
136–143, Feb. 1997. ments/33672_1014_580034.PDF.
[24] Y.-Y. Lee, “Improving electricity market efficiency: From market
monitoring to reserve allocation,” Ph.D. dissertation, Univ. Texas,
Austin, TX, USA, 2012.
[25] J. F. Benders, “Partitioning procedures for solving mixed-variables
programming problems,” Numer. Math., vol. 4, pp. 238–252, 1962.
[26] J. R. Birge and H. Tang, L-Shape Method for Two Stage Problems of
Stochastic Convex Programming, 1993, accessed Feb. 14, 2012, tech. Yen-Yu Lee (M’12) received the B.Sc. degree in electrical engineering from
rep. [Online]. Available: available from http://deepblue.lib.umich.edu/ the National Taiwan University, Taipei, Taiwan and the M.S. and Ph.D. degrees
bitstream/2027.42/3628/5/bbm0214.0001.001.pdf. in electrical and computer engineering from the University of Texas in Austin,
[27] A. M. Geoffrion, “Generalized Benders decomposition,” J. Optimiz. Austin, TX, USA, in 2010 and 2012, respectively.
Theory Appl., vol. 10, no. 4, pp. 237–260, 1972. He is currently a power system engineer at the optimization application team
[28] J. J. E. Kelley, “The cutting-plane method for solving convex pro- in Alstom Grid.
grams,” J. Soc. Ind. Appl. Math., vol. 8, no. 4, pp. 703–712, Dec. 1960.
[29] W. W. Hogan, A Model for Zonal Operating Reserve Demand Curve,
2009, accessed Feb. 12, 2012. [Online]. Available: http://www.hks.har-
vard.edu/fs/whogan. Ross Baldick (F’07) received the B.Sc. degree in mathematics and physics and
[30] W. W. Hogan, Connecting Reliability Standards and Electricity Mar- the B.E. degree in electrical engineering from the University of Sydney, Aus-
kets, 2005, accessed Feb. 14, 2012. [Online]. Available: http://www. tralia and the M.S. and Ph.D. degrees in electrical engineering and computer
hks.harvard.edu/fs/whogan. sciences from the University of California, Berkeley, Berkeley, CA, USA, in
[31] Argonne National Laboratory, A Survey on Wind Power Ramp Fore- 1988 and 1990, respectively.
casting, 2010, accessed Feb. 14, 2012. [Online]. Available: http://www. He is currently a Professor in the Department of Electrical and Computer
dis.anl.gov/pubs/69166.pdf. Engineering at The University of Texas at Austin, Austin, TX, USA.