You are on page 1of 11

800 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO.

2, MAY 2012

Robust Unit Commitment With Wind


Power and Pumped Storage Hydro
Ruiwei Jiang, Student Member, IEEE, Jianhui Wang, Member, IEEE, and Yongpei Guan, Member, IEEE

Abstract—As renewable energy increasingly penetrates into Start-up cost for thermal generator at bus .
power grid systems, new challenges arise for system operators to
keep the systems reliable under uncertain circumstances, while Shut-down cost for thermal generator at bus .
ensuring high utilization of renewable energy. With the naturally
intermittent renewable energy, such as wind energy, playing more Minimum-up time for thermal generator at bus .
important roles, system robustness becomes a must. In this paper,
we propose a robust optimization approach to accommodate wind Minimum-down time for thermal generator at
output uncertainty, with the objective of providing a robust unit bus .
commitment schedule for the thermal generators in the day-ahead
market that minimizes the total cost under the worst wind power Ramp-up limit for thermal generator at bus .
output scenario. Robust optimization models the randomness
Ramp-down limit for thermal generator at bus .
using an uncertainty set which includes the worst-case scenario,
and protects this scenario under the minimal increment of costs. In Minimal output of electricity if thermal generator
our approach, the power system will be more reliable because the
worst-case scenario has been considered. In addition, we introduce
at bus is on.
a variable to control the conservatism of our model, by which we Maximal output of electricity if thermal generator
can avoid over-protection. By considering pumped-storage units, at bus is on.
the total cost is reduced significantly.
Index Terms—Generation scheduling, pumped-storage, robust Forecasted mean wind power output at bus in
optimization, wind uncertainty. period .
Upper limit of the confidence interval for the wind
NOMENCLATURE power output at bus in period .
Lower limit of the confidence interval for the
A. Sets wind power output at bus in period .
Set of buses. Initial water reserve inventory of pumped-storage
Set of transmission lines linking two buses. unit at bus .
Set of generators at bus . Target water reserve inventory of pumped-storage
unit at bus at the end of planning horizon.
B. Parameters Efficiency of the pumping cycle of the pumped
Number of time periods in planning horizon, with storage station. It is dimensionless and represents
each period to be one hour. the ratio of energy absorbed into storage to energy
taken from the power system % .
Line flow distribution factor for transmission line
linking bus and bus due to the net injection Efficiency of the generating cycle of the
at bus . pumped storage station. It is dimensionless and
represents the ratio of energy injected into the
Transmission capacity on transmission line power system to the energy drawn from storage
linking bus and bus . %.
Demand at bus in time period .
Manuscript received February 19, 2011; revised June 15, 2011 and September
14, 2011; accepted September 19, 2011. Date of publication November 03,
2011; date of current version April 18, 2012. This work was supported in part by C. Decision Variables
the U.S. National Science Foundation under CAREER Award CMMI0748204
and by the U.S. Department of Defense under Office of Naval Research Young
Investigator Award N000141010749. Paper no. TPWRS-00147-2011. Binary decision variable to indicate if thermal
R. Jiang and Y. Guan are with the Department of Industrial and Systems generator at bus is on in time period .
Engineering, University of Florida, Gainesville, FL 32611 USA (e-mail:
guan@ise.ufl.edu). Binary decision variable to indicate if thermal
J. Wang is with the Argonne National Laboratory, Argonne, IL 60439 USA generator at bus is started up in period .
(e-mail: jianhui.wang@anl.gov).
Color versions of one or more of the figures in this paper are available online Binary decision variable to indicate if thermal
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TPWRS.2011.2169817
generator at bus is shut down in period .

0885-8950/$26.00 © 2011 IEEE


JIANG et al.: ROBUST UNIT COMMITMENT WITH WIND POWER AND PUMPED STORAGE HYDRO 801

Amount of electricity generated by thermal captured by a number of wind power scenarios. Each scenario
generator at bus in period . represents a possible realization of the wind power output. The
difficulty with stochastic UC is that the problem size and com-
Actual wind power output at bus in period .
putational requirement increases with the number of scenarios
Total amount of electricity absorbed by since a large number of scenarios are often required to ensure
pumped-storage units at bus in period . the quality of the UC solution. In addition, the accuracy of
Total amount of electricity generated by the algorithm is sensitive to the scenario generation technique
pumped-storage units at bus in period . used. An alternative is to impose the wind power scenarios as
constraints in the UC model [2], [7]. In this case, the UC results
Water reserve inventory of the pumped-storage have to satisfy all the wind power variability in the scenarios.
unit at bus at the end of period . A Benders’ decomposition technique can be used to reduce the
Binary decision variable to indicate if the wind computational requirement. But this method also suffers from
farm output at bus reaches its upper bound in its sensitivity to wind power scenario selection.
time period . Another drawback of stochastic UC is that it is not easy to
represent the wind power “ramp” events in the scenarios, which
Binary decision variable to indicate if the wind refer to the situations where wind power output increases or
farm output at bus reaches its lower bound in decreases significantly during a short period of time due to
time period . fast-moving weather phenomena. The reasons are as follows:
Dual variables corresponding to the economic 1) those “ramp” events are difficult to predict because such
dispatch constraints. See Section III-A for details. events may come from a variety of situations such as cold
frontal passage, thunderstorm outflow, and rapid intensification
of an area of low pressure [10], and 2) the normal scenario
generation method usually creates all the scenarios at one time
based on certain distribution assumption without specifying
I. INTRODUCTION what scenarios can represent the ramp events. If the ramp event
scenario is defined, it will be difficult to mingle it with the
IND power has maintained increasing penetration into others and assign an appropriate probability. Wind power ramp
W power systems in the past several years. Large amounts
of wind power has sophisticated implications on power system
events cause great challenges to power system operations since
the other conventional generators may not be able to follow this
operations. The system operators have the obligation to balance rapid change due to their technical constraints such as ramping
the system to take the most advantage of wind power. How- and min on/off constraints. For example, on April 4, 2009, the
ever, due to the inherent uncertainty and variability in wind, the Southwestern Public Service (SPS) balancing authority in the
fluctuation in wind power has a great impact on system oper- U.S. experienced a drop of wind power output from 650 MW
ations. Various wind power forecasting techniques have been to 450 MW within one hour and another approximately 140
developed to forecast the wind power output [1]. Wind power MW drop in the second hour. While the peak load in that area
forecasting usually starts with a large-scale numerical weather was only 3214 MW, this sudden change in wind output could
prediction model. Then a variety of physics-based models and have disturbed the system operations greatly if this was not
statistical tools such as artificial neural networks (ANN) and re- addressed and forecasted properly [10].
gression are used to generate the wind power forecasts. Due to In this paper, we present a robust optimization-based UC
a number of factors involved in the forecasting process such as model that takes into account the worst scenario of wind
terrain complexity, prediction horizon, and data quality, wind power fluctuation with deterministic loads in all the periods.
power forecasting has remained as a very challenging task. The robust optimization (RO) approach has received growing
Research has been done to investigate how to represent attention in both research and applications in the last decade.
the wind power uncertainty in the system operations. Various Instead of making assumptions on specific probability distri-
methods can be used to accommodate wind power variability butions, the RO approach puts the random problem parameters
including advanced unit commitment (UC), and balancing in a predetermined uncertainty set containing the worst-case
wind power variations with pumped-storage hydro and better scenario. The way of defining the uncertainty makes the RO
ancillary service procurement [2]–[4]. In this paper, we focus approach more applicable than the traditional approaches.
on how to deal with wind power forecast errors and ensure high On one hand, the decision makers can easily construct the
utilization of wind power through advanced UC. An improved uncertainty set from the historical data or estimate it with a
UC model considering wind power uncertainty will lead to confidence interval, rather than betting on the probability dis-
a better UC solution which can withstand the forecast errors tributions. On the other hand, the robust counterpart, obtained
in the real time. There are several methods to represent the by replacing the random parameters by their uncertainty sets in
wind power forecasting errors in the UC model. One way is the deterministic formulation, usually remains computationally
to accommodate wind power uncertainty through stochastic tractable even for medium or large scale problems in practice.
UC (for literature on stochastic programming, see [5]; for In view of this, the RO approach has recently been applied in
its application on stochastic UC, see [6]–[8] and [9], among various fields including network flows [11], [12], production
others). In such a model, the uncertainty of wind power is planning problems [12], evacuation transportation planning
802 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO. 2, MAY 2012

problems [13], portfolio management problems [14], and con- be far away from its forecasted value (i.e., deviations are
tingency-constrained unit commitment problems [15], among very large). In practice, the planning horizon equals 24 for
others. In this paper, we employ the RO approach to hedge the real-time market. If , then is forced to be close
the wind power output uncertainty. Specifically, the uncertain to (e.g., the deviations are small and we put them equal in
wind power output in each period is within an interval defined the formulation) throughout the next 24 hours. If , then
by its lower and upper bounds and can be obtained based on can take any value in in any arbitrary 12 hours,
the historical data or estimated with a confidence interval. This but is assumed to be close to in the remaining 12 hours. It
setting contains the worst possible scenario of wind power can be observed that we become more conservative as in-
fluctuation for a pre-specified set of quantiles. Our proposed creases, and we can view as the “budget of uncertainty.” In
method aims to provide a robust unit commitment schedule practice, the system operators can control the system robustness
for the day-ahead market that minimizes the total cost under by adjusting the value of . For example, in the event that a
the worst disturbance the wind power fluctuation can cause to weather front comes through, the system operators can define a
the system in the real time, while ensuring high utilization of certain number of hours when the actual wind power output will
wind power. This problem can be formulated as a two-stage significantly deviate from the forecasted values. The algorithm
min-max problem with the objective of minimizing the total can automatically simulate the worst scenario where the wind
cost under the worst wind power output scenario. In this paper, power output changes between upper and lower bounds and has
we apply the Benders’ decomposition algorithm to obtain a the greatest impact on system operations.
robust unit commitment solution for the day-ahead market, and In our model, the main objective is to derive a robust unit
a Monte Carlo sampling approach to obtain an upper bound of commitment schedule. For instance, through selecting the
the optimal objective value for the problem with the purpose budget of uncertainty (i.e., ), this model can guarantee that
of evaluating the solution quality. The algorithm can automati- the robust optimal solution is feasible with a high probability
cally simulate the worst scenario where the wind power output under wind power output uncertainty. Based on the bound
changes between upper and lower bounds and has the greatest derived in [16, formulation (20)], the robust optimal solution is
impact on system operations. Hence, the ramp events which feasible with probability approximately greater than 95% if
represent the inter-hour change in wind power can be modeled. exceeds , where represents the number of random
In the remaining part of this paper, we describe the mathe- variables. Hence, in our case, it states that .
matical formulation in Section II. In Section III, we develop Thus, we can decide the number of periods with big deviations
the solution approach to solve the problem. Case studies and (or equivalently, the budget of uncertainty ) by the feasibility
computational results are shown in Section IV, and finally, in probability guarantee, which can be further selected by the
Section V, we conclude our research. power system operators.
Once the budget of uncertainty is given, it is easy to observe
II. MATHEMATICAL FORMULATION that there exists an optimal solution in which reaches its
In this section, we build a two-stage robust optimization for- boundary or meets its forecasted value. Accordingly, we let
mulation to deal with the ramp events for the uncertain wind and represent if the wind power output reaches its upper
power outputs. In our setting, we model the random wind power bound , its lower bound , or meets its
output with an uncertainty set. We assume the uncertain wind forecasted value . For notation brevity, we let
power output is within an interval with the forecasted and . The uncertainty set
mean value to be . The interval can usually be can be described as
obtained based on the historical data or an interval forecast for
the wind power output. We can use quantiles to help us define
the interval if quantiles are known. For instance, we can set
and equal to the .95- and .05-quantiles of the random
wind power output, respectively. In this way, we might curtail (1)
the excess wind power if the wind power output is larger than
.95-quantile, and curtail partial loads if the wind power output
is less than .05-quantile. We can ensure at least 95% of the wind Based on this uncertainty set description, we can provide the
power output will be utilized. Under our setting, for each bus robust optimization formulation to address the wind power un-
in period , the actual wind power output can run freely in certainty as follows [denoted as formulation ]:
the range . It can be observed that the worse-case sce-
nario is at or for all and , and we are
apparently overconservative. Since small deviations which can
be balanced by dispatching other online conventional generating
units up or down will not cause serious issues to the system op- (2)
eration, we focus on the extreme cases where huge fluctuations
of wind power output between the hours occur (ramps), which
is defined as the possible worst case in the paper. To adjust the
degree of conservatism, we employ an integer (between 0 (3)
and ) to restrict the number of time periods that allow to
JIANG et al.: ROBUST UNIT COMMITMENT WITH WIND POWER AND PUMPED STORAGE HYDRO 803

(4) least units of time), start-up and shut-down constraints [(5),


(5) (6)], generation capacity constraints (8), ramping constraints
[(9), (10)] (in this paper, ramping constraints combine ramp-rate
(6) limit restriction when a generator is on and first-hour (last-hour)
minimum generation restriction as described in [17]), system
(7) power balance constraints (11), transmission constraints (12),
and hydro water inventory balance constraints (13) (the units
where for power generation and water reserve level are unified to be
MWh as described in [18]). Besides, constraints (14) and (15)
describe the upper and lower bounds of electricity absorbed and
generated by the pumped-storage units, and constraints [(16),
(17)] give the initial and target water inventory level for the
(8) pumped-storage units. Finally, constraints [(18), (19)] ensure
that the pumped-storage units cannot absorb and generate elec-
tricity simultaneously within any specific time period, by intro-
(9) ducing a binary variable to indicate whether the unit at bus
absorbs or generates electricity in time
period .

(10)
III. SOLUTION METHODOLOGY

(11) For the proposed formulation, we need to solve a


min-max-min problem, which could not be solved directly
by a commercial software such as CPLEX. Therefore, efficient
algorithms should be developed to solve the problem. If we
use a scenario-based approach, e.g., enumerating all possible
(12) outcomes of that could lead to the worst-case scenario for
the problem, then there are outcomes. For each outcome,
(13) we need to consider separate economic dispatch constraints
corresponding to the outcome. Thus, the problem size is huge
(14) and in the remaining part of this section, we develop a decom-
(15) position algorithm to solve the problem.
(16)
A. Second-Stage Problem Transformation
(17)
In practice, the fuel cost is a quadratic function and
(18)
can be approximated by a -piece piecewise linear function as
follows:
(19)

In the above formulation, the objective function (2) consists of (20)


start-up, shut-down, and operating costs (the operating costs for
absorbing or generating power by pumped storage hydros are Furthermore, constraints [(18), (19)] make the second-stage
usually very low and therefore not considered in this model). As problem a mixed integer program, which significantly increases
shown mathematically, we will make the unit commitment deci- the difficulty in solving to optimality. In view of this,
sions in the first stage, which represents the unit com- we first relax constraints [(18), (19)] and solve to get a
mitment for the day-ahead market, while in the second stage, lower bound of the optimal objective value. By dualizing the re-
we decide how much electricity to be generated by the thermal, maining constraints in , we can transform the second-stage
wind, and hydro, which corresponds to the economic dispatch in problem as follows [denoted as formulation ]:
the real time. Note that, in the formulation, wind power gener-
ation amounts are subject to uncertainty. They
are not deterministic, but random variables described by the un-
certainty set . Our objective is to minimize the total costs under
the worst-case scenario in terms of the uncertain wind power
output in the real time. The hourly UC constraints listed above
include the minimum-up and -down time constraints [(3), (4)]
(e.g., constraints [(3)] represent if a generator is turned on, it (21)
should be kept on for at least units of time; constraints [(4)]
represent if a generator is turned off, it should be kept off for at
804 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO. 2, MAY 2012

(22) (34)

where (34) is the replacement of (27).


(23) Next we employ the Benders’ decomposition algorithm to
solve . We solve the following master problem iteratively by
adding new constraints to cut off the infeasible or non-optimal
solutions [denoted as formulation ]:
(24)
(25)

(26)

(27) (35)

(28)
(36)
where is a linear function and the de-
tailed expression is shown in the Appendix. where constraints (35) represent the selected feasibility cuts
In the above formulation , and are the added to the master program, while constraints (36) represent
dual decision variables for constraints (8), [(9)–(10)], (11), (12), the optimality cuts.
(13), (14), (15), and (20), respectively. The objective of the dual
formulation is represented as (21). Constraints (22), (23), (24), B. Feasibility Cuts
and (25) correspond to the variables and in
. Note that the first term in the objective function is bilinear. We say a first-stage solution is infeasible if
For instance, there are bilinear terms in , implying that the program is un-
the objective function. To linearize these bilinear terms, we let bounded in value. From the duality theory, the unboundedness
. Then we have of indicates the infeasibility of the subproblem
, and thus the set
is empty. Intuitively, it means that given a unit commitment
decision , the load in power grid cannot be satisfied under the
restriction of transmission and generator capacity constraints,
however hard we try. To detect the infeasibility of the
subproblem, we construct the feasibility problem. To formulate
the dual problem, we do not need to consider the objective
function (20), and the corresponding formulation is shown as
follows [denoted as formulation ]:

subject to

(29)

(30)
(31)
(32)
(33)

By defining
, we have the
following mixed-integer-program formulation for the dual of
the second-stage problem of :
JIANG et al.: ROBUST UNIT COMMITMENT WITH WIND POWER AND PUMPED STORAGE HYDRO 805

where is a linear function and the detailed


expression is shown in the Appendix. The feasibility of a given
can be tested as follows:
• If is feasible;
• If , generate a feasibility cut in form of (35) with

where is an optimal solution to , and add the cut


to .

C. Optimality Cuts
In the formulation , since we denote as the optimal value
of the second-stage problem, we should have .
Hence, suppose in the th iteration we solve and obtain
and . If we substitute into and after optimized we have
, then we can claim that is not optimal and add an
optimality cut in the form of (36) with

Fig. 1. Framework of the Benders’ decomposition algorithm.

Third, we solve the optimization problem


with
and fixed.
Finally, suppose that we sample wind power output sce-
narios, namely , then we get a statistical upper
bound

where , and make up an op-


timal solution to .

D. Algorithm Framework
We finish this section with the framework of our algorithm,
as shown in Fig. 1. We first solve without constraints
[(18), (19)] following the procedure described in previous Note that we use Monte Carlo sampling and make an assump-
Subsections A, B, and C, and obtain a robust unit commitment tion of the wind power output distribution only for obtaining an
schedule for the day-ahead market and a lower bound of the upper bound to verify the quality of our proposed solution ap-
optimal objective value. proach. In general, the distribution of the wind power output
Then, we obtain a statistical upper bound of the optimal ob- is not required to obtain the robust unit commitment decisions.
jective value by Monte Carlo sampling of various wind power Note also that constraints [(18), (19)] are considered in the sam-
output scenarios. We sample the wind power output as fol- pling procedure to obtain the statistical upper bound and guar-
lows. For each bus in period , we generate a random variable antee a feasible solution.
rv subject to uniform distribution over [0, 1], and assign
a) and , if ;
b) and , if ; IV. COMPUTATIONAL RESULT
c) and , otherwise. In this section, we present the numerical experiments of
If , we go back to the beginning of this the proposed algorithm on two data sets. First, we introduce a
step and regenerate rv; otherwise, we go to the next step. six-bus system as a case study, and apply the robust optimization
806 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO. 2, MAY 2012

TABLE III
TRANSMISSION LINE DATA

Fig. 2. Six-bus system.

TABLE I
GENERATOR DATA

Fig. 3. Wind power output evolution over time.

TABLE II
FUEL DATA

approach to provide the operation of the small power grid. Be- Fig. 4. Worst-case scenario and water reserve level variation. Curve 1 depicts
sides, we demonstrate the contribution of the pumped-storage the deterministic load; curve 2 depicts the forecasted net load, i.e., the difference
between the load and the forecasted wind power output; curve 3 depicts the
units to system robustness by comparing the solutions and worst-case net load; curve 4 depicts the water reserve level.
values for the cases with and without pumped-storage units.
Second, we apply both deterministic and robust optimization
approaches on a modified IEEE 118-bus system, and compare wind power output is away from its forecasted value. Further,
their performance under the worst-case scenario for various the forecasted value and upper and lower limits of wind power
data settings. All the experiments are implemented using output in each period are shown in Fig. 3, which is provided
CPLEX 12.1, at Intel Quad Core 2.40 GHz with 8 GB memory. by the statistics from National Renewable Energy Laboratory
(NREL).
A. Six-Bus System The robust optimal unit commitment decisions and
As shown in Fig. 2, the six-bus system includes four thermal pumped-storage operations in the worst-case scenario of
generators at , and , one wind farm at , one the six-bus system are collected in Table IV. Since the wind
pumped-storage unit at , and seven transmission lines. The power output can be considered as negative system load, we
loads are deterministic with mean values 50 MW at , 80 MW can define the net load in a time period as the difference of the
at , and 250 MW at . The characteristics of the generators load and the wind power output in this period. If we depict the
and transmission lines are described in Tables I–III. net load under worst-case scenario and the corresponding water
For the pumped-storage unit, we assume reserve level as in Fig. 4, we observe that as time goes, starting
MW. Besides, we assume that the efficiency of the from period 10, is turned on due to net load increment.
pumped-storage unit is 0.8 on both absorbing and generating We also observe that the pumped-storage absorbs power when
electricity. For the wind farm, we assume that the budget of un- the net load is low (e.g., in periods 1 and 3–6) and generates
certainty is 6. That is, we allow six time periods in which the power when the net load is high (e.g., in periods 8, 9 and 24
JIANG et al.: ROBUST UNIT COMMITMENT WITH WIND POWER AND PUMPED STORAGE HYDRO 807

TABLE IV TABLE V
ROBUST OPTIMAL SOLUTION OF THE SIX-BUS COMPUTATIONAL RESULTS FOR AN IEEE 118-BUS SYSTEM UNDER
SYSTEM WITH PUMPED-STORAGE UNITS VARIOUS UNCERTAINTY LEVELS AND HYDRO CAPACITIES—OPTIMAL
VALUES AND WORST-CASE PERFORMANCE

TABLE VI
COMPUTATIONAL RESULTS FOR AN IEEE 118-BUS SYSTEM UNDER
VARIOUS UNCERTAINTY LEVELS AND HYDRO CAPACITIES—OPTIMALITY
GAPS, WORST-CASE VALUE GAPS, AND CPU TIME

to avoid putting online, and in periods 20 and 21 to avoid


putting online). With this observation, we can conclude
that the pumped-storage unit works as a power buffer for the
system, by charging/discharging in accordance with the net load
information. Then, we test the case without pumped-storage
units. For this case, as compared to the previous one with
pumped-storage units, the robust optimal solution turns on
from time periods 20 to 23, and extend the on status for
to periods 8, 9, and 24, in addition to the generator schedule
for the case with pumped-storage; the corresponding optimal
value is 111741.36. We can observe that in the case without
pumped-storage units, the system has to turn on more thermal
generators and for longer time to ensure the robustness, which
might incur larger generation costs and more greenhouse gas
emissions. In other words, the pumped-storage units can act
as substitutes of thermal generators on increasing the system or generate per period by its ratio over the average forecasted
robustness, and at the same time reduce greenhouse gas emis- wind power output. In other words, the pumped-storage unit be-
sions. comes more capable to absorb or generate power as the ratio
increases. Similar to , we allow the ratio to vary within the in-
B. Modified 118-Bus System terval [0, 0.5]. The computational results are shown in Tables V
and VI. In Table V, we provide the following lower and upper
A modified IEEE 118-bus system, based on the one given on- bounds:
line at motor.ece.iit.edu/data, is used to test the proposed algo- 1) Lower bound LB is obtained by applying the robust opti-
rithm. In this experiment, there are 118 buses, 33 thermal gen- mization approach to solve without constraints [(18),
erators, 186 transmission lines, 1 wind farm, and 1 pumped- (19)].
storage unit in the power grid and the time horizon is 24 hours. 2) Upper bound UB is a statistical upper bound obtained by
We assume that the forecasted wind power output follows the Monte Carlo simulation of the wind power output, as dis-
pattern illustrated in Fig. 3. cussed in Section III-D.
We first test the system performance under various uncer- 3) Upper bound WV represents the worst-case value, ob-
tainty levels and hydro capacities. We allow the budget of un- tained by evaluating the performance of the deterministic
certainty to vary within the interval [2, 10]. We measure the optimal solution under the worst-case scenario. In other
upper bound of electricity the pumped-storage units can absorb words, we first solve the deterministic problem, in which
808 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO. 2, MAY 2012

TABLE VII
COMPUTATIONAL RESULTS FOR AN IEEE 118-BUS SYSTEM UNDER
VARIOUS WIND POWER PENETRATION LEVELS—OPTIMAL VALUES,
GAPS, AND WORST-CASE PERFORMANCE

Fig. 5. Optimality gaps and worst-case performance under various wind power
penetration levels.

the wind power output in each period is fixed at its fore-


casted value, and then generate random large size scenarios
of wind power output to obtain an upper bound as we did
in obtaining UB. In the experiments, we found that the
deterministic optimal solutions are infeasible to for
all the instances. This observation indicates that it would
be risky to overlook the uncertainty of wind power output
and totally rely on the forecasted values. To compare
the performance between these solutions, we introduce
a linear penalty function for any unsatisfied demand or
transmission capacity/ramp-rate limit violations, and the
unit penalty cost is set to be $7947/MWh.
Correspondingly, we calculate the following gaps based on the
lower and upper bounds:
1) %. It estimates the Fig. 6. Optimality gaps and worst-case performance under various transmis-
optimality gap of the proposed algorithm. sion line capacities.
2) %. It estimates the dif-
ference between the performance of the robust and deter-
ministic optimal solutions under the worst-case scenario, lower and upper bounds of the robust optimal value and the cor-
when the linear penalty function is introduced in calcu- responding WV gaps under various penetration levels. In ad-
lating WV. dition, we depict the gaps over different penetration levels in
In the experiments, we first observe that statistically our algo- Fig. 5. From the computational results, we first observe that our
rithm provides a feasible solution for all instances with algorithm provides tight lower and upper bounds, which make
and Ratio . Besides, the optimality gaps are LB a reasonable estimate of the robust optimal value. Second, by
less than 0.1%, and the CPU times are less than one hour in all comparing LB and WV over various wind penetration levels, we
the instances. This result demonstrates that our algorithm can observe that LB decreases slightly with wind power increasingly
provide a solution that is very close to the optimal one within penetrating into the power system, while WV gap grows signif-
a reasonable amount of time. In other words, it is sufficient to icantly. This observation implies that RO methods can generate
apply our algorithm to solve large scale problems, and obtain reliable unit commitment schedules which are not only immune
near-optimal solutions. Second, we observe that WV gap in- against increasing wind power uncertainty, but also take advan-
creases with increasing and Ratio decreasing. This observa- tage of the low unit costs of wind power.
tion indicates that the deterministic optimal solution becomes Finally, we test the instances with various transmission line
worse with the wind power output uncertainty increased and capacities with other parameters fixed (e.g., and the hydro
the usage of pumped-storage units restricted. Hence, it becomes capacity ). Again, we measure the transmission line
even more valuable to emphasize the system robustness and the capacity by its ratio over the actual value for the given 118-bus
usage of power buffer, such as the pumped-storage units in our system, and allow the ratio to run from 0.5 to 1.5. We report
case, under such circumstances. the computational results in Table VIII and depict the gaps over
Then, we test the system performance under various wind different capacities in Fig. 6. In Table VIII, we list the lower
power penetration levels with fixed and hydro capacity and upper bounds of the robust optimal objective value and the
. Similar to the hydro capacity, we measure the corresponding gaps under various transmission capacities. From
wind power penetration level by the ratio of the nominal wind the computational results, we first notice that the problem be-
farm output over the forecasted value shown in Fig. 3. We allow comes infeasible when the transmission line capacity is suffi-
the ratio to run from 0.5 to 1.5 and display in Table VII the ciently low (e.g., half of the actual value). Second, in view that
JIANG et al.: ROBUST UNIT COMMITMENT WITH WIND POWER AND PUMPED STORAGE HYDRO 809

TABLE VIII where is a linear function and it is equal


COMPUTATIONAL RESULTS FOR AN IEEE 118-BUS SYSTEM UNDER to
VARIOUS TRANSMISSION LINE CAPACITIES—OPTIMAL VALUES,
GAPS, AND WORST-CASE PERFORMANCE

the bounds we provide are tight, we observe that the robust op-
timal value decreases slightly with increasing transmission line
capacities while WV gap decreases dramatically. This indicates
that RO method is stable for varying transmission line capaci-
ties, while the nominal optimal solution is highly sensitive when
the capacities decrease.

V. CONCLUSION
In this paper, we developed an innovative approach that in-
cludes applying robust optimization concepts and incorporating
pumped-storage units to hedge wind power output uncertainty.
We provided an uncertainty set description that can capture the
wind power “ramp” events and our proposed approach can pro- Following the similar procedure above for
vide a robust optimal solution that minimizes the total cost under with (20) not considered, we
the worst disturbance the wind power fluctuation can cause to have
the system in the real time, while ensuring high utilization of .
wind power. Meanwhile, this robust solution is feasible with a
high probability under wind power output uncertainty. In addi- ACKNOWLEDGMENT
tion, by incorporating pumped storage hydro units in the real
The authors would like to thank the editor and reviewers for
time, our robust optimization model contains discrete decision
their sincere suggestions on improving the quality of this paper.
variables in both the first and second stage problems, and we
developed an algorithm framework that can derive tight lower
REFERENCES
and upper bounds of the optimal objective value. Finally, our
[1] J. Wang, A. Botterud, V. Miranda, C. Monteiro, and G. Sheble, “Impact
computational results verify the tightness of the bounds and the of wind power forecasting on unit commitment and dispatch,” in Proc.
robustness of our solution under different simulation settings. 8th Int. Workshop Large-Scale Integration of Wind Power Into Power
Systems, Bremen, Germany, Oct. 2009.
[2] J. Wang, M. Shahidehpour, and Z. Li, “Security-constrained unit com-
mitment with volatile wind power generation,” IEEE Trans. Power
Syst., vol. 23, no. 3, pp. 1319–1327, Aug. 2008.
APPENDIX [3] J. Garcia-Gonzalez, R. de la Muela, L. Santos, and A. Gonzalez, “Sto-
DETAILED EXPRESSION OF AND chastic joint optimization of wind generation and pumped-storage units
in an electricity market,” IEEE Trans. Power Syst., vol. 23, no. 2, pp.
460–468, May 2008.
Let the decision variables be the dual vari- [4] R. Doherty and M. O’Malley, “A new approach to quantify reserve de-
ables for constraints (8), [(9)–(10)], (11), (12), (13), (14), (15), mand in systems with significant installed wind capacity,” IEEE Trans.
Power Syst., vol. 20, no. 2, pp. 587–595, May 2005.
and (20), respectively. Then, the objective function [i.e., ] [5] J. R. Birge and F. Louveaux, Introduction to Stochastic Program-
of the dual problem for ming. New York: Springer-Verlag, 1997.
can be written as follows: [6] R. Barth, H. Brand, P. Meibom, and C. Weber, “A stochastic unit-com-
mitment model for the evaluation of the impacts of integration of large
amounts of intermittent wind power,” in Proc. Int. Conf. Probabilistic
Methods Applied to Power Systems, 2006, pp. 1–8.
[7] F. Bouffard and F. D. Galiana, “Stochastic security for operations plan-
ning with significant wind power generation,” in Proc. Power and En-
ergy Society General Meeting—Conversion and Delivery of Electrical
Energy in the 21st Century, Pittsburgh, PA, 2008.
810 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 27, NO. 2, MAY 2012

[8] A. Tuohy, P. Meibom, E. Denny, and M. O’Malley, “Unit commitment Jianhui Wang (M’07) received the Ph.D. degree in
for systems with significant wind penetration,” IEEE Trans. Power electrical engineering from Illinois Institute of Tech-
Syst., vol. 24, no. 2, pp. 592–601, May 2009. nology, Chicago, in 2007.
[9] P. A. Ruiz, C. R. Philbrick, E. Zak, K. W. Cheung, and P. W. Sauer, Presently, he is a Computational Engineer with the
“Uncertainty management in the unit commitment problem,” IEEE Decision and Information Sciences Division at Ar-
Trans. Power Syst., vol. 24, no. 2, pp. 642–651, May 2009. gonne National Laboratory, Argonne, IL.
[10] W. Grant, D. Edelson, J. Dumas, J. Zack, M. Ahlstrom, J. Kehler, P. Dr. Wang is the chair of the IEEE Power & Energy
Storck, J. Lerner, K. Parks, and C. Finley, “Change in the air,” IEEE Society (PES) power system operation methods
Power Energy Mag., vol. 7, no. 6, pp. 47–58, 2009. subcommittee and co-chair of an IEEE task force
[11] D. Bertsimas and M. Sim, “Robust discrete optimization and network on the integration of wind and solar power into
flows,” Math. Program., vol. 98, no. 1, pp. 49–71, 2003. power system operations. He is an editor of the IEEE
[12] A. Atamturk and M. Zhang, “Two-stage robust network flow and design TRANSACTIONS ON SMART GRID, a guest editor of a special issue on Electrifi-
under demand uncertainty,” Oper. Res., vol. 55, no. 4, pp. 662–673, cation of Transportation of the IEEE POWER AND ENERGY MAGAZINE, and a
2007. guest editor of a special issue on Smart Grids, Renewable Energy Integration,
[13] T. Yao, S. R. Mandala, and B. D. Chung, “Evacuation transportation and Climate Change Mitigation—Future Electric Energy Systems of Applied
planning under uncertainty: A robust optimization approach,” Netw. Energy. He is the technical program chair of the IEEE Innovative Smart Grid
Spatial Econ., vol. 9, pp. 171–189, 2009. Technologies conference 2012.
[14] Y. Moon and T. Yao, “A robust mean absolute deviation model for port-
folio optimization,” Comput. Oper. Res., vol. 38, no. 9, pp. 1251–1258,
2011.
[15] A. Street, F. Oliveira, and J. M. Arroyo, “Contingency-constrained unit
0
commitment with n k security criterion: A robust optimization ap-
Yongpei Guan (M’10) received the B.S. degree in
mechanical engineering and the B.S. degree in eco-
proach,” IEEE Trans. Power Syst., vol. 26, no. 3, pp. 1581–1590, Aug. nomic decision science from the Shanghai Jiaotong
2011. University, Shanghai, China, in 1998, the M.S. de-
[16] D. Bertsimas and M. Sim, “The price of robustness,” Oper. Res., vol. gree in industrial engineering and engineering man-
52, no. 1, pp. 35–53, 2004. agement from the Hong Kong University of Science
[17] X. Guan, P. B. Luh, J. A. Yan, and J. A. Amalfi, “An optimization-based and Technology, Hong Kong, China, in 2001, and the
method for unit commitment,” Int. J. Elect. Power Energy Syst., vol. 14, Ph.D. degree in industrial and systems engineering
no. 1, pp. 9–17, 1992. from the Georgia Institute of Technology, Atlanta,
[18] P. D. Brown, J. A. P. Lopes, and M. A. Matos, “Optimization of GA, in 2005.
pumped storage capacity in an isolated power system with large He is the director of the computational and sto-
renewable penetration,” IEEE Trans. Power Syst., vol. 23, no. 2, pp. chastic optimization lab at the University of Florida, Gainesville, FL. Presently,
523–531, May 2008. his research interests include power grid security and energy policy analysis.
Ruiwei Jiang (S’11) received the B.S. degree in industrial engineering from Ts-
inghua University, Beijing, China, in 2009. He is currently pursuing the Ph.D.
degree in the Department of Industrial and Systems Engineering at the Univer-
sity of Florida, Gainsville, FL.
His research interests include stochastic optimization and power grid security
analysis.

You might also like