IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO.

3, MAY 2016 2383

Reliability-Constrained Robust Power

System Expansion Planning
Shahab Dehghan, Member, IEEE, Nima Amjady, Senior Member, IEEE, and Antonio J. Conejo, Fellow, IEEE

Abstract—This paper presents a tri-level reliability-constrained Annualized investment cost of unit and of line
robust power system expansion planning (SEP) framework mod- ($).
eling the uncertainty of electricity demand, wind power genera- Availability status of unit and of line (1/0
tion, and availability of units and lines. In the proposed tri-level stating availability/unavailability status).
SEP model, electricity demand and wind power generation being
continuous uncertain variables are modeled by bounded intervals Lower and upper bounds for the decomposition
and availability of units and lines being discrete uncertain vari- algorithm.
ables are modeled by probability distributions. Simulation results Sufficiently large positive constant.
on the Garver 6-bus, the IEEE 24-bus, and the IEEE 73-bus test Operation cost of unit ($).
systems confirm the effectiveness of the proposed tri-level frame-
work to cope with multi-fold planning uncertainties. Nominal estimate for electricity demand at bus
Index Terms—Reliability, robust optimization, stochastic pro- in operating condition (MW).
gramming, system expansion planning. Lower and upper estimates for electricity demand
at bus in operating condition (MW).
Penalty cost ($).
NOMENCLATURE Capacity of unit (MW).
1) Functions: Capacity of line (MW).
EENS Expected energy not supplied. Repair time of unit and of line (h).
2) Indices: Nominal estimate for wind power generation at
bus in operating condition (MW).
Index of operating conditions. Lower and upper estimates for wind power
Index of units. generation at bus in operating condition
Index of Monte Carlo simulation samples. (MW).
Susceptance of line (mho).
Index of iterations.
Maximum allowable limit for variables.
Index of lines.
Convergence tolerance.
Index of buses.
Degree of robustness.
Sending bus of lines .
Receiving bus of lines . 4) Sets:
Set of operating conditions.
3) Parameters:
Set of units and of lines.
Duration time of operating condition .
Set of units connected to bus .
at iteration (MWh).
Set of operation and of reliability cuts.
Target value of (MWh).
Set of buses.
Energy not supplied for sample (MWh).
Feasible region for investment decisions.
Failure rate of unit and of line (Occ./Yr.).
Feasible region for operation decisions.
Energy not supplied for sample (MWh).
Polyhedral uncertainty set.
Manuscript received January 31, 2015; revised June 03, 2015; accepted July 5) Variables:
26, 2015. Date of publication August 19, 2015; date of current version April 15,
2016. Paper no. TPWRS-00139-2015. Power generation of unit in operating condition
S. Dehghan is with the Department of Electrical, Biomedical, and Mecha- (MW).
tronics Engineering, Qazvin Branch, Islamic Azad University, Qazvin, Iran Power flow of line in operating condition
(e-mail: shahab.dehghan@ieee.org).
N. Amjady is with the Department of Electrical and Computer Engineering, (MW).
Semnan University, Semnan, Iran (e-mail: amjady@semnan.ac.ir). Uncertain electricity demand at bus in operating
A. J. Conejo is with the Integrated System Engineering and the Electrical and condition (MW).
Computer Engineering Departments, The Ohio State University, Columbus, OH
43210 USA (e-mail: conejonavarro.1@osu.edu).
Unserved electricity demand at bus in operating
Digital Object Identifier 10.1109/TPWRS.2015.2464274 condition (MW).

0885-8950 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2384 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 3, MAY 2016

Uncertain wind power generation at bus in others, most of previous research works deal with generation
operating condition (MW). and transmission expansion planning separately [1], [2]. How-
Status of unit and of line (e.g., ever, some SEP approaches have been presented in the liter-
constructed; otherwise). ature employing reliability cost-worth criterion [3], mixed-in-
Phase angle of bus in operating condition teger linear programming (MILP) [4]–[6], and mixed-integer
(radian). nonlinear programming (MINLP) [7].
Phase angle of reference bus in operating Any power system is subject to different low-frequency un-
condition (radian). certainty sources (e.g., annual peak load) and high-frequency
6) Vectors and Matrices: uncertainty sources (e.g., availability of units and lines) [8].
Moreover, in the past decade, increasing integration of intermit-
Matrices of coefficients.
tent renewable energy sources has resulted in increasing chal-
Vectors of requirements. lenges and uncertainties in planning studies. Accordingly, it is
Vectors of investment and operation costs. necessary to develop non-deterministic SEP frameworks coping
Vector of coefficients for the operation with the uncertain nature of power systems and finding a flexible
variables in the objective function of the expansion plan immunized against different realizations of un-
load-curtailment problem (i.e., certainty sources. Stochastic programming (SP) [9], modeling
). uncertain variables by means of scenarios [4]–[7], [10] and ro-
Vector of availability status of equipment bust optimization (RO) [11], modeling uncertain variables by
(e.g., ); when means of bounded intervals [12]–[16], have been used in the
the uncertainty of equipment availability is literature to handle a wide range of uncertainty sources. RO
considered, this vector is shown by . is particularly attractive in case of low-frequency uncertainty
Simulated vector of availability status of sources [13], [15], as it only requires the relative variation in-
equipment at iteration . tervals of uncertain variables rather than generating scenarios.
Vector of electricity demand; when the Besides, RO is more conservative than SP as the solution of
uncertainty of electricity demand is considered, RO is optimal for the worst-case realization of uncertain vari-
this vector is shown by . ables while the solution of SP is optimal on average for a set
of deterministic scenarios capturing the uncertainty spectrum.
Vectors of lower and upper estimates of .
Over-conservatism of the optimal solution may be a disadvan-
Vector of fixed obtained by solving the tage of RO-based optimization models. However, this issue is
operation sub-problem at iteration . more significant in the case of pioneering RO-based optimiza-
Vector of wind power generation; when the tion models, such as the robust counterpart of linear program-
uncertainty of wind power generation is ming problem introduced by Soyster [11], which are not capable
considered, this vector is shown by . of controlling the conservatism of the optimal solution. This
Vectors of lower and upper estimates of . issue is alleviated in the case of recent RO-based optimization
Vector of fixed obtained by solving the models, which have the ability of controlling the conservatism
operation sub-problem at iteration . of the optimal solution. For instance, the conservatism of the
Vector of investment variables (i.e., optimal solution in the proposed model can be controlled by a
). particular parameter named degree of robustness.
Vector of fixed obtained by solving the In [4] and [5], SP is used to cope with uncertainties pertaining
investment master problem at iteration . to electricity demands and availability of equipment within
Vector of operation variables (i.e., the SEP problem. In [6], SP is employed to handle the uncer-
tainties of SEP pertaining to correlated electricity demands
; ). and wind power generations. Also, the SEP problem under
Vector of operation variables added to the uncertain electricity demand is solved in [7] by incorporating
master problem to create operation primal cut a risk factor, based on the mean-variance theory, into SP. In
at iteration . [16], RO is employed to handle the uncertainties pertaining to
Vector of expected sensitivity of EENS with availability of equipment (i.e., units and lines) within the SEP
respect to at iteration . problem. Although the SEP problem has been solved previously
under different uncertainty sources, to the best of the authors'
I. INTRODUCTION knowledge, there is no non-deterministic SEP model for power
systems including wind units and considering both discrete and
A. Background and Motivation continuous uncertain variables simultaneously. Accordingly,

F ROM the viewpoint of the independent system operator

(ISO), or the regulator, the aim of system expansion plan-
ning (SEP) is to determine when, where, and what type of units
this paper presents a novel tri-level reliability-constrained
SEP model handling the inherent uncertainties pertaining to
correlated electricity demands and wind power generations as
and lines should be constructed throughout the system to supply well as availability of units and lines.
adequately the forecasted electricity demand during the plan-
ning horizon [1]. Due to excessive computational burden of co- B. Contributions
ordinated generation and transmission expansion planning and The main contributions of the paper are:
DEHGHAN et al.: RELIABILITY-CONSTRAINED ROBUST POWER SYSTEM EXPANSION PLANNING 2385

1) Joint consideration of continuous and discrete uncertainty between modeling accuracy and computation tractability [6].
sources. Previous research works in the area have used Without loss of generality, it is assumed that the location and
either bounded intervals [13]–[15] or probability distribu- capacity of wind farms are specified.
tions [4]–[7], [10] to characterize continuous and discrete
uncertain variables. Accordingly, there is no research B. Uncertainty Characterization
work characterizing the continuous uncertainties per- The intrinsic uncertainties pertaining to electricity demands
taining to correlated electricity demands and wind power and wind power generations as well as availability of equipment
generations by bounded intervals, through a polyhedral are considered in the proposed SEP model. Specifically, wind
uncertainty set, and the discrete uncertainties pertaining to power generation is not stochastic within a yearly window. In
availability of equipment by probability distributions. In other words, the profile of wind power generation (acquired by
addition, the -means clustering technique is used in this dividing the wind power generation by the capacity of wind
paper to construct a polyhedral uncertainty set pertaining farms) remains unaltered from year to year. Needless to say,
to the correlated electricity demands and wind power many scenarios are needed to characterize wind generation
generations. It is worth mentioning that the -means clus- throughout one year, but there are no changes from year to year.
tering technique has been previously used in [6] and [10] Analogous to wind power generation, the profile of electricity
to generate deterministic operation conditions considering demand (acquired by dividing the electricity demand by the
fixed values for the correlated electricity demands and peak demand of each year) is assumed to remain unaltered
wind power generations. On the contrary, such technique during one whole year and many scenarios are needed to
is used in this paper to generate non-deterministic opera- characterize it throughout one year [10]. However, the yearly
tion conditions considering bounded intervals rather than forecasted growth rate of demand is subject to uncertainty.
fixed values for the correlated electricity demands and We characterize the uncertainty pertaining to the profile of
wind power generations. electricity demand and wind power generation in terms of
2) Hybridizing SP and RO modeling to find a robust expan- uncertainty sets [18] rather than scenarios. Accordingly, we
sion plan satisfying a probabilistic reliability constraint, can find a flexible and robust expansion plan withstanding the
i.e., expected energy not supplied (EENS), under the worst-case realization of these uncertainty sources.
worst-case realization of correlated electricity demands Since electricity demand and wind power generation are sta-
and wind power generations. To the best of the authors' tistically correlated [19], the -means clustering technique [10]
knowledge, there is no work reported in the literature is used to construct a polyhedral uncertainty set. In this study,
jointly characterizing continuous and discrete uncertain the yearly historical data sets of wind/demand profiles are parti-
variables by hybridizing the notions of RO and SP. tioned by the -means clustering technique into a finite number
3) Introducing a new tri-level decomposition algorithm to of clusters [19]. The centroid of each cluster is chosen as the
solve the proposed reliability-constrained robust SEP nominal estimate of the electricity demand and wind power gen-
problem. The proposed decomposition algorithm employs eration in each operation condition (i.e., and ). Also,
primal and dual cutting planes submitted to the first level -quantile and -quantile of empirical cumulative prob-
expansion master problem by the second level operation ability distributions pertaining to electricity demand and wind
and third level reliability sub-problems, respectively. Note power generation of each cluster (e.g., for ) are chosen
that the traditional Benders decomposition only employs as lower and upper estimates of electricity demand and wind
dual cutting planes [17]. On the contrary, the proposed power generation in each operation condition (i.e.,
tri-level decomposition algorithm simultaneously benefits and ). In addition, the duration time of each oper-
from dual and primal cutting planes. ation condition (i.e., ) is calculated by dividing the total
number of historical data samples belonging to each cluster over
C. Paper Organization the total number of historical data samples. On the other hand,
The rest of the paper is organized as follows. In Section II the the discrete uncertainty pertaining to availability of equipment
main features of the proposed SEP model are described. In are characterized by their forced outage rate (FOR) [3]. In line
Section III the mathematical formulations of deterministic and with previous research works in the area [3]–[8], short-term op-
non-deterministic models are introduced. In Section IV, sim- erational uncertainties, such as the uncertainties pertaining to
ulation results comparing deterministic and non-deterministic short-term load forecast, are not considered in this paper. Note
SEP models are provided and discussed. Finally, the main that the time framework of the operational uncertainty sources
conclusions are given in Section V. (e.g., one to several hours) is completely different from the time
framework of the planning uncertainties considered in the paper
II. PROBLEM DESCRIPTION (e.g., one to several years).

A. Model Assumptions III. MATHEMATICAL FORMULATION

The proposed SEP problem is solved from the viewpoint of
the ISO, or the regulator. Accordingly, it seeks to find a co-ordi- A. Deterministic Model
nated generation and transmission expansion plan which is op- To provide a compact formulation, all units and lines are con-
timal for the society [6]. Due to the inherent complexity of a reli- sidered as candidates to expand the power system [16]. Accord-
ability-constrained non-deterministic SEP problem, a single-pe- ingly, for existing units and lines, investment related costs and
riod static model is considered constituting a suitable trade-off variables are fixed on 0 and 1, respectively (i.e., ,
2386 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 3, MAY 2016

, , and ). However, in the implemented

computer codes, existing units and lines are not considered as
decision variables with fixed values. Instead, they are consid-
ered as constants. Therefore, this assumption does not affect
the computation burden of the proposed planning model. Given
and , obtained from -means clustering
technique, as well as and , the deterministic SEP
(DSEP) model is

(1a)

(1b)
(1c)

(1d)
Fig. 1. Tri-level decomposition algorithm to solve the RSEP problem.
(1e)
(1f)
(1g) The objective function (2a) represents (1a). Also, (2b) corre-
(1h) sponds to (1h) and (1i), (2c) corresponds to (1b) and (1g), (2d)
corresponds to (1c)–(1f), and (2e) corresponds to (1j). Note that
(1i)
the equality constraint can be stated in terms of two op-
(1j) posite inequalities and . In (2d), , where is
a vector of ones. As the DSEP model only considers one real-
The objective function (1a) minimizes the total investment ization of the uncertainty sources, its expansion plan may not be
costs of candidate units and lines, total operation costs of units, sufficiently robust to withstand different realizations of the un-
and total penalty costs due to unserved electricity demands. certain variables. To remedy this limitation and find a robust and
Constraint (1b) states the power balance for each bus in every flexible expansion plan, the reliability-constrained robust SEP
operating condition. Constraints (1c)–(1e) represent the power (RSEP) model handling the uncertain nature of electricity de-
flow for every line and its limits, respectively, where is a suf- mands, wind power generations, and availability of equipment
ficiently large positive constant (i.e., [14]). Con- is introduced as follows:
straint (1f) limits the generation of units. Also, constraint (1g)
bounds the maximum value of unserved electricity demand for (3)
each bus. Constraint (1h) bounds phase angles. Constraint (1i)
sets the phase angle of reference bus on zero. Finally, constraint
(1j) represents the integrality conditions for investment vari- where ,
ables. In the deterministic model given by (1a)–(1j), and , and
indicate investment related variables and , , , , and .
indicate operation related variables. Min-max-min problem (3) minimizes the summation of
investment and operation costs subject to investment and
B. Non-Deterministic Model operation constraints (i.e., and ) under the worst-case
realization of uncertain electricity demands and wind power
The above-mentioned DSEP model can be rewritten in the generations (i.e., and ). Although min-max-min plan-
compact form given in the following: ning models have been solved by two-level decomposition
algorithms in [13]–[15], these algorithms cannot be used here
(2a) due to incorporating the probabilistic reliability constraint
(2b) of in (3). Accordingly, a
tri-level decomposition algorithm, outlined in Fig. 1, employing
(2c)
primal operation and dual reliability cuts is proposed in this
(2d) paper. Next, the investment master problem as well as the
(2e) operation and reliability sub-problems are presented.
DEHGHAN et al.: RELIABILITY-CONSTRAINED ROBUST POWER SYSTEM EXPANSION PLANNING 2387

1) Investment Master Problem: This problem finds the op- in (11) can be replaced by their equivalent expressions given
timum expansion plan considering the primal operation cuts, in (14) and (15), respectively. Thus, T2 in (11) can be rewritten
i.e., (5)–(7), and dual reliability cuts, i.e., (8): as follows:

(4)
Ξ (16)
(5)
As (16) only includes multiplication of binary and continuous
(6) variables, its nonlinear terms can be recast into linear ones using
(7) a set of new binary and continuous variables in addition to the
disjunctive constraints as given in the following:
(8)
(9) (17)
where Ξ . After solving the MILP where
investment master problem, the investment variables are fixed
in the operation and reliability sub-problems (i.e., ). Also, the
in Fig. 1 is updated in terms of the optimal solution of the
master problem obtained at iteration (i.e., ).
2) Operation Sub-Problem: The operation sub-problem only
considers the uncertainty pertaining to electricity demands and
(18)
wind power generations. Given the optimal investment vari-
ables obtained by solving the master problem (i.e., ),
Also, , , , and are auxiliary continuous mod-
the operation sub-problem is
eling variables. The linear terms in (17) function analogously to
(10) the nonlinear terms in (16) where . Consequently, the
linearized version of the bilinear operation sub-problem can be
rewritten as follows:
where
. Considering the dual of the inner minimization
problem of (10), this max-min problem can be rewritten as a (19)
Ξ
maximization model given in the following:
(20)
(11)
where Ξ
, and T2 is as given in (17). Also, the robustness
(12) of the operation sub-problem can be controlled by adding the
(13) following constraint (21) to the MILP model of (19) and (20):

where , , and represent dual variables pertaining to con- (21)

straints , , ,
respectively. Due to multiplication of optimization variables in where the binary variables , , , and represent the
T2 of (11) (i.e., and as well as and ), (11)–(13) rep- elements of binary vectors , , , and , respectively.
resent a bilinear optimization problem. However, the optimal Note that can adopt different integer values between zero and
solution of the operation sub-problem will be on the extreme the number of uncertain variables belonging to . Also, the
points of . Accord- binary modeling technique of , given in (14) and (15), can be
ingly, an exact MILP reformulation of (11)–(13) only exploring extended to consider non-integer values of at the expense of
the extreme points of the operation sub-problem is given in the more complexity. After solving the exact MILP reformulation
following. The uncertain variables of at any extreme point of the operation sub-problem, given in (19)–(20), the worst-case
can be modeled as realization of the uncertain electricity demands and wind power
generations at iteration (i.e., and ) can be obtained
in terms of (14) and (15). Also, the in Fig. 1 is updated
(14) in terms of the optimal solution of the operation sub-problem
and investment master problem obtained at iteration (i.e.,
(15) with and given in (4) and (9),
respectively).
where , , , and are auxiliary binary modeling vari- If , the reliability sub-problem (i.e., the third
ables. According to (14) and (15), the uncertain variables of level of the proposed tri-level decomposition algorithm) should
may adopt either their lower (e.g., ), nominal (e.g., be solved considering the optimal investment variables (i.e.,
), or upper (e.g., ) estimates. Therefore, and obtained in the first level) as well as the worst-case realization of
2388 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 3, MAY 2016

the uncertainty electricity demands and wind power generations TABLE I

(i.e., and obtained in the second level); otherwise, EXISTING UNITS FOR GARVER TEST SYSTEM
and should be submitted to the investment master problem
to add an additional primal operation cut, given in (5)–(7), as
shown in Fig. 1.
3) Reliability Sub-Problem: The reliability sub-problem
only considers the uncertainty pertaining to the availability
of equipment. In a real-world power system, there is a large
number of states for the availability of equipment and their enu-
meration is a very hard task (if not impossible). Accordingly,
state sampling in terms of Monte Carlo simulation (MCS)
[3] is used in the reliability sub-problem rather than state
enumeration. Given the optimal investment variables obtained
by solving the investment master problem (i.e., ) as well as TABLE II
the worst-case realization of uncertainty electricity demands EXISTING LINES FOR GARVER TEST SYSTEM
and wind power generations obtained by solving the operation
sub-problem (i.e., and ), the reliability sub-problem
simulates the availability of equipment aiming to estimate the
value of as follows:
Step 1) Set and .
Step 2) Set and employ MCS to simulate the avail-
ability of equipment for sample (i.e., ) consid-
ering their FORs [3].
Step 3) Solve the linear-programming (LP) load-curtailment
problem of (22)–(26) for sample of MCS:

(22)
(23) IV. CASE STUDIES
(24) In this section, the proposed RSEP methodology is used to
(25) expand the generation and transmission system of a modified
(26) Garver 6-bus test system [20], the IEEE 24-bus, and the IEEE
73-bus reliability test systems (RTS) [21] where . All
where, represents a vector of auxiliary variables case studies have been run using CPLEX within GAMS on a
and represents the vector of dual variables for the 64-bit windows-based server with 60 GB of RAM and 24 Intel
constraint of sample . By solving the LP Xenon processors clocking at 3.33 GHz.
problem for sample , the objective function of (22)
A. Garver 6-Bus Test System
gives .
Step 4) Calculate the value of and as given in The modified Garver 6-bus test system includes 10 units and
(27) and (28). If the coefficient of variation [12] for 8 lines as shown in Tables I and II, respectively. Moreover,
the MCS is less than , stop the algorithm and re- 72 candidate units and 45 transmission lines are considered.
port and ; otherwise, go to Step 2, gen- The characteristics of candidate units and lines are shown in
erate another sample , and continue the algorithm: Tables III and IV, respectively. In this study, the forecast of peak
electricity demand for each bus is assumed to be 2.5 times of the
current value given in [20]. Also, two wind farms with 40-MW
(27)
and 20-MW capacities are located at buses 1 and 5, respec-
tively. Wind units are not investment candidates, as mentioned
in Section II-A. The polyhedral uncertainty set of electricity de-
(28)
mand and wind power generation is constructed using the his-
torical data pertaining to the power system of Texas during year
After solving the reliability sub-problem, the probabilistic 2005 based on the method described in Section II-B with
reliability constraint (i.e., ) should be . The employed electricity demand profiles of buses 1,
checked. If , the decomposition algorithm 2, 3, 4, and 5 belong to North, East, Coast, West, and South re-
stops; otherwise, and are submitted to the invest- gions of electricity reliability council of Texas (ERCOT) [22],
ment master problem to incorporate a new dual reliability cut, respectively. Moreover, it is assumed that the wind farms lo-
as given in (8) and shown in Fig. 1. By terminating the proposed cated at buses 1 and 5 are equipped with 2-MW wind units
tri-level decomposition algorithm, the obtained optimal expan- (Vestas V80/2000 with a hub height of 80 m) and their wind
sion plan, i.e., , and its associated investment and operation power generation profiles belong to Northwestern and South-
costs are reported. eastern flat lands of Texas [23], respectively.
DEHGHAN et al.: RELIABILITY-CONSTRAINED ROBUST POWER SYSTEM EXPANSION PLANNING 2389

TABLE III TABLE V

CANDIDATE UNITS FOR GARVER TEST SYSTEM OPTIMAL SOLUTION VERSUS DEGREE OF ROBUSTNESS
FOR GARVER TEST SYSTEM

TABLE IV
CANDIDATE LINES FOR GARVER TEST SYSTEM

TABLE VI
OPTIMAL SOLUTION VERSUS FOR GARVER TEST SYSTEM

1) Optimal Solution versus Degree of Robustness: Given 2) Optimal Solution versus : Given and
10 operation conditions for the 5 electricity demand buses 10 operation conditions for correlated electricity demands and
and the 2 wind power generation buses, can adopt different wind power generations, the optimal expansion plans obtained
integer values between 0 and (i.e., the from RSEP-1 are provided in Table VI where of the
total number of uncertain variables belonging to ). The previous case (i.e., 0.01) is varied leading to 0.015 and
optimal expansion plans obtained from DSEP and RSEP are 0.005 of AED. According to Table VI, the expansion cost is
illustrated in Table V where is equal to 0, 35, and 70 in increased from $175.289 M to $187.352 M by decreasing the
RSEP-1, RSEP-2, and RSEP-3, respectively, and is value of from 0.015 of AED to 0.005. In other words,
0.01 of the annual energy demand (AED). In Table V, a more reliable expansion plan is obtained by adding more new
and denote number of operation and reliability cuts in units and lines at the expense of more expansion cost. Similar re-
the proposed tri-level decomposition algorithm required to sults are obtained from RSEP-2 and RSEP-3 for this case, where
find the optimal expansion plan, respectively. The expansion is set on 50% and 100% of its maximum value, respectively.
cost monotonically increases from DSEP to RSEP-3. The 3) Optimal Solution versus Operation Conditions: In this
lowest expansion cost is imposed by DSEP as it ignores all case, the initial 10 operation conditions are increased to 20 and
uncertainties. The expansion cost obtained by RSEP-1 is more 30. Considering 5 electricity demand and 2 wind power genera-
than DSEP because RSEP-1 includes the uncertainty pertaining tion buses, the maximum value of becomes
to the availability of equipment and requires satisfying the and , respectively. The optimal expansion
probabilistic reliability constraint. However, RSEP-1 excludes plans obtained from RSEP-2 for 20 and 30 operation conditions
the uncertainty pertaining to electricity demand and wind are illustrated in Table VII where adopts 50% of its max-
power generation as it considers . Accordingly, its imum allowable value. In this table, denotes number of op-
expansion cost is less than RSEP-2 and RSEP-3, which include eration conditions and the probabilistic reliability constraint is
the uncertainties pertaining to electricity demands and wind relaxed for simplicity leading to . The expansion plan
power generations in addition to satisfying the probabilistic of RSEP-2 for is relatively different with respect to
reliability constraint. Moreover, the expansion cost imposed by as more operation conditions model more realistically
RSEP-3 is larger than that of RSEP-2 because is equal to 70 the uncertainty pertaining to correlated electricity demands and
and 35 in RSEP-3 and RSEP-2, respectively. In other words, wind power generations.
RSEP-3 is the most expensive expansion plan as it satisfies the
probabilistic reliability constraint under the most conservative B. IEEE 24-Bus Test System
realization of uncertain electricity demands and wind power In the previous case studies, wind units are not considered
generations. as investment candidates. However, the proposed deterministic
2390 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 3, MAY 2016

TABLE VII TABLE X

OPTIMAL SOLUTION VERSUS OPERATION CONDITIONS OPTIMAL SOLUTION VERSUS DEGREE OF ROBUSTNESS FOR IEEE 24-BUS RTS
FOR GARVER TEST SYSTEM

TABLE VIII
CANDIDATE UNITS FOR IEEE 24-BUS RTS

TABLE IX
CANDIDATE LINES FOR IEEE 24-BUS RTS

to Southwestern, Southeastern, and Northwestern flat lands of

Texas in year 2005 [23], respectively. Given 5 operation condi-
and non-deterministic formulations can easily be modified to tions for 17 electricity demand buses and 3 wind power gener-
consider wind units as investment candidates. For the deter- ation buses, can adopt different integer values between 0 and
ministic model, the investment cost of the candidate wind units . For the sake of conciseness, only the re-
should be added to the objective function given in (1a). Also, sults of Section IV-A1, i.e., optimal solution versus degree of
in (1b) should be multiplied by an integer variable indi- robustness, are presented here.
cating the number of constructed wind units at bus . Similarly, The optimal expansion plans obtained from DSEP, RSEP-1
the non-deterministic model can be constructed. To further eval- with , RSEP-2 with and RSEP-3 with are
uate the performance of the proposed planning tool with and illustrated in Table X. The expansion cost imposed by RSEP-1
without considering wind units as investment candidates, we is slightly more than DSEP as RSEP-1 considers the uncertainty
analyze a modified IEEE 24-bus RTS including 32 units and 33 source of equipment unavailability and satisfies the probabilistic
lines [21]. The maximum capacity of lines is assumed to be 50% reliability constraint. By including the uncertainty sources of
of the original values. Moreover, 96 candidate units (without electricity demands and wind power generations, much more
considering candidate wind units) and 99 candidate lines are expensive expansion plans are obtained by RSEP-2 and RSEP-3
considered. In other words, there are three candidate units with with higher cost of RSEP-3 due to providing a more robust ex-
the same capacity of each existing unit and three candidate lines pansion plan against these uncertainty sources.
with the same capacity of each existing line. The characteristics 2) Including Wind Units as Investment Candidates: In this
of candidate units and lines are shown in Tables VIII and IX, case, the wind units of wind farms are also investment can-
respectively. In all case studies, the forecast of peak electricity didates where their annualized investment cost is kW.
demand for the Northern and Southern areas of the IEEE 24-bus It is assumed that up to 20 candidate wind units can be con-
RTS are considered as 1.5 and 2.5 times of the current values, structed at each bus of the IEEE 24-bus test system. Note that
respectively. The employed electricity demand profiles of buses the predetermined wind farms of the previous case are not con-
located in the Northern and Southern areas of the IEEE 24-bus sidered here. The employed wind power generation profiles of
RTS belong to North and South regions of ERCOT in year 2005, candidate wind units located in the Northern and Southern areas
respectively. The polyhedral uncertainty sets for all case studies of the IEEE 24-bus RTS belong to Southeastern and North-
are constructed as described for the previous case study. Also, western flat lands of Texas in year 2005 [23], respectively. For
it is assumed that all wind farms are equipped with the 2-MW brevity, only the results of Section IV-A3, i.e., optimal solution
wind units of the previous case study. For these case studies, versus operation conditions, are presented here. Accordingly,
is set to 0.01 of AED. the initial 5 operation conditions of the previous case are in-
1) Excluding Wind Units as Investment Candidates: In this creased to 20. Considering 17 electricity demand buses and 2
case, wind units are not investment candidates and it is assumed wind power generation areas (i.e., the Northern and Southern
that two wind farms with 200-MW capacity are located at buses areas of the IEEE 24-bus RTS), the maximum value of for 5
3 and 6, and a wind farm with 100-MW capacity is located at and 20 operation conditions becomes and
bus 20. Moreover, their wind power generation profiles belong , respectively. The optimal expansion plans
DEHGHAN et al.: RELIABILITY-CONSTRAINED ROBUST POWER SYSTEM EXPANSION PLANNING 2391

TABLE XI TABLE XII

OPTIMAL SOLUTION VERSUS OPERATION CONDITIONS FOR IEEE 24-BUS RTS OPTIMAL SOLUTION VERSUS DEGREE OF ROBUSTNESS FOR IEEE 73-BUS RTS

0.01 of AED. Given 5 operation conditions for 51 electricity de-

mand buses and 9 wind power generation buses, can adopt dif-
ferent integer values between 0 and . The op-
timal expansion plans for DSEP, RSEP-1 with , RSEP-2
with , and RSEP-3 with , are illustrated in
Table XII. The expansion cost imposed by RSEP-1 is similar to
DSEP as the optimal solution obtained from RSEP-1 satisfies
the probabilistic reliability constraint at the first iteration and ig-
nores the uncertainty pertaining to electricity demand and wind
power generation. By including the uncertainty sources of elec-
tricity demands and wind power generations, much more expen-
sive expansion plans are obtained by RSEP-2 and RSEP-3 with
higher cost of RSEP-3 due to higher immunization against the
uncertainty sources provided by RSEP-3. Also, the computation
obtained from RSEP-2 for 5 and 20 operation conditions are il- time increases from DSEP to RSEP-3 as shown in Table XII.
lustrated in Table XI where adopts 50% of its maximum al-
lowable value. According to Table XI, wind units are added for V. CONCLUSION
while no wind unit is added for . Also, the This paper presents a reliability-constrained robust SEP
total investment and operation costs of RSEP-2 for is model handling the uncertainty sources pertaining to electricity
less than . This is due to the fact that more operation demand and wind power generation as well as availability
conditions can more realistically characterize the uncertainties of equipment. The uncertain nature of correlated electricity
pertaining to the correlated electricity demands and wind power demand and wind power generation are modeled by bounded
generations. The average computation time of RSEP models for intervals. Also, the availability of equipment is characterized by
Garver and RTS test systems is about 1.5 h and 11 h, respec- probability distributions. Accordingly, the proposed approach
tively, measured on the previously mentioned hardware set. is capable to find a minimum cost expansion plan satisfying
a probabilistic reliability constraint under the worst-case
C. IEEE 73-Bus Test System realization of the uncertain variables. As the proposed relia-
bility-constrained minimax cost SEP model cannot be solved
To evaluate the scalability of RSEP, it is tested on the larger directly by available commercial optimization packages, a new
test case: the modified IEEE 73-bus RTS including 96 units and tri-level decomposition algorithm is also introduced in this
120 lines [21]. Analogous to the modified IEEE 24-bus RTS, paper to solve it. The effects of the degree of robustness, target
there are three candidate units with the same capacity of each value of EENS, and number of operation conditions on the
existing unit and three candidate lines with the same capacity optimal robust expansion plan are studied through the proposed
of each existing line in the modified IEEE 73-bus RTS. Accord- RSEP model. Considering reliability in its real probabilistic
ingly, 288 units and 360 lines are considered. Also, it is assumed nature as well as including more uncertainty sources leads to
that six 200-MW wind farms are located at buses 3, 6, 27, 30, 51, a more reliable and robust expansion plan at the expense of
and 52, and three 100-MW wind farms are located at buses 20, higher investment and operation costs.
44, and 68, respectively. All wind farms are equipped with the Comparing RO, SP, and the proposed approach (hybridizing
2-MW wind units of the previous case studies. Moreover, their RO and SP), it can be said that:
wind power generation profiles belong to Southwestern (buses 1) RO finds an optimal solution immunized against the
3, 27, and 51), Southeastern (buses 6, 30, and 52), and North- worst-case realization of uncertain variables and the con-
western (buses 20, 44, and 68) flat lands of Texas in year 2005 servatism of the optimal solution can be controlled by
[23]. No wind unit is considered as expansion candidate in this changing the degree of robustness. Also, SP finds an op-
case. The forecast of peak electricity demand for the Northern timal solution that is optimal on average for all scenarios
and Southern areas of the IEEE 73-bus RTS are considered as considered and the accuracy of the optimal solution can
1.5 and 2.5 times of the current values, respectively. Moreover, be controlled by changing the number of scenarios. It is
their employed electricity demand profiles belong to North and worth mentioning that the optimal solution of previous
South regions of ERCOT in year 2005, respectively [22]. The SP-based planning tools usually is optimal on average for
polyhedral uncertainty sets for all case studies are constructed all scenarios considered without satisfying a probabilistic
as described for the previous case study. Also, is set to reliability constraint such as EENS [5]. On the contrary,
2392 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 31, NO. 3, MAY 2016

the proposed reliability-constrained robust planning tool is [13] B. Chen, J. Wang, L. Wang, Y. He, and Z. Wang, “Robust optimiza-
capable of finding an optimal solution, immunized against tion for transmission expansion planning: Minimax cost vs. minimax
regret,” IEEE Trans. Power Syst., vol. 29, no. 6, pp. 3069–3077, Nov.
worst-case realization of uncertain electricity demands 2014.
and wind power generations, satisfying a probabilistic [14] R. A. Jabr, “Robust transmission network expansion planning with un-
reliability constraint. Additionally, the proposed approach certain renewable generation,” IEEE Trans. Power Syst., vol. 28, no.
4, pp. 4558–4567, Nov. 2013.
can control the robustness and reliability of the optimal [15] S. Dehghan, N. Amjady, and A. Kazemi, “Two-stage robust genera-
solution by means of two parameters named the degree tion expansion planning: A mixed integer linear programming model,”
of robustness and the target value of EENS. Accordingly, IEEE Trans. Power Syst., vol. 29, no. 2, pp. 584–597, Mar. 2014.
[16] R. L. Y. Chen, A. Cohn, N. Fan, and A. Pinar, “Contingency-risk in-
it is more flexible than solution methods based solely on formed power system design,” IEEE Trans. Power Syst., vol. 29, no. 5,
either RO or SP. pp. 2087–2096, Sep. 2014.
2) Joint consideration of continuous and discrete uncertain [17] A. J. Conejo, E. Castillo, R. Minguez, and R. Garcia-Bertrand, Decom-
variables using either RO or SP may result in intractability position Techniques in Mathematical Programming. New York, NY,
USA: Springer-Verlag, 2006.
of the optimization problem. Accordingly, RO has been [18] Y. Guan and J. Wang, “Uncertainty sets for robust unit commitment,”
previously used in the planning to only characterize either IEEE Trans. Power Syst., vol. 29, no. 3, pp. 1439–1440, May 2014.
continuous [13]–[15] or discrete [16] uncertain variables. [19] L. Baringo and A. J. Conejo, “Correlated wind-power production and
electric load scenarios for investment decisions,” App. Energy, vol.
On the contrary, SP has been previously used for joint 101, pp. 475–482, Jan. 2013.
consideration of continuous and discrete uncertain vari- [20] L. L. Garver, “Transmission network estimation using linear pro-
ables, but in small-scale test systems and not medium- and gramming,” IEEE Trans. Power App. Syst., vol. PAS-89, no. 7, pp.
1688–1697, Sep. 1970.
large-scale test systems [5]. Therefore, jointly considering [21] “The reliability test system task force, IEEE reliability test system,”
continuous and discrete uncertain variables of the planning IEEE Trans. Power Syst., no. 3, pp. 1010–1020, Aug. 1999.
model through either RO or SP is computationally cumber- [22] ERCOT [Online]. Available: http://www.ercot.com
[23] Alliance for Sustainable Energy System Advisor Model version, De-
some. On the other hand, the proposed hybrid approach can partment of Energy (DOE), National Renewable Energy Laboratory
more effectively characterize both continuous and discrete (NREL), Jan. 14, 2014.
planning uncertainties than either RO or SP.
Shahab Dehghan (M'14) received the M.S. and
Ph.D. degrees (with honors) in electrical engineering
REFERENCES from Iran University of Science and Technology,
[1] X. Wang and J. R. McDonald, Modern Power System Planning. New Tehran, Iran, in 2009 and 2014, respectively.
York, NY, USA: McGraw-Hill, 1994. He is currently an assistant professor at Qazvin Is-
[2] R. Hemmati, R. A. Hooshmand, and A. Khodabakhshian, “Compre- lamic Azad University, Qazvin, Iran.
hensive review of generation and transmission expansion planning,”
IET Gener. Transm. Distrib., vol. 7, no. 9, pp. 955–964, Sep. 2013.
[3] W. Li and R. Billinton, “A minimum cost assessment method for com-
posite generation and transmission system expansion planning,” IEEE
Trans. Power Syst., vol. 8, no. 2, pp. 628–635, May 1993.
[4] J. H. Roh, M. Shahidehpour, and L. Wu, “Market-based generation and
transmission planning with uncertainties,” IEEE Trans. Power Syst.,
vol. 24, no. 3, pp. 1587–1598, Aug. 2009.
[5] P. Jirutitijaroen and C. Singh, “Reliability constrained multi-area ad- Nima Amjady (SM'09) received the B.S., M.S., and
equacy planning using stochastic programming with sample-average Ph.D. degrees in electrical engineering from Sharif
approximations,” IEEE Trans. Power Syst., vol. 23, no. 2, pp. 504–513, University of Technology, Tehran, Iran, in 1992,
May 2008. 1994, and 1997, respectively.
[6] R. Dominguez, A. J. Conejo, and M. Carrion, “Toward fully renewable At present, he is a full professor with Semnan Uni-
electric energy systems,” IEEE Trans. Power Syst., vol. 30, no. 1, pp. versity, Semnan, Iran.
316–326, Jan. 2015.
[7] J. A. Lopez, K. Ponnambalam, and V. H. Quintana, “Generation and
transmission expansion under risk using stochastic programming,”
IEEE Trans. Power Syst., vol. 22, no. 3, pp. 1369–1378, Aug. 2007.
[8] B. G. Gorenstin, N. M. Campodonico, J. P. Costa, and M. V. F. Pereira,
“Power system expansion planning under uncertainty,” IEEE Trans.
Power Syst., vol. 8, no. 1, pp. 129–136, Feb. 1993.
[9] J. R. Birge and F. Louveaux, Introduction to Stochastic Program-
ming. New York, NY, USA: Springer, 1997. Antonio J. Conejo (F'04) received the M.S. degree
[10] L. Baringo and A. J. Conejo, “Risk-constrained multi-stage wind power from the Massachusetts Institute of Technology,
investment,” IEEE Trans. Power Syst., vol. 28, no. 1, pp. 401–411, Feb. Cambridge, MA, USA, in 1987 and the Ph.D. degree
2013. from the Royal Institute of Technology, Stockholm,
[11] A. Ben-Tal, L. E. Ghaoui, and A. Nemirovski, Robust Optimization. Sweden, in 1990.
Princeton, NJ, USA: Princeton Univ. Press, 2009. He is currently a full professor at The Ohio State
[12] S. Dehghan, A. Kazemi, and N. Amjady, “Multi-objective robust trans- University, Columbus, OH, USA.
mission expansion planning using information-gap decision theory and
augmented -constraint method,” IET Gener. Transm. Distrib., vol. 8,
no. 5, pp. 828–840, May 2014.