IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 6, NO.
3, SEPTEMBER 2019
Worst-Case Load Shedding in
Electric Power Networks
Fu Lin , Member, IEEE
Abstract—We consider the worst-case load-shedding
problem in electric power networks where a number of
transmission lines are to be taken out of service. The ob-
jective is to identify a prespecified number of line outages
that lead to the maximum interruption of power generation
and load at the transmission level, subject to the active
power-flow model, the load and generation capacity of the
buses, and the phase-angle limit across the transmission
lines. For this nonlinear model with binary constraints, we
show that all decision variables are separable except for the
nonlinear power-flow equations. We develop an iterative de-
composition algorithm, which converts the worst-case load-
shedding problem into a sequence of small subproblems.
We show that the subproblems are either convex problems
that can be solved efficiently or nonconvex problems that
have closed-form solutions. Consequently, our approach is
scalable for large networks. Furthermore, we prove conver-
gence of our algorithm to a critical point, and the objective
value is guaranteed to decrease throughout the iterations.
Numerical experiments with IEEE test cases demonstrate
the effectiveness of the developed approach.
Index Terms—Load shedding, power systems, vulnera-
bility analysis.
I. INTRODUCTION
EDUNDANCY of interconnection in power systems is
R known to help prevent cascade blackouts [1]. On the other
hand, recent studies suggest that excessive interconnectivity in
power grids can result in overloading the network capacity,
which, in turn, contributes to more severe blackouts [2]. There-
fore, a balance between the operational security and the network
interconnectivity is important for power grid operations.
Traditionally, contingency analysis in power grids has fo-
cused on the severity of line outages using linearized power-flow
models [3]. Recent years have seen vulnerability analysis of line
outages using nonlinear power-flow models [4]–[6]. Following
this line of research, we study the worst-case load-shedding
problem. Our objective is to identify a small number of trans-
mission lines, whose removal leads to the maximum damage
to the power systems. This problem contains binary decision
Manuscript received October 19, 2018; accepted May 25, 2019. Date
of publication August 15, 2019; date of current version September 17,
2019. This work was supported in part by the U.S. Department of
Energy under Contract DE-AC02-06CH11357 and in part by the Ad-
vanced Research Projects Agency Energy (ARPA-E) under Contract
DE-AR0000700. Recommended by Associate Editor B. Lesieutre.
The author is with the Autonomous and Intelligent Systems Depart-
ment, United Technologies Research Center, East Hartford, CT 06108
USA (e-mail:, linf@utrc.utc.com).
Digital Object Identifier 10.1109/TCNS.2019.2935559
2325-5870 © 2019 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.
Authorized licensed use limited to: Universitas Indonesia. Downloaded on July 25,2021 at 12:02:35 UTC from IEEE Xplore. Restrictions apply.
1269
variables for taking lines offline and the nonlinear power-flow
equations. As a result, it falls into the class of mixed-integer
nonlinear programs (MINLPs), which is beyond the capacity
of the state-of-the-art MINLP solvers even for small power
systems.
Our contributions can be summarized as follows. First, the
worst-case load-shedding model incorporates the nonlinear, ac-
tive power-flow equations, the generation and load capacities of
the buses, and the thermal constraints across the transmission
lines. This model is capable of providing more accurate oper-
ating conditions than models based on the linearized power-
flow equations. Second, we show that the decision variables
for taking transmission lines offline and for the generation,
load, and phase angles across buses are separable except for
the power-flow constraints. By exploiting this separable struc-
ture, we develop an algorithm that decomposes the worst-case
load-shedding problem into a sequence of subproblems. It turns
out that these subproblems are either convex problems or non-
convex problems that have closed-form solutions. As a result,
our approach is scalable for large networks. Third, we prove the
convergence of our algorithm to a critical point of the noncon-
vex problem. Furthermore, the objective value is proved to be
monotonically decreasing throughout the iterations. Our proof
techniques build upon convergence results from the proximal
alternating linearization method (PALM).
In our previous work [7], the alternating direction method of
multipliers (ADMM) is proposed to deal with the optimal load-
shedding problem with the linearized power-flow model. The
shortcoming of ADMM is that there is no theoretical guarantee
of convergence for nonconvex problems. In contrast, PALM
allows us to handle nonconvex, nonsmooth problems with a
provably convergence guarantee. PALM has found applications
in learning the low-complexity model with steady-state data [8],
[9], in sparse control of networked systems [10], and in the
codesign of output feedback controllers [11].
There is a large body of work on the load-shedding problem
in electric power networks [12]–[19]. We next provide a brief
literature review and put our contributions in context.
In [12] and [13], a bilevel optimization framework is proposed
to identify the critical components of the grid. While this bilevel
framework exposes security issues in long-term operation of
the power grid, efficient methods have yet to be developed for
the underlying mathematical problems. In contrast, we focus
on the worst-case impact with respect to a specified operating
point. As a result, we develop an efficient decomposition algo-
rithm by exploiting the separable structure.
1270 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 6, NO. 3, SEPTEMBER 2019

Another line of work studies efficient numerical methods for to the sum of generation
the load-shedding problem [14], [16]–[18]. In [14], a discretiza-
1T P = 0
tion technique is developed to convert the differential equations
to algebraic constraints. The resulting nonlinear programming where 1 is the vector of all ones.
(NLP) problem is solved by using standard NLP solvers. In [16], Let γ ∈ {0, 1}m denote whether a line is in service or not:
Newton’s method is employed to minimize the curtailment of γl = 1 if line l is in service and γl = 0 if line l is out of ser-
load service after severe faults. In [17], a quasi-Newton method vice. Let z = [zdT zgT ]T ∈ Rn , where zd ≥ 0 and zg ≤ 0 are
is proposed for the load-shedding problem with voltage and fre- the load-shedding vector and the generation reduction vector,
quency characteristics of load. In contrast to these NLP-based respectively. It follows that:
approaches, our formulation incorporates binary decision vari-
ables to model line removals in power networks. Thus, it falls P d ≤ P d + zd ≤ 0
in the class of a more general class of MINLP problems. where the upper bound 0 enforces Pd + zd to be a load vector.
Heuristics approaches are proposed for the nonconvex load- Similarly
shedding problem [15], [19]. In [15], a particle swarm-based
simulated annealing technique is introduced for the under- 0 ≤ P g + zg ≤ P g
voltage load-shedding problem. In [19], tree-like heuristics where the lower bound 0 enforces Pg + zg to be a generator
strategies are proposed for emergency situations to maintain vector. Since the load shed must be equal to the generation
reliability. In contrast to these heuristics approaches with no the- reduction, we have
oretical guarantees, we prove that our decomposition algorithm
converges to a critical point of the nonconvex load-shedding 1T z = 0.
problem. The active power-flow equation with possible line removal can
Our presentation is organized as follows. In Section II, we be written as
formulate the worst-case load-shedding problem. In Section III,
we describe the separable structure of the load-shedding prob- EDdiag(γ) sin(E T θ) = P + z
lem, which lends itself to the PALM algorithm in Section IV. where diag(γ) is a diagonal matrix with its main diagonal equal
We analyze the convergence behavior of PALM in Section V to γ.
and provide numerical results in Section VI. Section VII offers Our objective is to identify a small number of lines in the
concluding remarks and future directions. active power networks, whose removal results in the maximum
load shedding. Thus, we consider the following worst-case load-
II. WORST-CASE LOAD-SHEDDING PROBLEM shedding problem:
In this section, we formulate the worst-case load-shedding maximize LoadShedding = 1T zd (2a)
γ , θ, z
problem for electrical power grids with active power-flow
models. subject to EDdiag(γ) sin(E T θ) = P + z (2b)
Following [5], [6], we consider a lossless power network
with n buses and m lines. A line l connecting bus i and bus γ ∈ {0, 1}m , m − 1T γ = K (2c)
j can be described by a vector el ∈ Rn with 1 and −1 at the 1T z = 0, z = [ zdT zgT ]T (2d)
ith and jth elements, respectively, and 0 everywhere else. Let
E = [ e1 · · · em ] ∈ Rn ×m be the incidence matrix that describes 0 ≤ zd ≤ −Pd , −Pg ≤ zg ≤ 0 (2e)
m transmission lines of the network, and let D ∈ Rm ×m be
− π ≤ E T θ ≤ π. (2f)
the diagonal matrix with the lth diagonal element being the
admittance of line l. For a lossless power network with fixed The decision variables are the phase angle, θ, the reduction
voltage at the buses, the active power-flow equation can be of load, zd , the reduction of generation, zg , and the out-of-
written in a vector form [5], [6] service line indicator, γ. The problem data, {E, D, P, K}, are
the incidence matrix for the network topology, the admittance
ED sin(E T θ) = P (1) matrix for the transmission lines, the real power injection at the
buses, and the number of out-of-service lines, respectively.
where θ ∈ Rn is the phase angles, P ∈ Rn is the real power Our load-shedding problem is based on the model introduced
injection at buses, and (·)T denotes matrix transpose. Reactive in [4]. Related models have been employed for the contingency
power equation over networks can be written similarly in a analysis in [5] and vulnerability analysis in [6]. In particular, the
vector form [5]. One can also extend this model to include per- nonlinear model in [5] includes both active and reactive power
unit voltages of buses; see [5], [6]. flow equations with varying voltage magnitudes.
We enumerate buses such that the power injection P can be In this article, we focus on the active power-flow equation with
partitioned into a load vector Pd ≤ 0 and a generation vector fixed voltages. Note that the angle difference between buses
Pg > 0, thus, P = [PdT PgT ]T . The sequence of buses indexed takes a value between −π and π. This is in contrast to the
in P is the same as that of the columns of the incidence matrix assumption of small angle differences employed in dc power
E. Since the power system is lossless, the sum of load is equal flow models [8].

Authorized licensed use limited to: Universitas Indonesia. Downloaded on July 25,2021 at 12:02:35 UTC from IEEE Xplore. Restrictions apply.
LIN: WORST-CASE LOAD SHEDDING IN ELECTRIC POWER NETWORKS 1271

While we assume a lossless network, the lossless constraint We begin by introducing the following indicator functions of
1T z = 0 can be extended to 1T z ≤ 0 that takes into account the constraint sets:
loss over transmission. Similarly, the constraint on power gen- 
0, if γ ∈ {0, 1}m and m − 1T γ = K
eration zg ≤ 0 can be replaced by zg ≤ P̄g where P̄g > 0. This φ1 (γ) = (4)
allows increase in the power generation for redispatch flexibil- ∞, otherwise
ity of generators. These extensions can be accommodated in the ⎧
⎪
⎪ 0, if 0 ≤ zd ≤ − Pd
proposed approach in subsequent sections. ⎪
⎪
⎨ and − Pg ≤ zg ≤ 0
φ2 (z) = (5)
⎪
⎪ and 1T z = 0
III. SEPARABLE STRUCTURE ⎪
⎪
⎩
The worst-case load-shedding problem contains nonlinear ∞, otherwise
constraints and binary variables. One source of nonlinearity and
is the sinusoidal function and another source is the multiplica- 
tion between diag(γ) and sin(E T θ). Therefore, it falls into the 0, if − π ≤ E T θ ≤ π
φ3 (θ) = . (6)
class of MINLPs, which is the challenging problem. Finding a ∞, otherwise
feasible point for MINLPs can be computationally expensive or
Then the minimization problem (3) can be compactly expressed
even NP-hard [20].
as
The maximum load-shedding problem (2) turns out to have
a separable structure. In what follows, we exploit this structure minimize Φ(γ, z, θ) = φ1 (γ) + φ2 (z) + φ3 (θ)
γ, z, θ
and develop a decomposition algorithm based on the PALM. + Hρ (γ, z, θ). (7)
A closer look at (2) reveals that the only constraint that cou-
ples all decision variables, θ, z, and γ, is the active power Our algorithm, based on PALM, uses the following iterations:
flow (2b). Otherwise, the binary variable, γ, is subject only  ak 
to the cardinality constraint (2c). The load shedding and the γ k +1 ∈ argmin φ1 (γ) + γ − uk 22 (8a)
γ 2
generation-reduction variables, zl , zg , are subject to the loss-
lessness constraint (2d) and the box constraint (2e). The phase bk
z k +1 ∈ argmin φ2 (z) + z − v k 22 (8b)
angles of the buses, θ, are subject only to the linear inequality z 2
constraint (2f). Therefore, the constraints in the load-shedding  ck 
problem (2) are separable with respect to θ, z, and γ, provided θk +1 ∈ argmin φ3 (θ) + θ − wk 22 (8c)
θ 2
that the active power flow (2b) is relaxed.
Let us denote the coupling term as where ak , bk , and ck are positive coefficients. In other words,
we minimize Φ with respect to γ, z, and θ, one at a time,
c(γ, z, θ) = EDdiag(γ) sin(E T θ) − (P + z). while fixing the other variables constant. The quadratic proximal
terms penalize the deviation of decision variables (γ, z, θ) from
We penalize its deviation from 0 and include the penalty in the (uk , v k , wk ), which are given by
cost function
1
ρ uk = γ k − ∇γ Hρ γ k , z k , θk
minimize Hρ (γ, z, θ) := −1T zd + c(γ, z, θ)22 ak
γ, z, θ 2
1
subject to (2c), (2d), (2e), (2f) (3) v k = z k − ∇z Hρ γ k +1 , z k , θk
bk
where ρ is a positive number. Clearly, (3) is a relaxation of (2), 1
since the active power-flow equation, c(γ, z, θ) = 0, is no longer wk = θk − ∇θ Hρ γ k +1 , z k +1 , θk . (9)
ck
enforced. Note that we minimize the negative of the load shed-
ding and follow the convention of minimizing the constraint We refer to [22] for extensive discussions on the proximal algo-
violation. The penalty of the constraint violation is controlled rithms and [23] for the generic PALM algorithms.
by ρ. By solving the relaxed problem (3) with a sufficiently large
ρ, the solution of (3) converges to the solution of (2). Additional A. Efficient Solutions to Subproblems
background on penalty methods can be found in [21, Ch. 13]. The minimization problems (8) are projections on the corre-
sponding constraint sets in (4)–(6). In particular, the projection
IV. DECOMPOSITION ALGORITHM on the convex sets (5)–(6) can be computed efficiently. For the
projection on the nonconvex set (4), it turns out that the solution
In this section, we develop a decomposition algorithm that
has a closed-form expression.
exploits the separable structure of the worst-case load-shedding
We begin with the projection on the convex sets. The z-
problem. Roughly speaking, we minimize the cost function by
minimization problem (8b) can be expressed as
cycling through variables while keeping other variables fixed.
The original problem is thus broken down into a sequence of bk
minimize z − v k 22
partial problems that are more amenable to efficient algorithms 2
or even closed-form solutions. subject to L ≤ z ≤ U, 1T z = 0 (10)

Authorized licensed use limited to: Universitas Indonesia. Downloaded on July 25,2021 at 12:02:35 UTC from IEEE Xplore. Restrictions apply.
1272 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 6, NO. 3, SEPTEMBER 2019

where the lower bound is L = −[0T PgT ]T and the upper bound Algorithm 1: Proximal Alternating Linearization Method.
is U = −[PdT 0T ]T . The solution of this convex quadratic pro-
Start with any (γ k , z k , θk ) and set k ← 0.
gram with box constraints and a single equality constraint,
for k = 0, 1, 2, . . . until convergence do
1T z = 0, can be computed efficiently [24]. The θ-minimization
// γ-minimization:
problem (8c) can be expressed as
Set ak = r1 L1 (z k , θk ) where r1 > 1 and L1 in (15a).
ck Solve problem (12) via the closed-form expression (13)
minimize θ − wk 22 to get γ k +1 .
2
subject to −π ≤ E T θ ≤ π. (11) // z-minimization:
Set bk = r2 L2 (γ k +1 , θk ) where r2 > 1 and L2 in (15b).
This bound-constrained least-squares problem can be solved Solve the convex quadratic problem (10)
efficiently [24]. to get z k +1 .
We next provide a closed-form solution to the γ-minimization // θ-minimization:
problem (8a) Set ck = r3 L3 (γ k +1 , z k +1 ) where r3 > 1 and L3
in (15c). Solve the convex quadratic problem (11)
ak to get θk +1 .
minimize γ − uk 22 Set (γ k , z k , θk ) ← (γ k +1 , z k +1 , θk +1 ).
2
subject to γ ∈ {0, 1}m , 1T γ = m − K. (12) end for

Lemma 1: Let [uk ]K be the Kth smallest element of uk .

The positive coefficients ak , bk , and ck in (8) and (9) are
The ith element of the solution to (12) is given by
determined by
1 if uki ≥ [uk ]K ak = r1 L1 z k , θk
γi = (13)
0 otherwise
bk = r2 L2 γ k +1 , θk
for i = 1, . . . , m. ck = r3 L3 γ k +1 , z k +1
The proof can be found in Appendix A.
where positive constants ri > 1 for i = 1, 2, 3. The Lipschitz
constants Li for the partial gradients ∇Hρ are given by
B. Expressions for Partial Gradient ∇Hρ L1 (z k , θk ) = ρ (DE T ED)
Proximal algorithms typically rely on convexity assumptions
to guarantee convergence [22]. In contrast, the PALM algorithm ◦ (sin(E T θk ) sin(E T θk )T ) (15a)
does not require the objective or the constraints to be convex. L2 (γ k +1 , θk ) = ρ (15b)
PALM relies on the smoothness condition of the coupling term
k +1 k +1 2 k +1 k +1
Hρ and the Lipschitz conditions of the partial gradients ∇Hρ . L3 (γ ,z ) = ρE (2Q  + R ) (15c)
Another feature of PALM is that it does not require stepsize
where Qk = Γk DE T EDΓk , Rk = Γk DE T (P + z k ), and  ·
rules as in typical descent-based methods. This is because the
 denotes the maximum singular value of a matrix. The deriva-
Lipschitz conditions guarantee the descent of the objective value
tion of the Lipschitz constants are provided in Appendix C.
in each PALM iteration; see Section V.
We conclude this section by summarizing PALM in
We provide the expressions for ∇Hρ and discuss the choice
Algorithm 1.
of ak , bk , and ck in (8).
Lemma 2: The partial gradients ∇Hρ with respect to γ, z,
V. CONVERGENCE ANALYSIS
and θ are given by
In this section, we show that Algorithm 1 converges to a criti-
∇γ Hρ = ρ (DE ED) ◦ (sin(E θ) sin(E θ) ) γ
T T T T cal point of the nonconvex problem (3). This convergence behav-
ior is independent of the initial guess of the decision variables.
− ρ (sin(E T θ)(P + z)T ED) ◦ I 1, (14a) Furthermore, the objective value Φ is monotonically decreasing
with the number of iterates, that is
∇z Hρ = − [1T 0T ]T + ρ(P + z) − ρ(EDΓ sin(E T θ))
(14b) Φ γ k +1 , z k +1 , θk +1 ≤ Φ γ k , z k , θk .

∇θ Hρ = ρ Ediag(cos(E T θ))ΓDE T This feature of monotonic decreasing allows us to monitor the

progress of PALM. It also allows us to check if the implemen-
× (EDΓ sin(E T θ) − (P + z)) (14c) tation is correct in practice.
We begin with two technical lemmas on the Lipschitz prop-
where Γ := diag(γ) and ◦ denotes the element-wise product of erties of Φ.
matrices. Lemma 3: The objective function Φ in (7) satisfies the fol-
The derivation can be found in Appendix B. lowing properties:

Authorized licensed use limited to: Universitas Indonesia. Downloaded on July 25,2021 at 12:02:35 UTC from IEEE Xplore. Restrictions apply.
LIN: WORST-CASE LOAD SHEDDING IN ELECTRIC POWER NETWORKS
1) inf γ ,z ,θ Φ(γ, z, θ) > −∞, inf γ φ1 (γ) > −∞, inf z φ2
(z) > −∞, and inf θ φ3 (θ) > −∞.
2) For fixed (z, θ), the partial gradient ∇γ Hρ is globally
Lipschitz,
∇γ Hρ (γ1 , z, θ) − ∇γ Hρ (γ2 , z, θ)
≤ L1 (z, θ)γ1 − γ2 
for all γ1 and γ2 . Likewise, for fixed (γ, θ), the partial
gradient ∇z Hρ satisfies
∇z Hρ (γ, z1 , θ) − ∇z Hρ (γ, z2 , θ)
≤ L2 (γ, θ)z1 − z2 
for all z1 and z2 , and for fixed (z, γ)
∇θ Hρ (γ, z, θ1 ) − ∇θ Hρ (γ, z, θ2 )
≤ L3 (γ, z)θ1 − θ2 
for all θ1 and θ2 .
3) There exist positive constants s1 , s2 , s3 such that
sup{L1 (z , θ )} ≤ s1
k k
k The proof is complete by establishing Property 3). Since the
sup{L2 (γ , θ )} ≤ s2
k k
k matrices is upper bounded by the product of maximum singular
sup{L3 (γ k , z k )} ≤ s3 .
k
4) The entire gradient ∇Hρ (γ, z, θ) is Lipschitz continuous
on bounded subsets of Rm × Rn × Rn .
Remark 1: Property 1) is necessary for the minimization
problems in Algorithm 1, and thus the minimization of Φ, to
be well defined. Property 2) on the globally Lipschitz bounds is
critical for the convergence of PALM. Note that the block Lip-
schitz property of ∇Hρ is weaker than the globally Lipschitz
assumption of Φ in joint variables (γ, z, θ) in standard proximal
methods [23]. Property 3) guarantees that the Lipschitz con-
stants for partial gradients are upper bounded by finite numbers.
Property 4) is a mild condition which holds when Hρ is twice
continuously differentiable.
Proof: Property 1) is a direct consequence of the nonneg-
ativity of Hρ and the definition of the indicator functions φ1 ,
φ2 , and φ3 . Property 4) holds because Hρ is twice continuously
differentiable.
To show Property 2), recall that for fixed (z k , θk ) the Lipschitz
constant L1 (z k , θk ) of ∇γ Hρ is determined by
∇γ Hρ (γ1 , z k , θk ) − ∇γ Hρ (γ2 , z k , θk )
≤ L1 (z k , θk ) γ1 − γ2 
for all γ1 and γ2 . Since ∇γ Hρ is an affine function of γ
[see (14a)], it follows that:
L1 (z k , θk ) = ρ (DE T ED) ◦ (sin(E T θk ) sin(E T θk )T ) .
Authorized licensed use limited to: Universitas Indonesia. Downloaded on July 25,2021 at 12:02:35 UTC from IEEE Xplore. Restrictions apply.
1273
For fixed (γ k +1 , θk ), the Lipschitz constant L2 (γ k +1 , θk ) of
∇z Hρ is determined by
∇z Hρ (γ k +1 , z1 , θk ) − ∇z Hρ (γ k +1 , z2 , θk )
≤ L2 (γ k +1 , θk )z1 − z2 
for all z1 and z2 . Since ∇z Hρ is an affine function of z
[see (14b)], it follows that:
L2 (γ k +1 , θk ) = ρ.
For fixed (γ k +1 , z k +1 ), the Lipschitz constant L3 (γ k +1 , z k +1 )
of ∇θ Hρ is determined by
∇θ Hρ (γ k +1 , z k +1 , θ1 ) − ∇θ Hρ (γ k +1 , z k +1 , θ2 )
≤ L3 (γ k +1 , z k +1 )θ1 − θ2 
for all θ1 and θ2 . The Lipschitz constant for ∇θ Hρ is given
by (see Appendix C for derivation)
L3 (γ k , z k ) = ρE2 2Qk  + Rk 
where
Qk = Γk DE T EDΓk , Rk = Γk DE T (P + z k ).
maximum singular value of the element-wise product of two
values of individual matrices [25, Th. 5.5.1], it follows that:
(16)
L1 (z k , θk ) ≤ ρ DE T ED · sin(E T θk ) sin(E T θk )T
thus, s1 = ρmED2 . From (15b), we have s2 = ρ and
from (15c), we have s3 = ρE2 ED2 (2 + P ). 
The convergence of PALM relies on the so-called Kurdyka–
Lojasiewicz (KL) property. We refer to [26], [27], and [23] for
detailed discussions on the KL theory. We next recall a few
definitions needed for our PALM algorithm.
Definition 1: Let f : Rd → (−∞, +∞] be proper and lower
semicontinuous. The function f is said to have the KL prop-
erty at ū ∈ dom ∂f := {u ∈ Rd : ∂f (u) = ∅} if there exist
η ∈ (0, +∞], a neighborhood N of ū, and a function ψ such
that for all u ∈ N ∩ {f (ū) < f (u) < f (ū) + η}, the following
inequality holds:
ψ  (f (u) − f (ū)) · dist(0, ∂f (u)) ≥ 1
where ∂f denotes the subdifferential of f and dist(x, s) :=
inf{y − x : y ∈ s} denotes the distance from a point x ∈ Rd
to a set s ⊂ Rd . A function f is called a KL function if f satisfies
the KL property at each point of dom ∂f .
The KL property is a technical condition that controls the
difference in function value by its gradient. It turns out that a
large class of functions that arise in modern applications satisfy
the KL property [23], [26], [27]. One useful way of establishing
the KL property is via the connection with the semialgebraic
sets and the semialgebraic functions.
Definition 2: A subset S of Rd is a real semialgebraic set
if there exists a finite number of real polynomial functions gij
1274 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 6, NO. 3, SEPTEMBER 2019

and hij : Rd → R such that


p 
q
S= {u ∈ Rd : gij (u) = 0 and hij (u) < 0}.
j =1 i=1

Definition 3: A function h : Rd → (−∞, +∞] is called

semialgebraic function if its graph {(u, v) ∈ Rd+1 : h(u) = v}
is a semialgebraic subset of Rd+1 .
Given these definitions we show the KL property of Φ.
Lemma 4: The objective function Φ in (7) satisfies the KL
property.
Proof: Since analytic functions satisfy the Lojasiewicz in-
equality [26], [27] and since Hρ is the multiplication of polyno-
mial function and sinusoidal function, it follows that Hρ satisfies
the KL property.
The nonsmooth parts of Φ, namely, the indicator functions
φ1 , φ2 , and φ3 , are lower semicontinuous. Since a proper, lower
semicontinuous, and semialgebraic function satisfies the KL
property [23, Th. 3], it suffices to show that φ1 , φ2 , and φ3 Fig. 1. Diagram of IEEE 14-bus test case.
are semialgebraic functions. Because φ2 and φ3 are indica-
tor functions of the semialgebraic sets (5)–(6), they are semi-
TABLE I
algebraic functions. To show that φ1 is semialgebraic, note LOAD-SHEDDING STRATEGY FOR IEEE 14-BUS TEST CASE
that the binary constraint γi ∈ {0, 1} can be expressed as a
polynomial equation γi (γi − 1) = 0 for i = 1, . . . , m. Thus
{γ | γ ∈ {0, 1}, m − 1T γ = K} is a semialgebraic set. There-
fore the indicator function φ1 is semialgebraic, which completes
the proof. 
After establishing Lemma 3 and Lemma 4, the main con-
The lines to be removed are labeled in Fig. 1.
vergence results follow from the pioneering work by Bolte
et al. [23].
Proposition 1: Suppose that Φ is a KL function that satisfies A. IEEE 14-bus Test Case
conditions in Lemma 3. Let xk = (γ k , z k , θk ) be a bounded
sequence generated by PALM. The following results hold. Consider IEEE 14-bus test case shown in Fig. 1. This small
1) The sequence {xk } has finite length, that is system has five generator buses, nine load buses, and 20 trans-
mission lines. We compute the generation profile, Pg , and the
∞
 load profile, Pd , by solving the steady-state power-flow equa-
xk +1 − xk 2 < ∞. tions via MATPOWER [28].
k =1
We take out up to five lines to track the progress of the
2) The sequence {xk } converges to a critical point x∗ = worst-case load shedding in this small network. As the out-of-
(γ ∗ , z ∗ , θ∗ ) of Φ. service number of lines increases from K = 1 to K = 5, the
3) The sequence Φ(xk ) is nonincreasing amount of load shed increases from 18.3% to 65.3% of the
total power load; see Table I. It turns out that the set of lines
d k +1 to be taken out of service is a subset of the lines as K in-
x − xk 22 ≤ Φ(xk ) − Φ(xk +1 ), k≥0
2 creases. This implies the consistency in the set of critical trans-
where d is positive constant bounded as follows. mission lines for load shedding. The out-of-service lines are
Proof: The finite length property and the convergence to a highlighted in Fig. 1. It is worth mentioning that PALM is initial-
critical point follow from [23, Th. 1]. The monotonicity of the ized with (γ = 1, θ = 0, z = 0) for all K = 1, . . . , 5. In other
objective value is obtained from in [23, Lemma 3].  words, the algorithm starts with full service lines and zero load
shed.
Fig. 2 shows the convergence results of PALM when five lines
VI. NUMERICAL RESULTS are removed. The objective function decreases monotonically
In this section, we verify the convergence results of PALM with the PALM iterations, thereby confirming the prediction in
and examine its solution quality in two IEEE test cases. The Proposition 1. Furthermore, both the dual residuals and the pri-
first test case, IEEE 14-bus system, illustrates the convergence mal residual decrease monotonically. The fastest convergence
behavior of PALM. The second test case, IEEE 118-bus system, of PALM is in the first 200–300 iterations, in this case. The con-
demonstrates the scalability of the algorithm. vergence rate depends on the size of the problem and the choice

Authorized licensed use limited to: Universitas Indonesia. Downloaded on July 25,2021 at 12:02:35 UTC from IEEE Xplore. Restrictions apply.
LIN: WORST-CASE LOAD SHEDDING IN ELECTRIC POWER NETWORKS
Fig. 2. Convergence results of PALM for IEEE 14-bus test case: the objective value (left), the dual residuals (center), and the primal residual
(right). The markers show at every 50 iterations.
Fig. 3. Diagram of the IEEE 118-bus test case.
of parameter ρ. While a bigger ρ improves the primal conver-
gence rate, it slows down the dual convergence rate. In practice,
we find that ρ ∈ [104 , 106 ] achieves a good balance between the
primal and dual residuals.
Since we relax the constraint c(γ, z, θ) = 0 in (3), we check
the solution quality in satisfying the power-flow equation. As
shown in Fig. 2, the primal residual c(γ, z, θ) is monotoni-
cally decreasing with PALM iterations; in particular, we have
c(γ, z, θ) ≤ 3.5 × 10−3 after 1000 iterations. As discussed
above, one can further reduce the primal residual by increasing
the penalty parameter ρ.
B. IEEE 118-Bus Test Case
We next consider IEEE 118-bus test case as shown in Fig. 3.
This large power system has 54 generator buses, 64 load buses,
and 186 transmission lines. As in IEEE-14 bus system, the gen-
eration profile, Pg , and load profile, Pd , are obtained by solving
the steady-state power-flow equations via MATPOWER [28].
While the 118-bus system is much larger than the 14-bus
system, the convergence behavior of PALM is similar. The ob-
jective value, the dual residuals, and the primal residual all
Authorized licensed use limited to: Universitas Indonesia. Downloaded on July 25,2021 at 12:02:35 UTC from IEEE Xplore. Restrictions apply.
1275
decrease monotonically, as shown in Fig. 4. After 2000 itera-
tions, the primal residual is smaller than 1.3 × 10−2 and the
dual residual is smaller than 1.2 × 10−5 . The solution quality
is determined by the primal residual. The computational time
is less than 10 minutes on a laptop with 8GB RAM running
2.4 GHz CPU.
It is worth mentioning that the binary variable, γ, converges
well before the continuous variables, z and θ. This observa-
tion implies that after γ converges, one can utilize a local NLP
solver to further speed up convergence to the final values of the
continuous variables.
Table II shows the worst-case load-shedding scenarios in
which up to five transmission lines are taken offline. As ob-
served in 14-bus test case, the most critical lines to be taken out
of service form a subset of lines as K increases. For this large
system, the load shed percentage is less than 10.2% when five
lines are taken out. This is in contrast to the 14-bus system, in
which the load shed percentage is more than 65% when K = 5.
To gain some insight into the out-of-service lines, we consider
the types of buses with which the lines connect. As shown in
Table III, all critical lines connect the same types of buses, that
is, generator to generator and load to load buses. In particular,
1276 IEEE TRANSACTIONS ON CONTROL OF NETWORK SYSTEMS, VOL. 6, NO. 3, SEPTEMBER 2019
Fig. 4. Convergence results of PALM for IEEE 118-bus test case: the objective value (left), the dual residuals (center), and the primal residual
(right). The markers show at every 100 iterations.
TABLE II
LOAD-SHEDDING STRATEGY FOR IEEE 118-BUS TEST CASE
TABLE III
SET OF OUT-OF-SERVICE LINES AND THE BUS TYPES FOR IEEE
118-BUS TEST CASE
four out of the five critical lines connect generator buses. This
indicates the importance of lines between generator buses in
IEEE-118 system.
VII. CONCLUDING REMARKS
We formulate the worst-case load-shedding problem for elec-
trical power networks. We show that this nonconvex control
problem has a separable structure that can be exploited by
PALM. The PALM algorithm decomposes the load-shedding
problem into a sequence of subproblems that are amenable
to convex optimization or closed-form solutions. We prove
convergence of PALM to a critical point by leveraging the
KL theory. We believe that our proof techniques and the upper
bounds on the Lipschitz constants can be instrumental in de-
veloping other decomposition algorithms in large-scale power
networks.
For future directions, we intend to study a preventive prob-
lem formulation that includes system response to failures. One
approach is to employ the automatic generation control for the
generators to counteract the load shedding. Another direction of
interest is to allow the buses to have varying voltage magnitudes.
This extension requires further study of the generator’s reactive
power outputs and their nonlinear behaviors.
ACKNOWLEDGMENT
The author would like to thank the reviewers for their com-
ments and suggestions that improve the article.
Authorized licensed use limited to: Universitas Indonesia. Downloaded on July 25,2021 at 12:02:35 UTC from IEEE Xplore. Restrictions apply.
APPENDIX
A. Proof of Lemma 1
We prove by contradiction. Let γ satisfy 1T γ = m − K and
γi ∈ {0, 1}, but γ is different from the projection in (13). In
other words, there exists at least one element of γ, say, the lth
element such that γl = 1 with the corresponding ukl < [uk ]K ,
and at least one element, say, the jth element such that γj = 0
with the corresponding ukj ≥ [uk ]K . Consider
δlj = (γl − ukl )2 + (γj − ukj )2 = (1 − ukl )2 + (ukj )2 .
and the cost of the swapping the values of γl and γj
δj l = (ukl )2 + (1 − ukj )2 .
Since δlj − δj l = 2(ukj − ukl ) > 0, we conclude that the cost
function decreases if we choose (γl = 0, γj = 1) instead of
(γl = 1, γj = 0). In other words, we can reduce the cost by
swapping the values of γl = 1 with respect to ukl < [uk ]K and
γj = 0 with respect to ukl ≥ [uk ]K until (13) is satisfied for all
elements of γ. This completes the proof.
B. Proof of Lemma 2
The derivations of (14a) and (14b) are straightforward, as
they amount to taking the derivatives of quadratic functions,
thus omitted. The derivation of (14c) involves taking the first-
order variation for sine and cosine functions. We begin by taking
variation θ̃ around θ
sin(E T (θ + θ̃)) = sin(E T θ) ◦ cos(E T θ̃)
+ cos(E T θ) ◦ sin(E T θ̃)
where ◦ is the element-wise product. When θ̃ is small, we have
the first-order approximation
sin(E T (θ + θ̃)) ≈ sin(E T θ) + diag(cos(E T θ))E T θ̃.
It follows that the first-order approximation of Hρ (θ + θ̃) is
given by
Hρ (θ + θ̃) ≈ Hρ (θ) + ρ (EDΓ sin(E T θ) − (P + z))T
× EDΓ diag(cos(E T θ))E T θ̃.
Taking the transpose of the matrix multiplying θ̃ yields
∇θ Hρ (θ) = ρEdiag(cos(E T θ))[Q sin(E T θ) − R].
where Q = ΓDE T EDΓ and R = ΓDE T (P + z).
LIN: WORST-CASE LOAD SHEDDING IN ELECTRIC POWER NETWORKS
C. Lipschitz Constant of ∇θ Hρ
Recall that  sin(θ1 − θ2 ) ≤ θ1 − θ2  for all θ1 , θ2 . We
have
sin(E T θ1 ) − sin(E T θ2 )
= 2 cos(E T (θ1 + θ2 )/2) ◦ sin(E T (θ1 − θ2 )/2)
≤ 2 sin(E T (θ1 − θ2 )/2) ≤ Eθ1 − θ2 . (17)
The equality is the element-wise sum-to-product identity. The
first inequality follows from the fact that all cosine functions are
upper bounded by 1. Similar calculation yields:
 cos(E T θ1 ) − cos(E T θ2 ) ≤ Eθ1 − θ2 . (18)
Let f (θ) = sin(E T θ) and g(θ) = diag(cos(E T θ))Q. By
adding and subtracting the same term yields
g(θ1 )f (θ1 ) − g(θ2 )f (θ2 )
= g(θ1 )(f (θ1 ) − f (θ2 )) + (g(θ1 ) − g(θ2 ))f (θ2 ).
We calculate
g(θ1 )f (θ1 ) − g(θ2 )f (θ2 )
≤ diag(cos(E T θ1 ))Q(sin(E T θ1 ) − sin(E T θ2 ))
+ diag(cos(E T θ1 ) − cos(E T θ2 ))Q sin(E T θ2 )
≤ 2QEθ1 − θ2 
where we have used (17) and (18). It follows that the Lipschitz
constant for ∇θ Hρ is given by:
∇θ Hρ (θ1 ) − ∇θ Hρ (θ2 ) ≤ L3 (γ k +1 , z k +1 )θ1 − θ2 
where L3 (γ k +1 , z k +1 ) = ρE2 (2Qk +1  + Rk +1 ) and
Qk +1 = Γk +1 DE T EDΓk +1
Rk +1 = Γk +1 DE T P + z k +1 .
REFERENCES
[1] G. Andersson et al., “Causes of the 2003 major grid blackouts in
North America and Europe, and recommended means to improve sys-
tem dynamic performance,” IEEE Trans. Power Syst., vol. 20, no. 4,
pp. 1922–1928, Nov. 2005.
[2] C. D. Brummitt, R. M. D’Souza, and E. Leicht, “Suppressing cascades of
load in interdependent networks,” Proc. Nat. Acad. Sci., vol. 109, no. 12,
pp. E680–E689, 2012.
[3] F. D. Galiana, “Bound estimates of the severity of line outages in power
system contingency analysis and ranking,” IEEE Trans. Power App. Syst.,
vol. PAS-103, no. 9, pp. 2612–2624, Sep. 1984.
[4] V. Donde, V. López, B. Lesieutre, A. Pinar, C. Yang, and J. Meza, “Iden-
tification of severe multiple contingencies in electric power networks,” in
Proc. 37th Annu. North Amer. Power Symp., 2005, pp. 59–66.
[5] V. Donde, V. López, B. Lesieutre, A. Pinar, C. Yang, and J. Meza, “Severe
multiple contingency screening in electric power systems,” IEEE Trans.
Power Syst., vol. 23, no. 2, pp. 406–417, May 2008.
[6] A. Pinar, J. Meza, V. Donde, and B. Lesieutre, “Optimization strategies
for the vulnerability analysis of the electric power grid,” SIAM J. Optim.,
vol. 20, no. 4, pp. 1786–1810, 2010.
[7] F. Lin and C. Chen, “An ADMM algorithm for load shedding in electric
power grids,” in Proc. Amer. Control Conf., 2016, pp. 5002–5007.
[8] F. Lin and J. Chen, “Learning low-complexity autoregressive models
with limited time sequence data,” in Proc. Amer. Control Conf., 2017,
pp. 3153–3158.
Authorized licensed use limited to: Universitas Indonesia. Downloaded on July 25,2021 at 12:02:35 UTC from IEEE Xplore. Restrictions apply.
1277
[9] F. Lin and J. Chen, “Learning low-complexity autoregressive models
via proximal alternating minimization,” Syst. Control Lett., vol. 122,
pp. 48–53, 2018.
[10] F. Lin and S. D. Bopardikar, “Sparse linear-quadratic-Gaussian con-
trol in networked systems,” IFAC-PapersOnLine, vol. 50, no. 1,
pp. 10748–10753, 2017.
[11] F. Lin and V. Adetola, “Co-design of sparse output feedback and
row/column-sparse output matrix,” in Proc. Amer. Control Conf., 2017,
pp. 4359–4364.
[12] J. Salmeron, K. Wood, and R. Baldick, “Analysis of electric grid se-
curity under terrorist threat,” IEEE Trans. Power Syst., vol. 19, no. 2,
pp. 905–912, May 2004.
[13] J. M. Arroyo and F. D. Galiana, “On the solution of the bilevel program-
ming formulation of the terrorist threat problem,” IEEE Trans. Power
Syst., vol. 20, no. 2, pp. 789–797, May 2005.
[14] D. Xu and A. A. Girgis, “Optimal load shedding strategy in power systems
with distributed generation,” in Proc. Power Eng. Soc. Winter Meeting,
2001, vol. 2, pp. 788–793.
[15] N. Sadati, T. Amraee, and A. Ranjbar, “A global particle swarm-based-
simulated annealing optimization technique for under-voltage load shed-
ding problem,” Appl. Soft Comput., vol. 9, no. 2, pp. 652–657, 2009.
[16] L. P. Hajdu, J. Peschon, W. F. Tinney, and D. Piercy, “Optimum load-
shedding policy for power systems,” IEEE Trans. Power App. Syst.,
vol. PAS-87, no. 3, pp. 784–795, Mar. 1968.
[17] K. Palaniswamy, J. Sharma, and K. Misra, “Optimum load shedding taking
into account of voltage and frequency characteristics of loads,” IEEE
Trans. Power App. Syst., vol. PAS-104, no. 6, pp. 1342–1348, Jun. 1985.
[18] Y. Halevi and D. Kottick, “Optimization of load shedding system,” IEEE
Trans. Energy Convers., vol. 8, no. 2, pp. 207–213, Jun. 1993.
[19] S. Pahwa, C. Scoglio, S. Das, and N. Schulz, “Load-shedding strategies
for preventing cascading failures in power grid,” Elect. Power Compon.
Syst., vol. 41, no. 9, pp. 879–895, 2013.
[20] R. Fletcher and S. Leyffer, “Solving mixed-integer nonlinear programs
by outer approximation,” Math. Program., vol. 66, no. 1/3, pp. 327–349,
1994.
[21] D. G. Luenberger and Y. Ye, Linear and Nonlinear Programming. New
York, NY, USA: Springer, 2008.
[22] N. Parikh and S. P. Boyd, “Proximal algorithms.” Found. Trends Optim.,
vol. 1, no. 3, pp. 127–239, 2014.
[23] J. Bolte, S. Sabach, and M. Teboulle, “Proximal alternating linearized
minimization for nonconvex and nonsmooth problems,” Math. Program.,
vol. 146, no. 1/2, pp. 459–494, 2014.
[24] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, U.K.:
Cambridge Univ. Press, 2004.
[25] R. A. Horn and C. R. Johnson, Topics in Matrix Analysis. Cambridge,
U.K.: Cambridge Univ. Press, 1991.
[26] K. Kurdyka, “On gradients of functions definable in o-minimal structures,”
in Proc. Annales de l’institut Fourier, vol. 48, no. 3, 1998, pp. 769–784.
[27] J. Bolte, A. Daniilidis, O. Ley, and L. Mazet, “Characterizations of
Lojasiewicz inequalities: Subgradient flows, Talweg, convexity,” Trans.
Amer. Math. Soc., vol. 362, no. 6, pp. 3319–3363, 2010.
[28] R. D. Zimmerman, C. E. Murillo-Sánchez, and R. J. Thomas, “MAT-
POWER: Steady-state operations, planning, and analysis tools for power
systems research and education,” IEEE Trans. Power Syst., vol. 26, no. 1,
pp. 12–19, Feb. 2011.
Fu Lin (S’06–M’13) received the B.S. degree
in instrument science and engineering from
Shanghai Jiaotong University, Shanghai, China,
in 2005, and the Ph.D. degree in electrical and
computer engineering from the University of
Minnesota, Minneapolis, MN, USA, in 2012.
He was a Postdoctoral Research Associate
in the Mathematics and Computer Science Divi-
sion of Argonne National Laboratory, Lemont, IL,
USA, before joining the Autonomous and Intel-
ligent Systems Department of United Technolo-
gies Research Center, East Hartford, CT, USA, in 2016. His research
interests include distributed control, optimization, and machine learning.