Cascading Failures in Power Grids –

Analysis and Algorithms

Saleh Soltan Dorian Mazauric Gil Zussman

Electrical Engineering Laboratoire d’Informatique Electrical Engineering
Columbia University Fondamentale de Marseille Columbia University
New York, NY Marseille, France New York, NY
saleh@[Link] [Link]@[Link] gil@[Link]

ABSTRACT
This paper focuses on cascading line failures in the trans-
mission system of the power grid. Recent large-scale power
outages demonstrated the limitations of percolation- and
epidemic-based tools in modeling cascades. Hence, we study
cascades by using computational tools and a linearized power
flow model. We first obtain results regarding the Moore-
Penrose pseudo-inverse of the power grid admittance matrix.
Based on these results, we study the impact of a single line
failure on the flows on other lines. We also illustrate via sim- Figure 1: The first 11 line outages leading to the India black-
ulation the impact of the distance and resistance distance on out on July 2012 [2] (numbers show the order of outages).
the flow increase following a failure, and discuss the differ-
ence from the epidemic models. We use the pseudo-inverse of
1. INTRODUCTION
admittance matrix to develop an efficient algorithm to iden- Recent failures of the power grid (such as the 2003 and
tify the cascading failure evolution, which can be a building 2012 blackouts in the Northeastern U.S. [1] and in India [2])
block for cascade mitigation. Finally, we show that finding demonstrated that large-scale failures will have devastating
the set of lines whose removal results in the minimum yield effects on almost every aspect in modern life. The grid is
(the fraction of demand satisfied after the cascade) is NP- vulnerable to natural disasters, such as earthquakes, hurri-
Hard and introduce a simple heuristic for finding such a set. canes, and solar flares as well as to terrorist and Electro-
Overall, the results demonstrate that using the resistance magnetic Pulse (EMP) attacks [43]. Moreover, large scale
distance and the pseudo-inverse of admittance matrix pro- cascades can be initiated by sporadic events [1, 2, 42].
vides important insights and can support the development Therefore, there is a need to study the vulnerability of the
of efficient algorithms. power transmission network. Unlike graph-theoretical net-
work flows, power flows are governed by the laws of physics
and there are no strict capacity bounds on the lines [10].
Categories and Subject Descriptors Yet, there is a rating threshold associated with each line –
C.4 [Performance of Systems]: Reliability, availability, if the flow exceeds the threshold, the line will eventually ex-
and serviceability; G.2.2 [Discrete Mathematics]: Graph perience thermal failure. Such an outage alters the network
Theory—Graph algorithms, Network problems topology, giving rise to a different flow pattern which, in
turn, could cause other line outages. The repetition of this
Keywords process constitutes a cascading failure [19].
Previous work (e.g., [17, 18, 45] and references therein) as-
Power Grid; Pseudo-inverse; Cascading Failures; Algorithms sumed that a line/node failure leads, with some probabil-
ity, to a failure of nearby nodes/lines. Such epidemic based
modeling allows using percolation-based tools to analyze the
cascade’s effects. Yet, in real large scale cascades, a failure
of a specific line can affect a remote line and the cascade
does not necessarily develop in a contiguous manner. For
example, the evolution of the cascade in India on July 2012
Permission to make digital or hard copies of all or part of this work for personal or appears in Fig. 1. Similar non-contiguous evolution was ob-
classroom use is granted without fee provided that copies are not made or distributed served in a cascade in Southern California in 2011 [11, 42]
for profit or commercial advantage and that copies bear this notice and the full citation and in simulation studies [11, 12].
on the first page. Copyrights for components of this work owned by others than the
Motivated by this observation, we study the effects of edge
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission failures and introduce algorithms to identify the cascading
and/or a fee. Request permissions from permissions@[Link]. failure evolution and vulnerable lines. We employ the (lin-
e-Energy’14, June 11–13, 2014, Cambridge, UK.
Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM 978-1-4503-2819-7/14/067 ...$15.00.
[Link]
earized) direct-current (DC) power flow model,1 which is a a fast algorithm for identifying the evolution of the cascade,
practical relaxation of the alternating-current (AC) model, and (iii) develop a heuristic algorithm for the minimum yield
and the cascading failure model of [25] (see also [11–14]). problem.
Specifically, we first review the model and the Cascading This paper is organized as follows. Section 2 reviews re-
Failure Evolution (CFE) Algorithm that has been used to lated work. Section 3 describes the power flow, cascade
identify the evolution of the cascade [13,14,19] (its complex- model, metrics, and the graphs used in the simulations. In
ity is O(t|V |3 ), where |V | is the number of nodes and t is Section 4, we derive the properties of the admittance ma-
the number of cascade rounds). trix of the grid. Section 5 presents the effects of a single
Then, in order to investigate the impact of a single edge edge failure. Section 6 introduces the CFE-PB Algorithm.
failure on other edges, we use matrix analysis tools to study Section 7 discusses the hardness of the minimum yield prob-
the properties of the admittance matrix of the grid2 and lem and introduces the MVES-RB Algorithm. Section 8
Moore-Penrose Pseudo-inverse [4] of the admittance matrix. provides concluding remarks and directions for future work.
In particular, we provide a rank-1 update of the pseudo- The proofs appear in the Appendix.
inverse of the admittance matrix after a single edge failure.
We use these results along with the resistance distance and 2. RELATED WORK
Kirchhoff ’s index notions3 to study the impact of a single
Network vulnerability to attacks has been thoroughly stud-
edge failure on the flows on other edges. We obtain upper
ied (e.g., [3, 30, 37] and references therein). However, most
bounds on the flow changes after a single failure and study
previous computational work did not consider power grids
the robustness of specific graph classes. We also illustrate
and cascading failures. Recent work on cascades focused
via simulations the relation between the flow changes after
on probabilistic failure propagation models (e.g., [17,18,45],
a failure and the distance (in hop count) and resistance dis-
and references therein). However, real cascades [1,2,42] and
tance from the failure in the U.S. Western interconnection
simulation studies [11, 12] indicate that the cascade propa-
as well as Erdős-Rényi [26], Watts and Strogatz [44], and
gation is different than that predicted by such models.
Barábasi and Albert [9] graphs. These simulations show
In Sections 4 and 6, we use the admittance matrix of the
that there are cases in which an edge flow far away from the
grid to compute flows. This is tightly connected to the prob-
failure significantly increases. These observations are clearly
lem of solving Laplacian systems. Solving these systems can
in contrast to the epidemic-based models.
be done with several techniques, including Gaussian elim-
Once lines fail, there is a need for low complexity al-
ination and LU factorization [27]. Recently, [20] designed
gorithms to control and mitigate the cascade. Hence, we
algorithms that use preconditioning, to provide highly pre-
develop the low complexity Cascading Failure Evolution –
cise approximate solutions to Laplacian systems in nearly
Pseudo-inverse Based (CFE-PB) Algorithm for identifying
linear time. However, this approach only provides approx-
the evolution of a cascade that may be initiated by a fail-
imate solutions and is not suitable for analytical studies of
ure of several edges. The algorithm is based on the rank-1
the effects of edge failures.
update of the pseudo-inverse of the admittance matrix. We
In Section 5, we obtain upper bounds on the flow changes
show that its complexity is O(|V |3 + |Ft∗ ||V |2 ) (|Ft∗ | is the
after a single failure and study the robustness of graph classes
number of edges that eventually fail). Namely, if t = |Ft∗ |
based on resistance distance and Kirchhoff ’s index [16, 29].
(one edge fails at each round), the complexity of the CFE-
Recently, these notions have gained attention outside the
PB Algorithm is O(min{|V |, t}) times lower than that of the
Chemistry community. For instance, they were used in net-
CFE Algorithm. The main advantage of the CFE-PB Algo-
work science for detecting communities within a network,
rithm is that it leverages the special structure of the pseudo-
and more generally the strength of the connection between
inverse to identify properties of the underlying graph and to
nodes in a network [34,35]. Moreover, [22] recently used the
recompute an instance of the pseudo-inverse from a previous
resistance distance to partition power systems into zones.
instance.
The problem of identifying the set of failures with the
Finally, we prove that the problem of finding the set of
largest impact was studied in [13, 14, 32, 38]. In particu-
initial failures of size k that causes a cascade with the min-
lar, [14] studies the N − k problem which focuses on find-
imum possible yield (the fraction of demand satisfied af-
ing a small cardinality set of links whose removal disables
ter the cascade) is NP-hard. We introduce a very simple
the network from delivering a minimum amount of demand.
heuristic termed the Most Vulnerable Edges Selection – Re-
A broader network interdiction problem in which all the
sistance distance Based (MVES-RB) Algorithm. We numer-
components of the network are subject to failure was stud-
ically show that solutions obtained by it lead to a much lower
ied in [41]. A similar problem is studied in [38] using the
yield than the solutions obtain by selecting the initial edge
alternating-current (AC) model. However, none of the pre-
failures randomly. Moreover, in some small graphs with a
vious works consider the cascading failures. Moreover, while
single edge failure, it obtains the optimal solution.
the optimal power flow problem has been shown to be NP-
The main contribution of this paper is the development of
hard [31], the complexity of the cascade-related problems
new tools, based on matrix analysis, for assessing the impact
was not studied yet.
of a single edge failure. Using these tools, we (i) obtain upper
Finally, for the simulations, we use graphs that can rep-
bounds on the flow changes after a single failure, (ii) develop
resent the topology of the power grid. The structure of the
1 power grids has been widely studied [5, 6, 9, 18, 23, 24, 44].
The DC model is commonly used in large-scale contin-
gency analysis of power grids [13, 14, 38]. In particular, Watts and Strogatz [44] suggested the small-
2
An n × n admittance matrix represents the admittance world graph as a good representative of the power grid,
of the lines in a power grid with n nodes. based on the shortest paths between nodes and the clus-
3 tering coefficient of the nodes. Barabási and Albert [9,
These notions originate from Circuit Theory and are
widely used in Chemistry [29]. 18] showed that scale-free graphs are better representatives
based on the degree distribution. However, [23] indicated Algorithm 1 - Cascading Failure Evolution (CFE)
that none of these models can represent U.S. Western in- Input: A connected graph G = (V, E) and an initial edge failures
terconnection properly. Following these papers, we consider event F0 ⊆ E.
the Erdős-Rényi graph [26] in addition to these graphs. 1: F0∗ ← F0 and i ← 0.
2: while Fi 6= ∅ do
3: Adjust the total demand to equal the total supply within
3. MODELS AND METRICS each connected component of G = (V, E \ Fi∗ ).
4: Compute the new flows fe (Fi∗ ) ∀e ∈ E \ Fi∗ .
3.1 DC Power Flow Model 5: Find the set of new edge failures Fi+1 = {e|fe (Fi∗ ) >
We adopt the linearized (or DC) power flow model, which ce , e ∈ E \ Fi∗ }. Fi+1
∗ ← Fi∗ ∪ Fi+1 and i ← i + 1.
is widely used as an approximation for the more accurate 6: return t = i − 1, (F0 , . . . , Ft ), and fe (Ft∗ ) ∀e ∈ E\Ft∗ .
non-linear AC power flow model [10]. In particular we fol-
Fi ⊆ E the set of edge failures in the ith round and by
low [11–14] and represent the power grid by an undirected
Fi∗ = Fi−1 ∗
∪ Fi the set of edge failures until the end of
graph G = (V, E) where V and E are the set of nodes and
the ith round (i ≥ 1). We assume that before the initial
edges corresponding to the buses and transmission lines, re-
failure event F0 ⊆ E, the power flows satisfy (1)-(2), and
spectively. pv is the active power supply (pv > 0) or demand
fe ≤ ce ∀e ∈ E. Upon a failure, some edges are removed
(pv < 0) at node v ∈ V (for a neutral node pv = 0). We as-
from the graph, implying that it may become disconnected.
sume pure reactive lines, implying that each edge {u, v} ∈ E
Thus, within each component, the total demand is adjusted
is characterized by its reactance xuv = xvu > 0.
to be equal to the total supply by decreasing the demand
Given the power supply/demand vector P ∈ R|V |×1 and
(supply) by the same factor at all demand (supply) nodes
the reactance values, a power flow is a solution (f, θ) of:
(Line 3). This corresponds to the load shedding/generation
X
fuv = pu , ∀ u ∈ V (1) curtailing process. For any set of failures F ⊆ E, we denote
by fe (F ) the flow along edges in G0 = (V, E \ F ) after the
v∈N (u)
shedding/curtailing.
θu − θv − xuv fuv = 0, ∀ {u, v} ∈ E (2) Following an initial failure event F0 , the new flows fe (F0 ),
where N (u) is the set of neighbors of node u, fuv is the power ∀e ∈ E\F0 are computed (by (1)-(2)) (Line 4). Then, the
flow from node u to node v, and θu is the phase angle of node set of new edge failures F1 is identified (Line 5). Follow-
u. Eq. (1) guarantees (classical) flow conservation and (2) ing [12, 14, 25], we use a deterministic outage rule and as-
captures the dependency of the flow on the reactance values sume, for simplicity, that an edge e fails once the flow ex-
and phase angles. Additionally, (2) implies that fuv = −fvu . ceeds its capacity: fe (F0∗ ) > ce .5 Therefore, F1 = {e :
Note that the edge capacities are not taken into account in fe (F0∗ ) > ce , e ∈ E\F0∗ }.
determining the flows. When the total supply equals the If the set F1 of new edge failures is empty, then the cas-
total demand in each connected component of G, (1)-(2) has cade is terminated. Otherwise, the process is repeated while
a unique solution [14, lemma 1.1].4 Eq.(1)-(2) are equivalent replacing the initial event F0∗ = F0 by the failure event F1∗ ,
to the following matrix equation: and more generally replacing Fi∗ by Fi+1 ∗
at the ith round
(Line 5). The process continues until the system stabilizes,
AΘ = P (3) namely until no edges are removed. Finally, we obtain the
sequence (F0 , F1 , . . . , Ft ) of the sets of failures associated
where Θ ∈ R|V |×1 is the vector of phase angles and A ∈ with the initial event F0 , and the power flows fe (Ft∗ ) at
R|V |×|V | is the admittance matrix of the graph G, defined stabilization, where t is the number of rounds until the net-
as follows: work stabilizes. Since solving a system of linear equations
with n variables, requires O(n3 ) time [27], the output can

0
 if u 6= v and {u, v} ∈
/E
be obtained in O(t|V |3 ) time.
auv = −1/xuv if u 6= v and {u, v} ∈ E

− P An example of a cascade can be seen in Fig. 2. Initially,
w∈N (u) auw if u = v. the flows are fe = 0.5 for all edges. The initial set of failures
If there are (F0 ) disconnects a demand node from the graph. Hence,
P k multiple edges between nodes u and v, then intuitively, one may not expect a cascade. However, this
auv = − ki=1 1/xuvi . Notice that when xuv = 1 ∀{u, v} ∈
E, the admittance matrix A is the Laplacian matrix of the initial failure not only causes further failures but also causes
graph [15]. Once Θ is computed, the power flows, fuv , can failures in all edges except for two. This example can be
be obtained from (2). generalized to a graph with 2n nodes where with the same
Throughout this paper k.k denotes the Euclidean norm of set of initial failures, all the edges fail except for two.
the vector and the operator matrix norm. For matrix Q, qij For simplicity, when the initial failure event contains a
denotes its ij th entry, Qi its ith row, and Qt its transpose. single edge, F0 = {e0 }, we denote the flows after the failure
by fe0 ≡ fe ({e0 }) and the flow changes by ∆fe = fe0 −fe ∀e ∈
3.2 Cascading Failure Model E\{e0 }.
The Cascading Failure Evolution (CFE) Algorithm de-
scribed here is a slightly simplified version of the cascade 3.3 Metrics
model used in [12, 14, 25]. We define fe = |fuv | = |fvu | and To study the effects of a single edge (e0 ) failure after one
assume that an edge e = {u, v} ∈ E has a predetermined round, we define the ratio between the change of flow on an
power capacity ce = cuv = cvu , which bounds its flow (that edge, e, and its original value or the flow value on the failed
is, fe ≤ ce ). The cascade proceeds in rounds. Denote by edge, e0 :
4 5
The uniqueness is in the values of fuv -s rather than θu -s Note that [12, 14, 25] maintain moving averages of the
(shifting all θu -s by equal amounts does not violate (2)). fe values to determine which edges fail.
 0 0

0 0
F0 F2
F1
1/2 0
1/2
-1 1/2
1/2

(a) Initial flows and the fail- (b) Flows and failures due to (c) Flows and failures (F2 ). (d) Stable state.
ure event (F0 ). overload (F1 ).
Figure 2: An example of a cascading failure initiated by outages of the edges connecting a demand node to the network. The
edge capacities and reactance values are ce = 0.6, xe = 1. Numbers in nodes indicate power supply or demand (pv ), numbers
on edges indicate flows (fe ), and arrows indicate flow direction.

Edge flow change ratio: Se,e0 = |∆fe /fe |. Observation 1. If (3) has a feasible solution, Θ̂ = A+ P
Mutual edge flow change ratio: Me,e0 = |∆fe /fe0 |. is a solution for (3).7
Below, we define a metric related to the evaluation of the Proof. According to Theorem A.1, Θ̂ = A+ P minimizes
cascade severity for a given instance G, an initial failure kP − AΘk. On the other hand, since (3) has a solution,
event F0 ⊆ E, and an integer k ≥ 1. An instance is com- kP − AΘ̂k = minΘ kP − AΘk = 0. Thus, Θ̂ = A+ P is a
posed of a connected graph G, supply/demand vector P , solution for (3).
capacities and reactance values ce , xe ∀e ∈ E. For brevity, Jointly verifying whether an edge is a cut-edge and finding
an instance is represented by G. the connected components of the graph takes O(|E|) (us-
Yield (the ratio between the demand supplied at stabi- ing Depth First Search [21]). The following two Lemmas
lization and the original demand): Y (G, F0 ), Y (G, k) = show that by using the precomputed pseudo-inverse of the
minF0 ⊆E,|F0 |≤k Y (G, F0 ). admittance matrix, these operations can be done in O(1)
and O(|V |), respectively. The algorithm in Section ?? uses
3.4 Graphs Used in Simulations the results to check if the pseudo-inverse should be recom-
puted. Moreover, Lemma 1 is crucial for the proof of the
The simulation results are presented for the graphs de-
Theorem 1, below.
scribed below. All graphs have 1,374 nodes to correspond
Lemma 1 (Bapat [8]). Given G = (V, E) and A+ , all
the subgraph of the Western interconnection. The parame-
the cut-edges of the graph G can be found in O(|E|) time.
ters are as indicted below, unless otherwise mentioned.
Specifically, an edge {i, j} ∈ E is a cut-edge if, and only if,
Western interconnection: 1708-edge connected subgraph
a−1 + + +
ij − 2aij + aii + ajj = 0.
of the U.S. Western interconnection. The data is from the
Platts Geographic Information System (GIS) [39]. Lemma 2. Given G = (V, E), A+ , and the cut-edge {i, j},
Erdős-Rényi graph [26]: A random graph where each edge the connected components of G\{i, j} can be identified in
appears with probability p = 0.01. O(|V |).
Watts and Strogatz graph [44]: A small-world random In the following, we denote by A0 the admittance matrix of
graph where each node connects to k = 4 other nodes and the graph G0 = (V, E\{i, j}) and by P 0 the power vector af-
the probability of rewiring is p = 0.1. ter removing an arbitrary edge e0 = {i, j} from the graph G
Barábasi and Albert graph [9]: A scale-free random and conducting the corresponding load shedding/generation
graph where each new node connects to k = 3 other nodes at curtailing.
each step following the preferential attachment mechanism. Lemma 3 shows that after the removal of a cut-edge, A+
can be used to solve (3) and A0+ is not required.
Lemma 3. Given graph G = (V, E), A+ , and a cut-edge
4. ADMITTANCE MATRIX PROPERTIES {i, j}, then Θ̂ = A+ P 0 is a solution of (3) in G0 .
In this section, we use the Moore-Penrose Pseudo-inverse The following theorem gives an analytical rank-1 update of
of the admittance matrix [4] in order to obtain results that the pseudo-inverse of the admittance matrix. Using Theo-
are used throughout the rest of the paper. Specifically they rem 1 and Corollary 1, in Section 5 we provide upper bounds
are used in Section 5 to study the impact of a single edge on the mutual edge flow change ratios (Me,e0 ). We note that
failure on the flows on other edges and in Section 6 to intro- a similar result to Theorem 1 was independently proved in
duce an efficient algorithm to identify the evolution of the a very recent technical report [40].
cascade. We prove several properties of the Pseudo-inverse Theorem 1. Given graph G = (V, E), the admittance
of the admittance matrix A, denoted by A+ .6 A+ always ex- matrix A, and A+ , if {i, j} is not a cut-edge, then,
ists regardless of the structure of the graph G. Some proofs 1
and results that are used in the proofs appear in the Ap- A0+ = (A + aij XX t )+ = A+ − A+ XX t A+
a−1
ij + X t A+ X
pendix.
Observation 1 shows that the power flow equations can be in which X is an n × 1 vector with 1 in ith entry, −1 in j th
solved by using A+ . entry, and 0 elsewhere.
7
Recall from Section 3 that (1)-(2) have a unique solu-
6 + t t 2 −1
A = limδ→0 A (AA +δ I) [4]. For more information tion with respect to power flows but not in respect to phase
regarding the definition, see Appendix. angles. Therefore, the solution to (3) may not be unique.
 Resistance Distance (r(i,j))
35 0.022 3.0 1.0
30 0.020 2.5 0.8
25 0.018
20 2.0 0.6
0.016
15 1.5 0.4
0.014
10 1.0 0.2
5 0.012
0 0.010 0.5 0.0
−5 0.008 0.0 −0.2
−5 0 5 10 15 20 25 30 35 40 1 2 3 0 5 10 15 20 0 1 2 3 4 5 6 7
Distance (d(i,j)) Distance (d(i,j)) Distance (d(i,j)) Distance (d(i,j))
(a) Western interconnection (b) Erdos Renyi graph (c) Watts and Strogatz graph (d) Barabasi and Albert graph
Figure 3: Scatter plot showing the distance versus the resistance distance between nodes in the graphs defined in Subsection 3.4.

For the following, recall from Section 3 that A+ = [a+

rs ].
5. EFFECTS OF A SINGLE EDGE FAILURE
In this section we provide upper bounds on the flow changes
Corollary 1.
after a single edge failure and study the robustness of differ-
0 ars (a+ + + +
ri − arj ) − (asi − asj )
ent graph classes. For simplicity, in this section, we assume
frs = frs − −1 fij . that xe = 1 ∀e ∈ E, unless otherwise indicated. As men-
aij aij − 2(a+ )ij + (a+ )ii + (a+ )jj
tioned in Section 3, in this case the admittance matrix of
Finally, Lemma 4, gives the complexity of the rank-1 update the graph, A, is equivalent to the Laplacian matrix of the
provided in Theorem 1. This is used in the computation of graph. However, all the results can be easily generalized.
the running time of the algorithm in Section 6.
Lemma 4. Given graph G = (V, E), A+ , and an edge
5.1 Flow Changes
{i, j}, which is not a cut-edge of the graph, A0+ can be com-
puted from A+ in O(|V |2 ).
5.1.1 Edge Flow Change Ratio
In order to provide insight into the effects of a single edge
We now define the notion resistance distance [29]. In re- failure, we first present simulation results. The simulations
sistive circuits, the resistance distance between two nodes is have been done in Python using NetworkX library. Fig. 4
the equivalent resistance between them. It is known that shows the edge flow change ratios (Se,e0 ) as the function
the resistance distance, is actually a measure of distance be- of distance (d(e, e0 )) from the failure for over 40 different
tween nodes of the graph [8]. For any network, this notion random choices of an initial edge failure, e0 . The power
can be defined by using the pseudo-inverse of the Lapla- supply/demand in the Western interconnection is based on
cian matrix of the network. Specifically, it can be defined the actual data. In other graphs, the power supply/demand
in power grid networks by using the pseudo-inverse of the at nodes are i.i.d. Normal random variables with a slack
admittance matrix, A+ . node to equalize the supply and demand. Notice that if the
Definition 1. Given G = (V, E), A, and A+ , the resis- initial flow in an edge is close to zero, the edge flow change
tance distance between two nodes i, j ∈ V is r(i, j) := a+ ii + ratio on that edge can be very large. Thus, to focus on
a+ +
jj − 2aij . Accordingly, the resistance distance between two the impact of an edge failure on the edges with reasonable
edges e = {i, j}, e0 = {p, q} is r(e, e0 ) = min{r(i, p), r(i, q), initial flows, we do not illustrate the edge flow change ratios
r(j, p), r(j, q)}. for the edges with flow below 1% of the average flow. Yet,
we observe that such edges that experience a flow increase
When all the edges have the same reactance, xij = 1 ∀{i, j} ∈
after a single edge failure, are within any arbitrary distance
E, the resistance distance between two nodes is a measure
from the initial edge failure.
of their connectivity. Smaller resistance distance between
Fig. 4 shows that after a single edge failure, there might
nodes i and j indicates that they are better connected. Fig. 3
be a very large increase in flows (edge flow change ratios up
shows the relation between the distance and the resistance
to 80, 14, 50, and 24 in Fig. 4-(a), (b), (c), and (d), respec-
distance between nodes in the graphs defined in Subsec-
tively) and sometimes far from the initial edge failure (edge
tion 3.4 (notice that xij = 1 ∀{i, j} ∈ E).8 As can be
flow change ratio around 10 for edges 11- and 4-hops away
seen, there is no direct relation between these two measures
from the initial failure in Fig. 4-(a) and (c), respectively).
in Erdős-Rényi and Barábasi-Albert graphs. However, in
Moreover, as we observed in all of the four graphs, there
the Western interconnection and Watts-Strogatz graph the
are edges with positive flow increase from zero, far from the
resistance distance increases with the distance.
initial edge failure.
In Chemistry, the sum over the resistance distances be-
Finally, we show that by choosing the parameters in a
tween all pairs of nodes in the graph G is referred to as the
specific way, the edge flow change ratio can be arbitrarily
Kirchhoff index [16] of G and denoted by Kf (G). We use
large.
this notion in Subsection 5.2.2 to study the robustness of
different graph classes to single edge failures. Observation 2. For any xe1 , xe2 ∈ R+ \{0}, there exists
a graph G = (V, E) and two edges e1 , e2 ∈ E such that
Definition 2. Given GP= (V, E) and A, the Kirchhoff Se2 ,e1 = xe2 /xe1 .
index of G is Kf (G) = 12 i,j∈V r(i, j).
5.1.2 Mutual Edge Flow Change Ratio
8
While in the Western interconnection the reactance val- We use the notion of resistance distance to find upper
ues depend on the line characteristics (see values in [12]), bounds on the mutual edge flow change ratios (Me,e0 ). The
for comparison and consistency, we used xij = 1 ∀{i, j} ∈ E following Lemma provides a formula for computing the flow
in all the graphs. changes after a single edge failure based on the resistance
 3.0 90 0.9 14 3.0 50 1.0 25
80 0.8 12 2.5
2.5 40 0.8 20
70 0.7
10
2.0 60 0.6 2.0

Max Se,e′
Average Average 8 Average 30 0.6 Average 15
50 0.5
Se,e′

1.5 STD STD 1.5 STD STD

40 0.4 6
Max Max Max 20 0.4 Max 10
1.0 30 0.3 1.0
4
20 0.2 10 0.2 5
0.5 2 0.5
10 0.1
0.0 0 0.0 0 0.0 0 0.0 0
0 5 10 15 20 25 30 35 40 0 1 2 3 4 0 2 4 6 8 10 12 14 16 18 0 1 2 3 4 5
Distance (d(e,e′)) Distance (d(e,e′)) Distance (d(e,e′)) Distance (d(e,e′))
(a) Western interconnection (b) Erdos-Renyi graph (c) Watts and Strogatz graph (d) Barabasi and Albert graph
Figure 4: The average, standard deviation, and maximum edge flow change ratios (Se,e0 ) as the function of distance (d(e, e0 ))
from the failure. The right y-axis shows the values for the maximum edge flow change ratios (max Se,e0 ). The data points are
obtained for 40 different random choices of an initial edge failure.

1.0 With the very same idea, the following corollary gives an
upper bound on the flow changes in a specific edge {i, j} ∈ E
0.9
after a failure in the non-cut-edge {p, q} ∈ E.
0.8
Corollary 3. Given G = (V, E), A, and A+ , the flow
0.7 changes in an edge e = {i, j} ∈ E after a failure in a non-
0.6 cut-edge e0 = {p, q} ∈ E and the mutual edge flow change
(a) Western interconnection (b) Erdos-Renyi (c) Watts and Strogatz ratio M (e, e0 ) can be bounded by,
0.5
0.4 r(e, e0 ) r(e, e0 )
|∆fij | ≤ |fpq |, Me,e0 ≤ .
0.3 1 − r(p, q) 1 − r(p, q)
Corollary 3 directly connects the resistance distance between
0.2
two edges (r(e, e0 )) to their mutual edge flow change ratio
0.1 (Me,e0 ). It shows that the resistance distance, in contrast to
(d) Barabasi and Albert (e) Complete (f) Cycle
0.0 the distance, can be used for assessing the influence of an
edge failure on other edges.
Figure 5: The mutual edge flow change ratios (Me,e0 ) after We present simulations to show the relations between the
an edge failure (represented by black wide line) in different mutual flow change ratios and the two distance measures.
graph classes. All the graphs have 14 nodes, except for (a) Figs. 6 and 7 show the mutual edge flow change ratio (Me,e0 )
which has 28 nodes. In (b) p = 0.1. as the function of distance (d(e, e0 )) and resistance distance
(r(e, e0 )) from the failure, respectively. The figures show
distances. It is independent of the power supply/demand that increasing number of edges (increasing p in Erdős-Rényi
distribution. graph and increasing k in Watts and Strogatz, and Barábasi
Lemma 5. Given G = (V, E), A, and A+ , the flow change and Albert graphs) affects the Me,e0 -r(e, e0 ) relation more
and the mutual edge flow change ratio for an edge e = than the Me,e0 -d(e, e0 ) relation. This suggests that the re-
{i, j} ∈ E after a failure in a non-cut-edge e0 = {p, q} ∈ E sistance distance better captures the information hidden in
are, the structure of a graph. Both figures show a monotone re-
1 −r(i, p) + r(i, q) + r(j, p) − r(j, q) lation between the mutual edge flow change ratios and the
∆fij = fpq , distances/resistance distances. However, this monotonicity
2 1 − r(p, q)
is smoother in the case of the distance.
1 −r(i, p) + r(i, q) + r(j, p) − r(j, q)
Me,e0 = . Moreover, Fig. 6, unlike Fig. 4, shows that after a single
2 1 − r(p, q) edge failure, the mutual edge flow change ratios decrease
Proof. It is an immediate result of Corollary 1. as the distance from the initial failure increases. Thus, it
Fig. 5 illustrates the mutual edge flow change ratios after
suggests that probabilistic tools may be used to model the
an edge failure. Recall that Me,e0 describes the distribu-
mutual edge flow change ratios (Me,e0 ) better than the edge
tion of the flow that passed through e0 on the other edges.
flow change ratios (Se,e0 ).
These values are differently distributed for different graph
classes. In the next subsection, we study in detail the re- 5.2 Graph Robustness
lation between the mutual edge flow change ratios and the
graph structure. We now use the upper bounds provided in Corollaries 2
The following Corollary gives an upper bound on the flow and 3 to study the robustness of some well-known graph
changes after a failure in a non-cut-edge {p, q} ∈ E by using classes to single edge failures. We use the average mutual
the triangle inequality for resistance distance and Lemma 5. edge flow change ratio, Me,e0 , as the measure of the ro-
bustness. The small value of Me,e0 indicates that the flow
Corollary 2. Given G = (V, E), A, and A+ , the flow changes in edges after a single edge failure is small compared
changes in any edge e = {i, j} ∈ E after a failure in a non- to the original flow on the failed edge. In other words, the
cut-edge e0 = {p, q} ∈ E can be bounded by, network is able to distribute additional load after a single
edge failure uniformly between other edges.
r(p, q) r(p, q) We show that (i) graphs with more edges are more ro-
|∆fij | ≤ |fpq |, Me,e0 ≤ .
1 − r(p, q) 1 − r(p, q) bust to single edge failures and (ii) the Kirchhoff index can
 0.08 0.045 0.18 0.06
0.040 p=0.01 0.16 k=4 k=2
0.07
p=0.02 k=6 0.05 k=3
0.06 0.035 0.14
p=0.05 k=8 k=4
Average Me,e′

0.030 0.12 0.04

0.05 p=0.08 k=10 k=5
0.025 0.10
0.04 0.03
0.020 0.08
0.03
0.015 0.06 0.02
0.02 0.010 0.04
0.01
0.01 0.005 0.02
0.00 0.000 0.00 0.00
0 2 4 6 8 10 0 1 2 3 0 2 4 6 8 10 12 14 16 0 1 2 3 4 5 6
Distance (d(e,e′)) Distance (d(e,e′)) Distance (d(e,e′)) Distance (d(e,e′))
(a) Western interconnection (b) Erdos-Renyi graph (c) Watts and Strogatz graph (d) Barabasi and Albert graph
Figure 6: The average mutual flow change ratios (Me,e0 ) versus the distance from the initial edge failure. Each point represents
the average of 40 different initial single edge failure events.
0.08 0.07 0.18 0.05
p=0.01 0.16 k=4 k=2
0.07 0.06
p=0.02 k=6 0.04 k=3
0.06 0.14
0.05 p=0.05 k=8 k=4
Average Me,e′

0.12
0.05 p=0.08 k=10 0.03 k=5
0.04 0.10
0.04
0.03 0.08
0.03 0.02
0.06
0.02
0.02 0.04 0.01
0.01 0.01 0.02
0.00 0.00 0.00 0.00
0 1 2 3 4 5 0.00 0.05 0.10 0.15 0.20 0.25 0.0 0.5 1.0 1.5 2.0 2.5 0.0 0.1 0.2 0.3 0.4 0.5
Resistance distance (r(e,e′)) Resistance distance (r(e,e′)) Resistance distance (r(e,e′)) Resistance distance (r(e,e′))
(a) Western interconnection (b) Erdos-Renyi graph (c) Watts and Strogatz graph (d) Barabasi and Albert graph
Figure 7: The average mutual flow change ratios (Me,e0 ) versus the resistance distance from the initial edge failure. Each
point represents the average of 40 different initial single edge failure events. For clarity, the markers appear for every 5 data
points.
Table 1: The Kirchhoff indices and the average mutual edge
be used as a measure for the robustness of different graph flow change ratios (Me,e0 ) for some well-known graphs. The
classes. values that were previously known [33] are highlighted by
grey cells.
5.2.1 Robustness Based on Number of Edges
Average mutual
Using Corollary 2, it can be seen that a failure in an edge Graph Class Kirchhoff index edge flow change
with small resistance distance between its two end nodes ratio (Me,e0 )
1
leads to a small upper bound on the mutual edge flow change Complete graph n−1 O( n )
Complete 1
ratios, Me,e0 , on the other edges. Thus, the average r(i, j) bipartite graph
4n − 3 O( n )
for {i, j} ∈ E is relatively a good measure of the average Complete 1
2 (9n − 5) O( n12 )
mutual edge flow change ratio. The following Observation tripartite graph
shows that graphs with more edges have smaller average Cycle graph 1
12 (n − 1)n(n + 1) O(n2 )
Cocktail party 2n2 −2n+1
r(i, j) for {i, j} ∈ E, and therefore, smaller average mutual graph n−1
1
O( n )
edge flow change ratio. Erdős-Rényi 1
Θ( n
p)
O( np 2)
graph
Observation 3. Given G = (V, E), the average r(i, j)
for {i, j} ∈ E is |V|E|
|−1
. Lemma 6. For an Erdős-Rényi random graph, G(n, p),
Observation 3 implies that for a fixed number of nodes, the Kf (G) is of Θ( np ), and therefore the average resistance dis-
average resistance distance gets smaller as the number of tance between all pairs of nodes is of Θ( np1 2 ).
edges increases. Therefore, graphs with more edges are more
This Lemma shows that the average resistance distance be-
robust against a single edge failure.
tween all pairs of nodes of an Erdős-Rényi graph is related
to 1/p2 . Since as p grows, the average number of edges in a
5.2.2 Robustness Based on the Graph Class
Erdős-Rényi graph increases, this Lemma also suggests that
Another way of computing the average mutual edge flow graphs with more edges are more robust to a single edge
change ratio is to use Corollary 3 which implies that graphs failure. Thus, the results in this subsection are aligned with
with low average resistance distance over all pairs of nodes the result in Subsection 5.2.1 indicating that graphs with
have the small average mutual edge flow change ratios. On more edges are more robust to a single edge failure.
the other hand, recall from Definition 2 that the average
resistance distance over all pair of nodes is equal to Kirchhoff
index of the graph divided by the number of edges. Hence, 6. EFFICIENT CASCADING FAILURE EVO-
table 1 summarizes the Kirchhoff indices and corresponding LUTION COMPUTATION
average mutual edge flow change ratios for some well-known
Based on the results obtained in Section 4, we present the
graph classes. To complete the table, in the following lemma
Cascading Failure Evolution – Pseudo-inverse Based (CFE-
we compute the Kirchhoff index of the Erdős-Rényi graph
PB) Algorithm which identifies the evolution of the cascade.
as a function of p.
The CFE-PB Algorithm uses the Moore-Penrose Pseudo-
Algorithm 2 - Cascading Failure Evolution – Pseudo- Algorithm 3 - Most Vulnerable Edges Selection – Resis-
inverse Based (CFE-PB) tance distance Based (MVES-RB)
Input: A connected graph G = (V, E) and an initial edge failures Input: A connected graph G = (V, E) and an integer k ≥ 1.
event F0 ⊆ E. 1: Compute A+ .
1: Compute A+ , F0∗ ← F0 and i ← 0. 2: Compute the phase angles Θ̂ = A+ P and compute flows fe
2: while Fi 6= ∅ do from the phase angles.
3: for each {r, s} ∈ Fi do 3: Compute the resistance distance r(i, j) = r(e) ∀e = {i, j} ∈
4: if {r, s} is a cut-edge (see Lemma 1) then E.
5: Find the connected components after removing {r, s}. 4: Sort edges e1 , e2 , . . . , e|E| such that p ≤ q iff fep r(ep ) ≥
(see Lemma 2) feq r(eq ).
6: Adjust the total demand to equal the total supply 5: return e1 , e2 , . . . , ek .
within each connected component.
7: else update A+ after removing {r, s}. (see Lemma 4) k that causes a cascade resulting with the minimum possi-
8: Compute the phase angles Θ̂ = A+ P and compute new ble yield (minimum yield problem). We numerically show
flows fe (Fi∗ ) from the phase angles. that solutions obtained by the heuristic algorithm lead to
9: Find the set of new edge failures Fi+1 = {e|fe > ce , e ∈ a much lower yield than the solutions obtain by selecting
E \ Fi∗ }. Fi+1
∗ ← Fi∗ ∪ Fi+1 and i ← i + 1.
the initial edge failures randomly. Moreover, in some small
10: return t = i − 1, (F0 , . . . , Ft ), and fe (Ft∗ ) ∀e ∈ E\Ft∗ .
graphs with a single edge failure, this algorithm obtains the
inverse of the admittance matrix for solving (3). Comput- optimal solution.
ing the pseudo-inverse of the admittance matrix requires First, we show that deciding if there exists a failure event
O(|V |3 ) time. However, the algorithm obtains the pseudo- (of size at most a given value) such that the yield after sta-
inverse of the admittance matrix in round i from the one bilization is less than a given threshold, is NP-complete.
obtained in round (i − 1), in O(|Fi ||V |2 ) time. Moreover, in Lemma 7. Given a graph G, a real number y, 0 ≤ y ≤ 1,
some cases, the algorithm can reuse the pseudo-inverse from and an integer k ≥ 1, the problem of deciding if Y (G, k) ≤ y
the previous round. Since once lines fail, there is a need for is NP-complete.
low complexity algorithms to control and mitigate the cas- We now present a heuristic algorithm for solving this prob-
cade, the CFE-PB Algorithm may provide insight into the lem. We refer to it as the Most Vulnerable Edge Selection
design of efficient cascade control algorithms. – Resistance distance Based (MVES-RB) Algorithm. From
We now describe the CFE-PB Algorithm. It initially Corollary 2, it seems that edges with large r(i, j) × |fij |
computes the pseudo-inverse of the admittance matrix (in have greater impact on the flow changes on the other edges.
O(|V |3 ) time) and this is the only time in which it computes Based on this result, the MVES-RB Algorithm selects the k
A+ without using a previous version of A+ . Next, starting edges with highest r(i, j) × |fij | values as the initial set of
from F0 , at each round of the cascade, for each e ∈ Fi , it failures.
checks whether e is a cut-edge (Line 4). This is done in O(1) The MVES-RB Algorithm is in the same category as the
(Lemma 1). If yes, based on Lemma 3, in Lines 5 and 6, the algorithms that identify the set of failures with the largest
total demand is adjusted to equal the total supply within impact (i.e., algorithms that solve the N − k problem [14,32,
each connected component (in O(V ) time). Else, in Line 7, 38]). However, none of the previous works focusing on the
A+ after the removal of e is computed in O(|V |2 ) time (see N −k problem, considers cascading failures. The MVES-RB
Lemma 4). After repeating this process for each e ∈ Fi , the Algorithm is simpler than most of the algorithms proposed
phase angles and the flows are computed in O(|V |2 ) time in the past. However, it is not possible to compare its per-
(Line 8). The rest of the process is similar to the CFE Al- formance to that of algorithms in [14, 32, 38, 41] since they
gorithm. use different formulations of the power flow problem.
The following theorem provides the complexity of the al- We first compare via simulation the MVES-RB Algorithm
gorithm (the proof is based on the Lemmas 1–4). We show to the optimal solution in small graphs and for a single initial
that the algorithm runs in O(|V |3 + |Ft∗ ||V |2 ) time (com- edge failure. Fig. 8 shows the yield after stabilization when
pared to the CFE Algorithm which runs in O(t|V |3 )). Namely, selecting a single edge failure based on the MVES-RB Al-
if t = |Ft∗ | (one edge fails at each round), the CFE-PB Algo- gorithm, randomly, and optimally. All the graphs have 136
rithm outperforms the CFE Algorithms by O(min{|V |, t}). nodes. For all the edges the reactance, xe = 1,9 and the ca-
pacity ce = 1.1fe ,10 where fe is the initial flow on the edge.
Theorem 2. CFE-PB Algorithm runs in O(|V |3 +|Ft∗ ||V |2 ) At each point, equal number of power supply and demand
time. nodes are randomly selected and assigned values of 1 and -1.
We notice that a similar approach (the step by step rank-1 As can be seen, the MVES-RB Algorithm obtains the op-
update) can also be applied to other methods for solving lin- timal solution in Erdős-Rényi and Barábasi-Albert graphs.
ear equations (e.g., LU factorization [27]). However, as we However, it does not achieve the optimal solution in the
showed in Section 5, using the pseudo-inverse allows devel- Western interconnection and Watts-Strogatz graph.
oping tools for analyzing the effect of a single edge failure. Finding the optimal solution for the minimum yield prob-
Moreover, it supports the development of an algorithm for lem in the general case is impossible in practice. Therefore,
finding the most vulnerable edges. 9
While in the Western interconnection the reactance val-
ues depend on the line characteristics (see values in [12]),
7. HARDNESS AND HEURISTIC for comparison and consistency, we used xij = 1 ∀{i, j} ∈ E
in all the graphs.
In this section, we prove that the decision problem asso- 10
Following [12], we assume that the capacities are K
ciated with the minimum yield is NP-complete. Using the times the initial flows on the edges. K is often referred
results from Section 5, we introduce a heuristic algorithm to as the Factor of Safety (FoS) of the grid. Here, K = 1.1
for the problem of finding the set of initial failures of size as in [12].
 0.7 0.35 0.25 0.16
Random
0.6 0.30 0.14
MVES-RB
0.20
0.12 OPT
0.5 0.25
0.15 0.10
0.4 0.20
Yield

0.08
0.3 0.15 0.10 0.06
0.2 0.10
0.04
0.05
0.1 0.05 0.02
0.0 0.00 0.00 0.00
0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25
Number of power supply/demand nodes Number of power supply/demand nodes Number of power supply/demand nodes Number of power supply/demand nodes
(a) Western interconnection (b) Erdos-Renyi graph (p =0.2) (c) Watts-Strogatz graph (d) Barabasi-Albert graph
Figure 8: The yield after stabilization when selecting a single edge failure based on the MVES-RB Algorithm, randomly, and
optimally. All graphs have 136 nodes. Every data point is the average over 20 trials, each composed of a different set of
supply/demand nodes.
0.20 0.7 0.012 0.016
MVES-RB
0.6 0.010 0.014 Random
0.15
0.5
0.008 0.012
0.4
Yield

0.10 0.006 0.010

0.3
0.004 0.008
0.2
0.05
0.1 0.002 0.006

0.00 0.0 0.000 0.004

0 5 10 15 20 25 30 0 2 4 6 8 10 12 14 0 5 10 15 20 25 30 0 2 4 6 8 10 12 14
Number of power supply/demand nodes Number of initial edge failures Number of power supply/demand nodes Number of initial edge failures
(a) Western interconnection (b) Western interconnection (c) Watts and Strogatz graph (d) Watts and Strogatz graph
Figure 9: The yield after stabilization when selecting an initial set of edge failures randomly and based on the MVES-RB
Algorithm. In (a) and (c) number of edge failures is 10, and in (b)and (d) the number of supply/demand nodes is 20. Every
data point is the average over 20 trials, each composed of a different set of supply/demand nodes.

to get better insight into the performance of the MVES-RB Hence, there are still many open problems. In particular,
Algorithm, we compare it with the case that k edges are se- we plan to study the effect of failures on the interdependent
lected randomly. As can be seen in Fig. 8, the MVES-RB Al- grid and communication networks. Moreover, while due to
gorithm outperforms the random selection most of the time. its relative simplicity, most previous work in the area of grid
Fig 9 depicts this comparison for larger initial failures in vulnerability is based on the DC model, this model does
the Western interconnection and the Watts-Strogatz graph. not capture effects such as voltage collapse that may occur
The power supplies and demands, the reactances, and the during a cascade. Hence, we plan to develop methods to
capacities are as above. It can be seen that the MVES-RB analyze the cascades using the more realistic AC model.
Algorithm can perform significantly better than the random
selection (Fig. 9-(a) and (b)), and in some cases obtains simi- Acknowledgement
lar performance to the random selection (Fig. 9-(c) and (d)).
Notice that in these cases, both methods perform relatively This work was supported in part by CIAN NSF ERC under
good (lead to yield less than 0.02). grant EEC-0812072, NSF grant CNS-1018379, and DTRA
To conclude, despite the simplicity and low complexity of grant HDTRA1-13-1-0021.
the MVES-RB Algorithm, simulations indicate that it out-
performs the random selection and in simple cases obtains 9. REFERENCES
the optimal solution. [1] U.S.-Canada Power System Outage Task Force. report on the
August 14, 2003 blackout in the United States and Canada:
Causes and recommendations. [Link]
(2004).
8. CONCLUSIONS [2] Report of the enquiry committee on grid disturbance in
Northern region on 30th July 2012 and in Northern, Eastern
We studied properties of the admittance matrix of the grid and North-Eastern region on 31st July 2012, Aug. 2012.
and provided analytical tools for studying the impact of a [Link]
single edge failure on the flows on the other edges. Based on [3] P. Agarwal, A. Efrat, S. Ganjugunte, D. Hay, S. Sankararaman,
these tools, we derived upper bounds on the flow changes and G. Zussman. The resilience of WDM networks to
probabilistic geographical failures. IEEE/ACM Trans. Netw.,
after a single edge failure and discussed the robustness of 21(5):1525–1538, 2013.
different graph classes against single edge failures. We il- [4] A. Albert. Regression and the Moore-Penrose pseudoinverse,
lustrated via simulations the impact of a single edge failure. volume 3. Academic Press, 1972.
Then, we introduced a pseudo-inverse based efficient algo- [5] R. Albert, I. Albert, and G. L. Nakarado. Structural
vulnerability of the North American power grid. Phys. Rev. E,
rithm to identify the evolution of the cascade. Finally, we 69(2):025103, 2004.
proved that the minimum yield problem is NP-hard and in- [6] L. A. N. Amaral, A. Scala, M. Barthélémy, and H. E. Stanley.
troduced a simple heuristic algorithm to detect the most Classes of small-world networks. PNAS, 97(21):11149–11152,
2000.
vulnerable edges.
[7] T. Aura, M. Bishop, and D. Sniegowski. Analyzing single-server
This is one of the first steps in using computational tools network inhibition. In IEEE Proc. Computer Security
for understanding the grid resilience to cascading failures. Foundations Workshop (CSFW-13), 2000.
 [8] R. Bapat. Graphs and matrices. Springer, 2010. [36] J. L. Palacios and J. M. Renom. Bounds for the Kirchhoff index
[9] A.-L. Barabási and R. Albert. Emergence of scaling in random of regular graphs via the spectra of their random walks. Int. J.
networks. Science, 286(5439):509–512, 1999. Quantum Chem., 110(9):1637–1641, 2010.
[10] A. R. Bergen and V. Vittal. Power Systems Analysis. [37] C. Phillips. The network inhibition problem. In Proc. ACM
Prentice-Hall, 1999. STOC’93, May 1993.
[11] A. Bernstein, D. Bienstock, D. Hay, M. Uzunoglu, and [38] A. Pinar, J. Meza, V. Donde, and B. Lesieutre. Optimization
G. Zussman. Sensitivity analysis of the power grid vulnerability strategies for the vulnerability analysis of the electric power
to large-scale cascading failures. ACM SIGMETRICS grid. SIAM J. Optimiz., 20(4):1786–1810, 2010.
Perform. Eval. Rev., 40(3):33–37, 2012. [39] Platts. GIS Data. [Link]
[12] A. Bernstein, D. Bienstock, D. Hay, M. Uzunoglu, and [40] G. Ranjan, Z.-L. Zhang, and D. Boley. Incremental
G. Zussman. Power grid vulnerability to geographically computation of pseudo-inverse of Laplacian: Theory and
correlated failures - analysis and control implications. In Proc. applications. arXiv:1304.2300, Apr. 2013.
IEEE INFOCOM’14, Apr. 2014. [41] J. Salmeron, K. Wood, and R. Baldick. Analysis of electric grid
[13] D. Bienstock. Optimal control of cascading power grid failures. security under terrorist threat. IEEE Trans. Power Syst.,
Proc. IEEE CDC-ECC, Dec. 2011. 19(2):905–912, 2004.
[14] D. Bienstock and A. Verma. The N − k problem in power grids: [42] The Federal Energy Regulatory Comission (FERC) and the
New models, formulations, and numerical experiments. SIAM North American Electric Reliability Corporation (NERC).
J. Optimiz., 20(5):2352–2380, 2010. Arizona-Southern California Outages on September 8, 2011.
[15] N. Biggs. Algebraic graph theory. Cambridge university press, [Link]
1994. [Link].
[16] D. Bonchev, A. T. Balaban, X. Liu, and D. J. Klein. Molecular [43] U.S. Federal Energy Regulatory Commission, Dept. of
cyclicity and centricity of polycyclic graphs. i. cyclicity based Homeland Security, and Dept. of Energy. Detailed technical
on resistance distances or reciprocal distances. Int. J. report on EMP and severe solar flare threats to the U.S. power
Quantum Chem., 50(1):1–20, 1994. grid, Oct. 2010.
[17] S. Buldyrev, R. Parshani, G. Paul, H. Stanley, and S. Havlin. [44] D. J. Watts and S. H. Strogatz. Collective dynamics of
Catastrophic cascade of failures in interdependent networks. small-world networks. Nature, 393(6684):440–442, 1998.
Nature, 464(7291):1025–1028, 2010. [45] H. Xiao and E. M. Yeh. Cascading link failure in the power
[18] D. P. Chassin and C. Posse. Evaluating North American grid: A percolation-based analysis. In Proc. IEEE Int. Work.
electric grid reliability using the Barabási–Albert network on Smart Grid Commun., June 2011.
model. Phys. A, 355(2-4):667 – 677, 2005.
[19] J. Chen, J. S. Thorp, and I. Dobson. Cascading dynamics and
mitigation assessment in power system disturbances via a APPENDIX
hidden failure model. Int. J. Elec. Power and Ener. Sys.,
27(4):318 – 326, 2005. A. PRELIMINARIES AND PROOFS
[20] P. Christiano, J. A. Kelner, A. Madry, D. A. Spielman, and
S.-H. Teng. Electrical flows, Laplacian systems, and faster
In this appendix we restate results related to the Moore-
approximation of maximum flow in undirected graphs. In Proc. Penrose pseudo-inverse of matrix and the proofs for the re-
ACM STOC’11, June 2011. sults in Sections 4, 5, 6, and 7.
[21] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. In the following, matrices I and J denote the identity and
Introduction to algorithms. MIT press, 2009.
[22] E. Cotilla-Sanchez, P. Hines, C. Barrows, S. Blumsack, and
the all-1 matrices, respectively.
M. Patel. Multi-attribute partitioning of power networks based Theorem A.1 (Moore-Penrose [4]). For any n × m
on electrical distance. IEEE Trans. Power Syst., matrix H, Moore-Penrose pseudo-inverse of H,
28(4):4979–4987, 2013.
[23] E. Cotilla-Sanchez, P. D. Hines, C. Barrows, and S. Blumsack. H+ = lim (H t H + δ 2 I)−1 H t = lim H t (HH t + δ 2 I)−1
Comparing the topological and electrical structure of the North δ→0 δ→0
American electric power infrastructure. IEEE Syst. J.,
6(4):616–626, 2012. always exists. And for any n-vector z, x̂ = H + z is the vector
[24] P. Crucitti, V. Latora, and M. Marchiori. A topological analysis of minimum norm among those which minimize kz − Hxk.
of the Italian electric power grid. Phys. A, 338(1):92–97, 2004.
[25] I. Dobson, B. Carreras, V. Lynch, and D. Newman. Complex Theorem A.2 (Albert [4]). For any matrices U, V ,
systems analysis of series of blackouts: cascading failure, critical
points, and self-organization. Chaos, 17(2):026103, 2007. (U U + V V t )+ = (CC t )+ + [I − (V C + )t ]
t

[26] P. Erdős and A. Rényi. On random graphs. Publicationes ×[(U U t )+ − (U U t )+ V (I − C + C)KV t (U U t )+ ]

Mathematicae Debrecen, 6:290–297, 1959.
[27] G. H. Golub and C. F. Van Loan. Matrix Computations. Johns ×[1 − V C + ]
Hopkins Studies in Mathematical Sciences, 4th edition, 2012.
[28] I. Gutman and B. Mohar. The quasi-wiener and the Kirchhoff where C and K are defined as follows
indices coincide. J. Chem. Inf. Comput. Sci., 36(5):982–985,
1996. C = [I − (U U t )(U U t )+ ]V
[29] D. J. Klein and M. Randić. Resistance distance. J. Math.
Chem., 12(1):81–95, 1993. K = {I + [(I − C + C)V t (U U t )+ V (I − C + C)]}−1 .
[30] J. Kleinberg, M. Sandler, and A. Slivkins. Network failure
detection and graph connectivity. In Proc. ACM-SIAM Proof of Lemma 2. Suppose that {i, j} is a cut-edge
SODA’04, Jan. 2004. of the connected graph G, and G\{i, j} = G1 ∪ G2 . Assume
[31] J. Lavaei and S. Low. Zero duality gap in optimal power flow that i ∈ G1 and j ∈ G2 . We show below that for any
problem. IEEE Trans. Power Syst., 27(1):92–107, 2012.
{r, s} ∈ G\{i, j}, a+ + + +
ir − ajr = ais − ajs . Moreover, for any
[32] X. Liu, K. Ren, Y. Yuan, Z. Li, and Q. Wang. Optimal budget
deployment strategy against power grid interdiction. In Proc. r ∈ G1 and s ∈ G2 , air − ajr 6= ais − a+
+ + +
js . Suppose that
IEEE INFOCOM’13, Apr. 2013. {r, s} ∈ G\{i, j} is an arbitrary edge. Then, the solution
[33] I. Lukovits, S. Nikolić, and N. Trinajstić. Resistance distance in
regular graphs. Int. J. of Quantum Chem., 71(3):217–225, to (1)-(2) for the power vector P̂ with p̂r = −p̂s = 1 and zero
1999. elsewhere is frs = −fsr = 1 and zero elsewhere. Therefore,
[34] B. H. McRae. Isolation by resistance. Evolution, fij = 0. On the other hand, from Observation 1, Θ̂ = A+ P̂
60(8):1551–1561, 2006. is a solution to the equivalent matrix equation (3). Since
[35] M. E. Newman and M. Girvan. Finding and evaluating
community structure in networks. Phys. rev. E, 69(2):026113, the solution with respect to power flows is unique, 0 = fij =
2004. −aij (θ̂i − θ̂j ) = −aij (A+ + + + +
i P̂ − Aj P̂ ) ⇒ 0 = (air − ais − ajr +
+ + + + + + +
ajs ) ⇒ air − ajr = ais − ajs . From this and since aii − aji 6=
 Pi−1
a+ + +
mk < j ≤ ik=1 mk , and zero elsewhere. Now similar
P
ij − ajj (Lemma 1), for any r ∈ G1 and s ∈ G2 , air − k=1
a+ + +
jr 6= ais − ajs . Thus, by using the precomputed pseudo-
to previous part,
inverse of the admittance matrix, computing A+ +
i − Aj , and AA+ = I − U t [U1t |U2t | . . . |Ukt |0| . . . |0]t = I − Jk .
dividing the entries into two groups with equal values, the
connected components of G\{i, j} can be identified. This Now we can prove the theorem. A is a real and symmetric
process requires O(|V |) time. matrix, therefore there exist an n × n matrix B such that
BB t = A. Now using Theorem A.2,
Proof of Lemma 3. First, from Observation 1, Θ̂ = √
A+ P 0 is a solution to (3) for the power vector P 0 in the (A + aij XX t )+ = (CC t )+ + [I − ( aij XC + )t ]
graph G. Since the solution to (1)-(2) with respect to power ×[A+ − aij A+ X(I − C + C)KX t A+ ]
flows is unique, if fij = 0, then Θ̂ = A+ P 0 is also a solution √
×[1 − aij XC + ].12
to (3) for the power vector P 0 in the graph G0 . Therefore, we
only need to prove that θ̂i = θ̂j from Θ̂ = A+ P 0 . To prove Therefore, all we need to compute is matrices C and K.
this, we prove that θ̂i − θ̂j = (A+ + 0 Using previous part,
i − Aj )P = 0. However,
from the proof of Lemma 2, since {i, j} is a cut-edge, the C = [I − AA+ ]X = [I − I + Jk ]X = Jk X.
entries of A+ +
i − Aj have equal values at the entries in the
same connected component. On the other hand, since P 0 is Since {i, j} ∈ E, nodes i and j should be in the same con-
the power vector after load shedding/generation curtailing, nected component of G. Therefore, from the structure of
then the sum of the supplies and demands at each connected Jk , Jk X = 0 and so C = 0. Using this,
0
component is zero. Thus, (A+ +
i − Aj )P = 0. K = {I + aij [(I − C + C)X t A+ X(I − C + C)]}−1
Proof of Theorem 111 . First we show that if G is con- = {I + aij [IX t A+ XI]}−1 = {1 + aij X t A+ X}−1 .
nected, then AA+ = I − n1 J. A is a real and symmetric ma-
Notice that X is an n × 1 vector, therefore X t A+ X is an
trix, therefore there exist an orthogonal and unitary matrix
scaler and I in the second equation is 1 × 1. This is why it
U such that A = U t DU , in which D = diag(λ1 , λ2 , . . . , λn )
is written 1 instead of I in the last equation. Since {i, j} is
is the diagonal matrix of eigenvalues of A and Ui is the
not a cut edge, from Lemma 1 we have, 1 + aij X t A+ X =
normalized eigenvector related to eigenvalue λi . It is well-
aij [a−1 + + +
ij − 2(a )ij + (a )ii + (a )jj ] 6= 0, therefore K is well-
known that when G is connected and unweighted, then the
defined. Replacing K and C,
multiplicity of eigenvalue 0 of the Laplacian matrix is 1 [15].
Exactly the same result with the same approach can be ob- (A + aij XX t )+
tained for weighted graph, therefore we can assume that
= A+ − aij A+ X{1 + aij X t A+ X}−1 X t A+
λ1 = 0 and all other eigenvalues are nonzero. In this case
U1 = [ √1n , √1n , . . . , √1n ]. On the other hand, A+ = U t D+ U , 1
= A+ − −1 A+ XX t A+
therefore aij + X t A+ X

AA+ = U t DU U t D+ U = U t DD+ U which is what we wanted to prove.

= U t diag(λ1 λ+ + +
1 , λ2 λ2 , . . . , λn λn )U
Proof of Corollary 1. It is easy to see from Theorem 1,
t
= U (I − diag(1, 0, . . . , 0))U (a+ +
ri − arj )
1 A0+ +
r = Ar − (A+ +
i − Aj ).
= I − U t [U1t |0| . . . |0]t = I − J a−1
ij − 2(a )ij + (a )ii + (a )jj
+ + +
n
0
Using this in frs = −ars (A0+ 0+
r −As )P completes the proof.
in which [U1t |0| . . . |0]t is an n × n matrix with U1 in the first
row and 0 elsewhere. Proof of Lemma 4. Based on Corollary 1, after the re-
Similarly we show that if G has k connected components moval of a non-cut edge {i, j}, each entry of the pseudo in-
with m1 , m2 , . . . , mk nodes, then AA+ = I − Jk in which verse of the admittance matrix can be updated in O(1) time.
Thus, computing A0+ from A+ takes O(|V |2 ) time.
1 1 1
Jk = diag( Jm ×m , Jm ×m , . . . , Jm ×m ) Proof of Observation 2. We construct the graph G =
m1 1 1 m2 2 2 mk k k
(V, E) as follows, V = {s, t}, Ps = −Pt = 1, and there
is a block matrix with matrices on the diagonal entries (with are two parallel edges e1 and e2 between s and t. Set the
proper node indexing). Suppose G has k ≤ n connected capacities ce1 = ce2 = 1. Assume the reactances xe1 , xe2 are
components. Again it is well-known that when G is un- such that 0 < xe1 < xe2 .
xe 2 xe 1
weighted, multiplicity of eigenvalue 0 of the Laplacian ma- By Eq. (1)-(2), we get fe1 = xe +x and fe2 = xe +x .
e 2 1 e 1 2
trix is equal to the number of connected components of xe2
If F0 = {e1 }, then fe2 (F0 ) = 1 and Se2 ,e1 = xe1
.
graph G [15]. With exactly the same reasoning it can be
shown that it is also the case for weighted graph. There-
fore, in this case λ1 = λ2 = · · · = λk = 0. Suppose Proof of Corollary 2. Using triangle inequality for re-
mi is the size of the ith connected component. With a sistance distance, we can write,
proper indexing of nodes, it is easy to verify that Ui = −r(i, p) + r(i, q) ≤ r(p, q)
1 1 1
[0, . . . , 0, √m , . . . , √m , 0, . . . , 0], in which uij = √m for
i i i r(j, p) − r(j, q) ≤ r(p, q).
11
The proof provided could be simplified, if the form of Apply these to Lemma 5 completes the proof.
the A0+ was known in advance. However, the proof shows
12 √
the derivation of A0+ . aij might be an imaginary number.
 m
Proof of Corollary 3. Notice that r({i, j}, {p, q}) = E with |F0 | ≤ k such that Y (G, F0 ) ≤ 1 − |T |
. First, since
min{r(i, q), r(i, p), r(j, q), r(j, p)}. The proof is exactly the all the edges have capacity equal to |T | which is an upper
same as the proof of Corollary 2. bound for a flow in an edge, after initial set of failures, there
PProof of Observation 3. From [8, Lemma 9.9], we have is no cascade. Therefore, there is no further edge failures.
{i,j}∈E r(i, j) = |V | − 1 [8].
Second, with the same reason, as long as a demand node is
connected to the supply node, its demand can be satisfied.
Proof of Lemma 6. It is known that the Kirchhoff in- m
Now since Y (G, F0 ) ≤ 1− |T , with initial set of failure F0 , at
|
dex of the graph G can be written in terms of the eigen-
least m of the demand nodes are disconnected from supply
values
Pn−1of1 the Laplacian matrix of the graph as Kf (G) = node s. In problem 1 choose A = F0 , since the graphs in
n i=1 λi
[28]. On the other hand,
two problems are the same, by removing set of edges A from
n−1 n−1 n−1 G, at least m of the sink nodes are disconnected from source
X 1 X X 1
n2 ≤ ( )( λi ) = ( )tr(A). node s. Since |A| = |F0 | ≤ k, the answer to problem 1 is
i=1
λi i=1 i=1
λi also yes.
However, when n is relatively big, then each node has the It can be concluded from this claim that problems 1 and
degree equal to Θ(np), therefore tr(A) = Θ(n2 p). Combin- 2 are equivalent. Therefore, problem 2 is also NP-complete.
ing this with the equations above, we can easily see that Now since problem 2 is an special case of the minimum yield
Kf (G) = Ω(n/p). Thus, the average resistance distance is problem, the minimum yield problem is NP-hard, and hence
of Kf (G)/|E| = Ω( np1 2 ). its decision version is NP-complete.
As for the upper bound, it is shown in [36] that for a
2
d-regular graph H with n nodes, Kf (H) ≤ 3nd . Using
this bound for Erdős-Rényi graph, we can write Kf (G) =
O(n/p). Thus, the average resistance distance is of O( np1 2 ).
Proof of Theorem 2. Finding the pseudo inverse of
the matrix requires O(|V |3 ) time. Therefore, Line 1 takes
O(|V |3 ) time. Lines 5 and 6 in the algorithm take O(|V |)
time and Line 7 takes O(|V |2 ), therefore the whole for loop
takes at most O(|Fi ||V |2 ) time at each step. Using A+ com-
puted in the for loop, Lines 8 and 9 take O(|V |2 ) time. Thus,
the total running time of the algorithm is at most O(|V |3 ) +
O((|F0 |+|F1 |+· · ·+|Ft |)|V |2 ) = O(|V |3 )+O(|Ft∗ ||V |2 ).
Proof of Lemma 7. Consider following problem:
Problem 1. Suppose G = (V, E) is an instance of the
classical flow problem, with a single source node {s} and set
of sink nodes T . Assume demands are equal to 1 and lines
have unbounded capacity (O(|V |)). Does a subset of edges
A ⊆ E with |A| ≤ k exist such that |Tf ail | ≥ m? (Tf ail
is set of sink nodes which get disconnected from the source
node s after removing set of edges A.)
It is proved in [7, Theorem 7], that problem 1 is NP-complete.
We want to use this result to proof Lemma 7. For this rea-
son we provide a polynomial time reduction from problem
above to minimum yield problem.
Problem 2. Suppose G = (V, E) is an instance of the
power flow problem, with set of supply node S = {s} and set
of demand nodes T . Assume Pt = −1 for all t ∈ T , and
Ps = |T |. Assume all the lines have capacities equal to |T |
m
and reactances equal to 1. Is Y (G, k) ≤ 1 − |T |
?
Claim 1. Suppose the graphs in problems 1 and 2 are the
same, then the answer to problem 1 is yes if, and only if,
the answer to problem 2 is yes.

Proof. (⇒) Assume the answer to problem 1 is yes. It means

that there exists a set of edges A ⊆ E with |A| ≤ k such
that their removal disconnects at least m of the sink nodes
from the source node. Now in problem 2, choose F0 = A.
Since two graphs are the same, at least m of the demand
nodes are disconnected from the supply node s. As a result,
final yield is at most |T | − m. Since initial yield was |T |,
m m
Y (G, F0 ) ≤ 1 − |T |
. Hence, Y (G, k) ≤ 1 − |T |
.
(⇐) Now the other way, assume the answer to problem 2 is
yes. It means that there is an initial set of edge failures F0 ⊆