3224 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 35, NO.

4, JULY 2020

A Markovian Influence Graph Formed From Utility

Line Outage Data to Mitigate Large Cascades
Kai Zhou , Student Member, IEEE, Ian Dobson , Fellow, IEEE, Zhaoyu Wang , Member, IEEE,
Alexander Roitershtein , and Arka P. Ghosh

Abstract—We use observed transmission line outage data to risk [2] and detailed outage data can be used to identify critical
make a Markovian influence graph that describes the probabilities lines [7]. However it is clear that proposed mitigation cannot
of transitions between generations of cascading line outages. Each
be tested and evaluated with historical data. In this paper, we
generation of a cascade consists of a single line outage or multiple
line outages. The new influence graph defines a Markov chain and process historical line outage data to form a Markovian influence
generalizes previous influence graphs by including multiple line graph that statistically describes the interactions between the
outages as Markov chain states. The generalized influence graph observed outages. The Markovian influence graph can quantify
can reproduce the distribution of cascade size in the utility data. the probability of different sizes of cascades, identify critical
In particular, it can estimate the probabilities of small, medium
lines and interactions, and assess the impact of mitigation on the
and large cascades. The influence graph has the key advantage
of allowing the effect of mitigations to be analyzed and readily probability of different sizes of cascades.
tested, which is not available from the observed data. We exploit
the asymptotic properties of the Markov chain to find the lines most
involved in large cascades and show how upgrades to these critical A. Literature Review
lines can reduce the probability of large cascades.
We review the previous literature on influence graphs for
Index Terms—Cascading failures, power system reliability, power grid cascading outages and related topics. There is in-
mitigation, Markov, influence graph.
creasing interest in graphs to represent cascading outages, in
which the graph describes the interaction between outaged
I. INTRODUCTION components and is not the power grid topology. These graphs
of interactions have differences in how they are formed and
ASCADING outages in power transmission systems can
C cause widespread blackouts. These large blackouts are
infrequent, but are high-impact events that occur often enough to
have different names, such as the influence graph, the interaction
graph, the correlation network, and the cascading faults graph.
The idea of a graph of interactions can be traced back to [8] which
pose a substantial risk to society [1], [2]. The power industry has
has a stochastic process at each graph node that interacts with
always analyzed specific blackouts and taken steps to mitigate
different strengths along the graph edges joining to that node to
cascading. However, and especially for the largest blackouts
the other nodes. Rahnamay-Naeini [9] generalizes the model of
of highest risk, the challenges of evaluating and mitigating
interacting and cascading nodes in [8] to include interactions
cascading risk in a quantitative way remain.
within and between two interdependent networks. This type
There are two main approaches to evaluating cascading risk:
of interacting particle system model has some nice properties
simulation and analyzing historical utility data. Cascading sim-
allowing analysis, but remains a somewhat abstract model for
ulations can predict some likely and plausible cascading se-
power system cascading because it is not known how to estimate
quences [3], [4]. However, only a subset of cascading mech-
the model parameters from data.
anisms can be approximated, and simulations are only starting
Influence graphs in their present form were introduced by
to be benchmarked and validated for estimating blackout risk
Hines and Dobson [10], and further developed by Qi, Hines, and
[5], [6]. Historical outage data can be used to estimate blackout
Dobson [11], [12]. These influence graphs describe the statistics
of cascading data with networks whose nodes represent outages
Manuscript received August 13, 2019; revised November 5, 2019 and Decem-
ber 20, 2019; accepted January 26, 2020. Date of publication January 30, 2020; of single transmission lines and whose directed edges represent
date of current version June 22, 2020. This work was supported by NSF under probabilistic interactions between successive line outages. The
Grants 1609080 and 1735354. Paper no. TPWRS-01198-2019. (Corresponding more probable edges correspond to the interactions between line
author: Ian Dobson.)
K. Zhou, I. Dobson, and Z. Wang are with the Department of Electrical and outages that appear more frequently in the data. Cascades in the
Computer Engineering, Iowa State University, Ames IA 50011 USA (e-mail: influence graph start with initial line outages at the nodes and
kzhou@[Link]; dobson@[Link]; wzy@[Link]). spread probabilistically along the directed graph edges. Once the
A. Roitershtein is with the Department of Statistics, Texas A&M University,
College Station TX 77843 USA (e-mail: alexander@[Link]). influence graph is formed from the simulated cascading data, it
A. P. Ghosh is with the Department of Statistics, Iowa State University, Ames, can be used to identify critical components and test mitigation of
IA 50011 USA (e-mail: apghosh@[Link]). blackouts by upgrading the most critical components [11]–[13].
Color versions of one or more of the figures in this article are available online
at [Link] As well as outages of single lines, cascading data typically
Digital Object Identifier 10.1109/TPWRS.2020.2970406 includes multiple line outages that occur nearly simultaneously.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see [Link]
ZHOU et al.: MARKOVIAN INFLUENCE GRAPH FORMED FROM UTILITY LINE OUTAGE DATA TO MITIGATE LARGE CASCADES 3225

When the states are single line outages, these multiple simulta- Various measures are proposed for the identification of critical
neous outages cause problems in obtaining well-defined Markov components based on the influence graph. [11], [12], [17],
chain transitions between states. For example, if the outage of [23] form their specific measures based on their own influ-
two lines causes an outage in the next generation, it is hard to ence/interaction graph. Ma [17] uses a modified page-rank al-
tell which line caused the subsequent outage or whether the two gorithm to find critical lines. Nakarmi [20] forms the influence
lines caused the subsequent outage together. To address this, [12] graph using methods of both [12] and [18], and proposes a
assigns an equal share to the two lines. The resulting influence community-based measure to identify critical components. [20]
graph is then approximated to enable analysis. Qi [11] assumes compares its measure with other centrality measures based on
that the subsequent outage is caused by the most frequent network theory, and concludes that its method performs better
line outage. Improving on this assumption, Qi [14] considers than other methods in most cases. In this paper, our influence
the causal relationships among successive outages as hidden graph is a rigorous Markov chain, and the identification of criti-
variables and uses an expectation maximization algorithm to cal lines is based on the asymptotic quasi-stationary distribution.
estimate the interactions underlying the multiple outage data. The quasi-stationary distribution has a clear interpretation of
In this paper, we solve this problem in a novel way by defining specifying the probabilities that each of the lines is involved in
an additional state for each multiple line outage. Thus our new large cascades.
influence graph generalizes the interaction between single lines The graph of interactions also provides useful information
to multiple line outages, so we do not need to make assumptions about mitigation actions in power system operation. Ju [21]
or approximations when calculating the interactions between extends the interaction graph to a multi-layer graph, in which
two single lines. This enables a Markov chain to be cleanly and the three layers reflect the number of line outages, load shed,
clearly defined. and electrical distance of the cascade spread, respectively. This
Considering the different types of graphs of interactions more multi-layer graph is suggested to mitigate cascades in system
generally, there are three methods of quantifying interactions operation by providing the critical lines at different states of
between components which are the edges in the graph of inter- cascades. Chen [22] proposes a dynamic interaction graph to
actions. First, as explained in the preceding paragraph, in [10]– better support online mitigation actions than a static interaction
[12], the edge corresponds to the conditional probability of a graph. During the propagation of a specific cascade, this dynamic
single line outage given that the previous line has outaged. Sec- interaction graph removes the interactions involving already
ond, in [15]–[17], the edge weight is calculated based on the line outaged lines, and optimal power flow controls the power flow
flow changes due to a single line outage applied to the base case on the critical lines indicated by the dynamic interaction graph.
using a DC load flow (In contrast to [10]–[12] and this paper, this The dynamic interaction graph model reduces the risk of large
implies that the edge weights do not change during the cascade.). cascades more than the static interaction graph.
In Merrill [16], the edge weight is obtained from the line outage As expected, the graph of interactions and any conclusions
distribution factors. In Zhang [15] and Ma [17], the directed drawn depend on the outage data from which the graph is formed.
edge weights are obtained from both the line flow changes and If the outage data is simulated, the selection of initial system
the remaining margin in the line the power is transferred to. states matters. For example, Nakarami [20] shows that different
Then Zhang [15] combines the directed edges to give undirected system states lead to different influence graphs. In this paper, we
edges. On the other hand, Ma [17] retains the directed edges and form our influence graph from fourteen years of public outage
also represents hidden failures by additional nodes. Third, in data of a specific area, so that our influence graph reflects the
Yang [18], the edge corresponds to the correlation between any initial faults and states encountered over that period of time in
two lines. In [19], Carreras constructs a synchronization matrix that power system area. The textbook [24] includes material on
from simulation data from the OPA model to identify the lines both influence and interaction graphs.
with higher overloading probabilities. Other papers [13], [14], Another related line of research is fault chains. A fault chain
[20]–[22] form their graph of interactions similarly to the above as described in [25] is one cascading sequence of line outages.
methods. In this paper, we base the influence graph edges on Each initial line fault gives a fault chain of lines most stressed
conditional probabilities. However, the edges are different than at each step until outage or instability. Usually only the most
the edges in [10]–[12] as they directly correspond to transition stressed or most likely next line outage is selected to form fault
probabilities in a rigorously defined Markov chain. chains. By taking each line in the system as the initial outage
Influence graphs describing the interactions between succes- of each fault chain, Wei [23] obtains a set of fault chains using
sive cascading outages were developed using simulated data a branch loading index to select the most stressed next line to
(Zhou [13] is the exception, but [13] differs from this paper outage. Each fault chain is expressed as a subgraph whose nodes
because it applies the methods of [12] to utility data). But even are transmission lines, and directed edges are branch loading
for simulated cascade data, there remain challenges in extracting assessment indexes, and the union of the subgraphs forms a
good statistics for the influence graph from limited data. Hines, cascading faults graph. The edge weights depend on the sum
Dobson and Qi [10]–[12] estimate the conditional probabilities of the branch loading indices, each scaled by the length of the
of transitions with empirical probabilities. In this paper, we fault they are in. Then critical lines are identified according to
mitigate the limited historical cascading data by using a Bayesian the in- or out-degree of the cascading faults graph. Luo [26]
method and carefully combining the sparser data of the later also forms a cascading faults graph with weights depending on
stages of cascading in a sophisticated way. load loss in the chain, and then uses hypertext-induced topic
3226 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 35, NO. 4, JULY 2020

search to select critical lines. The edge weights of [23], [26] cascades. This allows checking that the estimated proba-
differ from those in influence graphs because they are not based bilities of small, medium and large cascades are accurate
on conditional probabilities. Li and Wu [27] combine simulated enough to be useful.
fault chains into a network and use reinforcement learning to r calculates critical lines most involved in large cascades
explore, evaluate, and find chains most critical to load shed. In directly from the Markov chain as the quasi-stationary
further work, Li and Wu [28] combine simulated fault chains distribution of the Markov chain.
into a state-failure network from which expected load shed can All of these advances clearly distinguish this paper from the
be computed for each state and failure by propagating load previous work reviewed above.
shed backwards accounting for the transition probabilities of
the edges. The transition probabilities are estimated similarly to II. FORMING THE MARKOVIAN INFLUENCE GRAPH FROM
an influence graph by the relative frequency of that transition HISTORICAL OUTAGE DATA
at that stage of the data. However, in contrast to the practice in
We use detailed historical line outage data consisting of
influence graphs, the state transition data for the later stages is not
records of individual automatic transmission line outages that
combined together to get better estimates. Moreover, fault chains
specify the lines outaged and the outage times to the nearest
differ from this paper in only considering single line outages one
minute. We emphasize that this data is routinely recorded by
after another.
utilities worldwide, for example in the North American Trans-
There are also approaches to modeling cascading with
mission Availability Data System.
continuous-time Markov processes. Wang [29] drives line load-
The first step in building an influence graph is to take many
ings with generator and load power fluctuations to determine
cascading sequences of transmission line outages and divide
overloads and outages that change the Markov state and hence
each cascade1 into generations of outages as detailed in [34].
simulate the cascading. Rahnamay-Naeini [30] constructs, using
Each cascade starts with initial line outages in generation 0, and
simulated cascading data and fitted functional forms, a Markov
continues with subsequent generations of line outages 1,2,3, …
process with states highly aggregated into 3 quantities, namely
until the cascade stops. Each generation of line outages is a
the number of failed lines, the maximum of the capacities of all
set of line outages that occur together on a fast time scale of
of the preceding failed lines, and a cascade stopping index. The
less than one minute. Often there is only one line outage in
aggregated Markov process can model the time evolution of the
a generation, but protection actions can act quickly to cause
cascade and the distribution of cascade size. In further work,
several line outages in the same generation. (Sometimes in a
Rahnamay-Naeini reduces the aggregated model to a discrete
cascading sequence an outaged line recloses and outages in a
time Markov chain and generalizes it to model cyber and power
subsequent generation. In contrast to [13], [34], here we neglect
interdependent network cascading interactions in [31] and to
the repeats of these outages.)
model operator actions interacting with cascading in [32].
The influence graph represents cascading as a Markov chain
For another, independent perspective on the literature,
X0 , X1 , . . . , in which Xk is the set of line outages in generation
Nakarmi’s review paper [33] surveys various methods of con-
k of the cascade. We first illustrate the formation of the influence
structing interaction graphs and the reliability analysis based on
graph from artificial cascading data with the simple example of
interaction graphs.
four observed cascades involving three lines shown in Fig. 1. The
first cascade has line 1 outaged in generation 0, line 3 outaged in
B. Contributions of Paper generation 1, line 2 outaged in generation 2, and then the cascade
stops with no lines (indicated by the empty set {}) outaged in
The new influence graph generalizes and improves previous
generation 3. All cascades eventually stop by transitioning to
work in several ways. In particular, this paper
r uses real data observed and routinely collected by utilities and remaining in the state {} for all future generations. The five
states observed in the data are {}, {line 1}, {line 2}, {line 3},
rather than simulated data.
r obtains a clearly defined influence graph that solves the and {line 1, line 3}, where this last state is lines 1 and 3 outaging
together in the same generation, as in generation 1 of cascade 2.
problem of multiple simultaneous outages by using addi-
Introducing the state {line 1, line 3} with two line outages avoids
tional states with multiple outages. This generalized influ-
the problems in previous work in accounting for transitions to
ence graph rigorously defines a Markov chain.
r mitigates the problems of limited cascading data with and from the simultaneous outages of line 1 and line 3.
We can estimate the probabilities of transitioning from state i
several new methods; in particular, it combines Bayesian
to state j in the next generation by counting the number of those
methods of estimation with elaborate methods of distin-
transitions in all the cascades and dividing by the number of oc-
guishing and combining different events. This better esti-
currences of state i. For example, the probability of transitioning
mates the transition matrices of the influence graph while
from state {line 1} to state {line 3} is 1/3 and the probability of
matching the increasing cascade propagation and retaining
possibilities of analysis.
r computes the probabilities of small, medium and large 1 The grouping of line outages into cascades uses the simple method of [34]:

cascades, and these match the historical data statistics. The grouping is done by looking at the gaps in start time between successive line
r makes innovative use of the bootstrap to estimate the outages. If successive outages have a gap of one hour or more, then the outage
after the gap starts a new cascade. More elaborate methods of grouping real line
variance of the probabilities of small, medium and large outages into cascades could be developed and applied.
ZHOU et al.: MARKOVIAN INFLUENCE GRAPH FORMED FROM UTILITY LINE OUTAGE DATA TO MITIGATE LARGE CASCADES 3227

Fig. 1. Simple example forming influence graph from artificial data (real utility
data is shown in Fig. 2).

transitioning from state {line 2} to state {line 1, line 3} is 1/2.

The probability of transitioning from state {line 1} to {}; that is,
stopping after the single outage of line 1, is 2/3. The probabilities
of the edges out of each state sum to 1. By working out all
the transition probabilities, we can make the network graph of
the Markov chain as shown in Fig. 1. The transitions between
states with higher probability are shown with thicker lines. In this
generalized influence graph, the nodes are sets of line outages
and the edges indicate transitions or interactions between sets
Fig. 2. The gray network is the system network and the red network is
of line outages in successive generations of cascading. The the influence graph showing the main influences between lines. The red edge
influence graph is different than the physical grid network and thickness indicates the strength of the influence.
cascades are generated in the influence graph by moving along
successive edges, selecting them according to their transition
is postponed to Section VI, and until Section VI we assume
probabilities.
that P k has already been estimated for each generation k from
In the general case, there are many states s0 , s1 , . . . , and we
the utility data. Forming the Markov chain transition matrix
describe the transitions between them. Let P k be the Markov
from the data in this way makes the Markovian assumption
chain transition matrix for generation k. The P k matrix entry
that the statistics of the lines outaged in a generation only
Pk [i, j] is the conditional probability that the set of outaged lines
depend on the lines outaged in the previous generation. This
is sj in generation k + 1, given that the set of outaged lines is
is a pragmatic assumption that yields a tractable data-driven
si in generation k; that is,
probabilistic model of cascading.
Pk [i, j] = P[Xk+1 = sj | Xk = si ]. (1) One way to visualize the influence graph interactions between
line outages in P k is to restrict attention to the interactions
The key task of forming the Markov chain is to estimate the between single line states, and show these as the red network
transition probabilities in the matrix P k from the cascading data. in Fig. 2. The gray network is the actual grid topology, and the
If one supposed that P k does not depend on k, a straightforward gray transmission lines are joined by a red line of the thickness
way to do this would first construct a counting matrix N whose proportional to the probability of being in successive genera-
entry N [i, j] is the number of transitions from si to sj among tions, if that probability is sufficiently large. The interactions in
all generations in all the cascades. Then P k would be estimated Fig. 2 reflect a very wide range of mechanisms. The longer-range
as mechanisms include redistributions of power due to line and
N [i, j] generator outages, remedial action schemes, and bad weather
Pk [i, j] =  . (2) across the grid.
j N [i, j]
Let the row vector π k be the probability distribution of states
However, we find that P k must depend on k in order to reproduce in generation k. The π k entry πk [i] is the probability that the set
the increasing propagation of outages observed in the data [34]. of outaged lines is si in generation k; that is,
On the other hand, there is not enough data to accurately estimate
P k individually for each k > 0. Our solution to this problem πk [i] = P[Xk = si ]. (3)
involves both grouping together data for higher generations Then the propagation of sets of line outages from generation k
and having P k varying with k, as well as using empirical to generation k + 1 is given by
Bayesian methods to improve the required estimates of cascade
stopping probabilities. The detailed explanation of this solution π k+1 = π k P k (4)
3228 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 35, NO. 4, JULY 2020

and, using (4), the distribution of states in generation k depends

on the initial distribution of states π 0 according to

π k = π 0 P 0 P 1 · · ·P k−2 P k−1 . (5)

III. ILLUSTRATIVE HISTORICAL OUTAGE DATA

While our method applies generally to the detailed outage
data routinely collected by utilities, we illustrate our method
with a specific publicly available data set, which is the automatic
transmission line outages recorded by a large North American
utility over 14 years starting in 1999 [35]. We group the 9,741
line outages into 6,687 cascades [34]. Most of the cascades Fig. 3. Survival functions of the number of generations from real data and
from the Markov chain.
(87%) have one generation because initial outages often do not
propagate further. There are 614 lines and the observed cascades
have 1094 subsets of these lines that form the 1094 states s0 ,
s1 , . . . , s1093 . Among these 1094 states, 50% have multi-line
outages. And among these multi-line outage states, about 20%
are comprised of lines sharing no common buses. While in theory
there are 2614 subsets of 614 lines, giving an impractically large
number of states, we find in practice with our data that the
number of states is less than twice the number of lines. Note
that our statistical modeling approximates the power grid as
unchanging over the time span of the data [36]. In practice a
utility would have the records of changes to partially mitigate
the effects of this approximation.
Fig. 4. Survival function of cascade sizes. Red crosses are from Markov chain,
and blue lines indicate the 95% confidence interval estimated by bootstrap.
IV. COMPUTING THE DISTRIBUTION OF CASCADE SIZES
AND ITS CONFIDENCE INTERVAL

We compute the distribution of cascade sizes from the Markov sample resamples the observed cascades with replacement, re-
chain and check that it reproduces the empirical distribution constructs the Markov chain, and recomputes the probabilities
of cascade sizes, and estimate its confidence interval with a of cascade sizes. Note that each bootstrap resampling amounts
bootstrap. to a different selection of the cascades observed in the data. The
We can measure the cascade size by its number of generations. variance of the probabilities of cascade sizes is then obtained
Define the survival function of the number of generations in a as the empirical variance of the bootstrap samples. We use 500
cascade as bootstrap samples to ensure a sufficiently accurate estimate of
the variance of the probabilities.
S(k) = P[number of cascade generations > k] (6)
The risk of a given size of blackout is estimated as risk =
S(k) = 1 − πk [0], where πk [0] is the probability that a cascade (estimated probability p̂ of that size of blackout) × (cost of
is in state s0 = {} in generation k and also the probability that that size of blackout). Knowing the multiplicative uncertainty
the cascade stops at or before generation k. Hence in p̂ is useful. For example, if we know p̂ to within a factor of
2, then this contributes a factor of 2 to the uncertainty of the
S(k) = 1 − πk [0] = π k (1 − e0 ) risk. Therefore, it is appropriate to use a multiplicative form of
= π 0 P 0 P 1 · · ·P k−2 P k−1 (1 − e0 ), (7) confidence interval for p̂ specified by a parameter κ. A 95%
multiplicative confidence interval for an estimated probability p̂
where 1 is the column vector (1, 1, 1, . . ., 1) , and e0 is the means that the probability p satisfies P[p̂/κ ≤ p ≤ p̂ κ] = 0.95.
column vector (1, 0, 0, 0, . . ., 0) . The initial state distribution The confidence interval for the estimated survival function is
π 0 can be estimated directly from the cascading data. shown in Fig. 4. Since larger cascades are rarer than small
Then we can confirm that the influence graph reproduces the cascades, the variation increases as the number of generations
statistics of cascade size in the cascading data by comparing increases.
the survival function S(k) computed from (7) with the empirical To apply and communicate the probability distribution of
survival function computed directly from the cascading data as cascade size, it is convenient to combine sizes together to get
shown in Fig. 3. The Markov chain reproduces the statistics the probabilities of small, medium, and large cascades, where
of cascade size closely, with a Pearson χ2 goodness-of-fit test a small cascade has 1 or 2 generations, a medium cascade has
p-value of 0.99. 3 to 9 generations, and a large cascade has 10 or more genera-
We use bootstrap resampling [37] to estimate the variance tions. (The respective probabilities are calculated as 1 − S(2),
of our estimates of probabilities of cascade sizes. A bootstrap S(2) − S(9), and S(9)). The 95% confidence intervals of the
ZHOU et al.: MARKOVIAN INFLUENCE GRAPH FORMED FROM UTILITY LINE OUTAGE DATA TO MITIGATE LARGE CASCADES 3229

TABLE I
95% CONFIDENCE INTERVALS USING BOOTSTRAP

estimated probabilities of small, medium, and large cascades are

shown in Table I. The probability of large cascades is estimated
within a factor of 1.5, which is adequate for the purposes of
estimating large cascade risk, since the cost of large cascades
is so poorly known: estimates of the direct costs of cascading
Fig. 5. Quasi-stationary distribution of transmission lines eventually involved
blackouts vary by more than a factor of 2. in propagating cascades. Red dots are ten critical lines.
We now discuss tracking cascades by their number of gen-
erations. The number of generations is the same concept as
the number of tiers in commercial cascading software [38]. entries in d∞ are the critical states most involved in the latter
Basic to cascading analysis is the grouping of line outages into portion of large cascades. Since d∞ does not depend on the initial
successive generations within each cascade. This grouping is outages, the Markov chain is supplying information about the
usually done by outage timing as in this paper, or by simulation eventual cascading for all initial outages.
loops naturally producing generations of outages. This paper We now find the critical lines corresponding to these critical
is structured in terms of these generations, so that propagation states by projecting the states onto the lines in those states. Let
is determined by the probability of a next generation (i.e. the k be the row vector whose entry k [j] is the probability that
cascade not stopping at the current generation), and cascade line j outages in generation k. Then
size is measured by number of cascade generations. In contrast, 
some previous papers [7], [12], [13], [34] are structured in terms k [j] = πk [i] or k = π k R, (8)
of the line outages in the generations, so that, according to i:j∈si
the branching process model [34], each line outage in each where the matrix R projects states to lines according to
generation propagates independently to form line outages in 
the next generation. Then the propagation is determined by the 1; line j ∈ si
R[i, j] = (9)
number of line outages per line outage in the previous generation, 0; line j ∈ / si
and it is natural to use the total number of lines outaged as
a measure of cascade size. While it is not yet clear which Then the probability distribution of lines eventually involved
approach is better, there may be some advantages to tracking in the propagating cascades that are not stopped is c∞ = d∞ R
cascades by generations rather than line outages. Generations and the critical lines most involved in the latter portion of large
are simpler and more general than line outages, and can more cascades correspond to the highest probability entries in c∞ .
easily encompass other outages significant in cascading such as Fig. 5 shows the probabilities in c∞ in order of decreasing
transformer outages. Also, it may be that the statistics of the probability. We identify the top ten lines as critical and as
number of generations is more simply described, as in the Zipf candidates for upgrading to decrease the probability of large
distribution observed in utility data in [39]. cascades.

B. Modeling and Testing Mitigation in the Markov Chain

V. CRITICAL LINES AND CASCADE MITIGATION
A transmission line is less likely to fail due to other line
A. The Transmission Lines Involved in Large Cascades
outages after the line is upgraded, its protection is improved, or
The lines eventually most involved in large cascades can be its operating limit is reduced. These mitigations have the effect
calculated from the asymptotic properties of the Markov chain. of decreasing the probability of transition to states containing
While all cascades eventually stop, we can consider at each the upgraded line, and are an adjustment of the columns of the
generation those propagating cascades that are not stopped at transition matrix corresponding to these states. The mitigation is
that generation. The probability distribution of states involved represented in the Markov chain by reducing the probability of
in these propagating cascades converges to a probability distribu- transition to the state s containing the upgraded line by (r/|s|)%,
tion d∞ , which is called the quasi-stationary distribution. d∞ can where |s| is the number of lines in the state. The reduction is r%
be computed directly from the transition matrices (as explained if the state contains only the upgraded line, and the reduction is
in Appendix, d∞ is the left eigenvector corresponding to the less if the state contains multiple lines.
dominant eigenvalue of the transition submatrix Q̄1+ ). That is, We demonstrate using the Markov chain to quantify the
except for a transient that dies out after some initial generations, impact of mitigation by upgrading the ten lines critical for
the participation of states in the cascading that continues past large cascades identified in Section V-A with r = 80%. The
these initial generations is well approximated by d∞ . Thus the effect of this mitigation on cascade probabilities is shown in
high probability states corresponding to the highest probability Fig. 6. It shows that upgrading the critical lines reduces the
3230 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 35, NO. 4, JULY 2020

and the transition matrix, but here it is convenient to discuss

them separately.
This paper aims to select the lines critical for large cascades
and quantify the impact on cascade probability of generic up-
grades to these lines. Once the critical lines are selected, an en-
gineering process of much wider scope is required to determine
the possible approaches to upgrade each of the lines, quantify the
benefits other than reducing large cascades and balance the costs
and feasibilities of the upgrading approaches against the total
benefits of upgrading. One part of this process is that for each
line, the percentage reduction in outage probability for the best
Fig. 6. Cascade size distribution before (red) and after (light green) mitigating
lines critical in propagating large cascades. approach to line upgrade is estimated and the Markov chain is
used to quantify the corresponding reduction in large, medium,
and small cascade probabilities. However, cascade mitigation
is only one of the many factors to be considered in justifying
probability of large cascades by 45%, while the probability of upgrade. Evaluating and costing specific upgrading approaches
medium cascades is slightly decreased and the probability of for specific lines requires utility expertise, including details of
small cascades is slightly increased. the line construction and right of way, maintenance history, and
To show the effectiveness of the method of identifying critical operation.
lines, we compare the mitigation effect of upgrading critical
lines and upgrading ten random lines. Randomly upgrading ten VI. ESTIMATING THE TRANSITION MATRIX
lines only decreases the probability of large cascades by 11% The Markov chain has an absorbing first state s0 = {}, indi-
on average. cating no lines outaged as the cascade stops and after the cascade
So far we have only considered upgrading the lines critical for stops. Therefore the transition matrix has the structure
propagating large cascades. Now, in order to discuss this miti- ⎡ ⎤
gation of large cascades in a larger context, we briefly consider 1 0 ··· 0
⎢ ⎥
and contrast a different mitigation tactic of upgrading lines that ⎢ ⎥
are critical for initial outages. Since initial outages are caused by Pk = ⎢ ⎥ (10)
⎣ uk Qk ⎦
external causes such as storm, lightning, or misoperation, they
often have different mechanisms and different mitigations than
for propagating outages. A straightforward method to identify where uk is a column vector of stopping probabilities; that is,
lines critical for initial outages selects the ten lines in the data uk [i] = Pk [i, 0]. Qk is a submatrix of transition probabilities
with the highest frequencies of initial outage [13]. Upgrading between transient states which contains the non-stopping prob-
these ten lines will reduce their initial outage frequencies and abilities. The first row of P k is always e0 , so the transition
hence reduce the overall cascade frequency. In the Markov chain, probabilities to be estimated are uk and Qk for each generation
this upgrading is represented by reducing in the first generation k. The rows and columns of P k are indexed from 0 to |S| − 1
the frequency of states s that contain the critical lines for initial and the rows and columns of Qk are indexed from 1 to |S| − 1,
outages by r/|s|%, where r = 80%. The main effect is that where |S| is the number of states.
by reducing the initial outage frequencies of the critical lines As summarized in Section II after (1), we need to both
by 80%, we reduce the frequency of all cascades by 19%. In group together multiple generations to get sufficient data and
addition, this mitigation will change the probabilities of states account for variation with generation k. The statistics of the
π 0 after renormalizing the frequencies of states. It turns out for transition from generation 0 to generation 1 are different than
our case that there is no overlap between critical lines for initial the statistics of the transitions between the subsequent gen-
outages and for propagation. erations. For example, stopping probabilities for generation 0
Changing the initial state distribution π 0 has no effect on are usually larger than stopping probabilities for subsequent
the distribution of cascade sizes in the long-term. However, it generations [13]. Also, the data for the subsequent generations
directly reduces the frequency of all cascades. In contrast, mit- is sparser. Therefore, we construct from counts of the number
igating the lines critical for propagating large cascades reduces of transitions from generation 0 to generation 1 a probability
the probability of large cascades relative to all cascades but has transition matrix P̄ 0 , and construct from the total counts of
no effect on the frequency of all cascades. (Note that Fig. 6 the number of transitions from all the subsequent generations a
shows the distribution of cascade sizes assuming that there is probability transition matrix P̄ 1+ . Specifically, we first use the
a cascade, but gives no information about the frequency of all right-hand side of (2) to construct two corresponding empirical
cascades.) transition matrices, and then we update stopping probabilities
In practice, a given mitigation measure can affect both the ini- by the empirical Bayes method and adjust non-stopping proba-
tial outages and the propagation of outages into large cascades. bilities to obtain P̄ 0 and P̄ 1+ . Finally, we adjust P̄ 0 and P̄ 1+
The combined mitigation effects can also be represented in the to match the observed propagation rates to obtain P k for each
influence graph by changing both the initial state distribution generation k.
ZHOU et al.: MARKOVIAN INFLUENCE GRAPH FORMED FROM UTILITY LINE OUTAGE DATA TO MITIGATE LARGE CASCADES 3231

A. Bayesian Update of Stopping Probabilities

The empirical stopping probabilities are improved by an
empirical Bayes method [40], [41] to help mitigate the sparse
data for some of these probabilities. Since the method is applied
to both P̄ 0 and P̄ 1+ , we simplify notation by writing P̄ for
either P̄ 0 or P̄ 1+ .
The matrix of empirical probabilities obtained from the tran-
sition counts N [i, j] is
N [i, j]
P̄ counts [i, j] =  (11)
j N [i, j]
counts
We construct P̄ from P̄ in two steps. First, Bayesian
updating is used to better estimate stopping probabilities and
bayes
form a matrix P̄ . Second, the non-stopping probabilities
bayes
in P̄ are adjusted to form the matrix P̄ to account for the
fact that some independent outages are grouped into cascading
outages when we group outage data into cascades.
We need to estimate the probability of the cascade stopping
at the next generation for each state encountered in the cascade.
For some of the states, the stopping counts are low, and cannot
give good estimates of the stopping probability. However, by
pooling the data for all the states we can get a good estimate of
the mean probability of stopping over all the states. We use this
Fig. 7. Stopping probabilities before and after Bayesian updating.
mean probability to adjust the sparse counts in a conservative
way. In particular, we form a prior that maximizes its entropy
subject to the mean of the prior being the mean of the pooled
data. This maximum entropy prior can be interpreted as the prior We want to estimate hyperparameters β1 , β2 to make the
distribution that makes the least possible further assumptions beta distribution have maximum entropy subject to the mean
about the data [42], [43]. being the average stopping probability of the pooled data û =
|S|−1 |S|−1
1) Finding a Maximum Entropy Prior: Assuming the stop- ( i=1 ci )/( i=1 ni ). Then we can obtain hyperparameters
ping counts are independent with a common probability, the β1 , β2 by finding the m > 0 that maximizes Ent(m, û) and
stopping counts follow a binomial distribution. Its conjugate evaluating β1 = mû and β2 = m(1 − û). The hyperparameters
bayes
prior distribution is the beta distribution, whose parameters are used for P̄ 0 are (β1 , β2 ) = (2.18, 0.32), and the hyperpa-
estimated using the maximum entropy method. bayes
rameters for P̄ 1+ are (β1 , β2 ) = (1.10, 0.93).
Let stopping counts Ci be the observed number of transitions 2) Updating the Observed Data Using the Prior: The pos-
from state si to s0 (i = 1, . . . , |S| − 1). Then Ci = N [i, 0]. Let terior distribution of the stopping probability Ui is a beta distri-
|S|−1
ni = j=0 N [i, j] be the row sum of the counting matrix N . bution with parameters ci + β1 , ni − ci + β2 . We use the mean
The stopping counts Ci follow a binomial distribution with of the posterior distribution as a point estimate of the stopping
parameter Ui , with probability mass function probability:
ni !
fCi |Ui (ci |ui ) = uci (1 − ui )ni −ci (12) c i + β1
ci !(ni − ci )! i P̄ bayes [i, 0] = E(Ui |Ci = ci ) = (15)
ni + β1 + β2
The conjugate prior distribution for the binomial distribution is
the beta distribution. Accordingly, we use the beta distribution Fig. 7 shows a comparison between the empirical stopping
with hyperparameters β1 , β2 for the stopping probability Ui : probabilities and the updated stopping probabilities. Black dots
are the empirical probabilities sorted in ascending order (if two
fUi (ui ) = B(β1 , β2 )uiβ1 −1 (1 − ui )β2 −1 (13) probabilities are equal, they are sorted according to the total
Γ(β1 +β2 ) counts observed). Red dots are the updated stopping probabil-
where B(β1 , β2 ) = Γ(β 1 )Γ(β2 )
. Alternative parameters for the
ities. As expected, the empirical probabilities with the fewest
beta distribution are its precision m = β1 + β2 and its mean
counts move towards the mean the most when updated. As the
μ = β1β+β
1
. The entropy of the beta distribution is
2 counts increase, the effect of the prior decreases and the updated
Ent(m, μ) = ln B(mμ, m(1 − μ)) − (mμ − 1)ψ(mμ) probabilities tend to the empirical probabilities.
bayes
Equation (15) forms the first column of P̄ . Then
− (m(1 − μ) − 1)ψ(m(1 − μ)) + (m − 2)ψ(m) (14)
the nonstopping probabilities in the rest of the columns of
counts
where ψ(x) is the digamma function. the P̄ matrix are scaled so that they sum to one minus the
3232 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 35, NO. 4, JULY 2020

TABLE II The matrix Ak has the effect of transferring a fraction of

PROPAGATIONS OF GENERATIONS k = 0 TO 17
probability from the transient to stopping transitions and has the
following form:
⎛ ⎞
1 0 ... 0
⎜ ⎟
⎜ak 1 − ak . . . 0 ⎟
⎜
Ak = ⎜ . ⎟ (19)
.. ⎟
⎝ .. . ⎠
bayes
stopping probabilities of (15) to complete the matrix P̄ : ak 0 . . . 1 − ak
1 − P̄ bayes [i, 0] ak is determined from the estimated propagation rate ρ̂k as
P̄ bayes [i, j] = |S|−1 P̄ counts [i, j], j > 0 (16)
P̄ counts [i, r] follows. Using (18), we have
r=1
bayes π k P̄ Ak (1 − e0 ) 1 − π k P̄ e0
This Bayesian updating is applied to form P̄ 0 for the first ρ̂k = = (1 − ak ) (20)
bayes
transition and P̄ 1+ for the subsequent transitions. π k (1 − e0 ) 1 − π k e0
and we solve (20) to obtain ak for each generation k.
B. Adjust Nonstopping Probabilities for Independent Outages
The method explained in Section II that groups outages into VII. DISCUSSION AND CONCLUSION
cascades has an estimated 6% chance that it groups independent We process observed transmission line outage utility data to
outages into cascading outages [36]. These 6% of outages occur form a generalized influence graph and the associated Markov
independently while the cascading of other outages proceeds and chain that statistically describe cascading outages in the data.
do not arise from interactions with other outages. The empirical Successive line outages, or, more precisely, successive sets of
data for the nonstopping probabilities includes these 6% of near simultaneous line outages in the cascading data correspond
outages, and we want to correct this. Therefore, the non-stopping to transitions between nodes of the influence graph and tran-
probabilities are modified by shrinking the probabilities in tran- sitions in the Markov chain. The more frequently occurring
sition matrix by 6%, and sharing this equally among all the states. successive line outages in the cascading data give a stronger
That is, influence between nodes and higher transition probabilities. The
0.06 generalized influence graph introduces additional states corre-
P̄ [i, j] = 0.94P̄ bayes [i, j] + (1 − P̄ bayes [i, 0]) (17) sponding to multiple line outages that occur nearly simultane-
|S| − 1
ously. This innovation adds a manageable number of additional
where P̄ bayes indicates the transition matrices after the Bayesian states and solves some problems with previous influence graphs,
updateof Section VI-A. Notice that P̄ is a probability matrix making the formation of the Markov chain clearer and more
since j P̄ (i, j) = 1 for each i. A benefit is that this adjustment rigorous.
makes the submatrix Qk have non-zero off-diagonal entries, One of the inherent challenges of cascading is the sparse data
making P̄ irreducible. for large cascades. We have used several methods to partially
alleviate this when estimating the Markov chain transition ma-
C. Adjustments to Match Propagation trices, including combining data for several generations, conser-
vatively improving estimates of stopping probabilities with an
The average propagation ρk for generation k [34] is estimated
empirical Bayes method, accounting for independent outages
from the data using
during the cascade, and matching the observed propagation
Number of cascades with > k + 1 generations for each generation. The combined effect of these methods is
ρ̂k =
Number of cascades with > k generations to improve estimates of the Markov chain transition matrices.
Although some individual elements of these transition matri-
S(k + 1) π k+1 (1 − e0 )
= = (18) ces are nevertheless still poorly estimated, what matters is the
S(k) π k (1 − e0 )
variability of the results from the Markov chain, which are the
An important feature of the cascading data is that average probabilities of small, medium and large cascades. We assess
propagation ρk increases with generation k as shown in Table II. the variability of these estimated probabilities with a bootstrap
To do this, we need to form transition matrices for each of these and find them to be estimated to a useful accuracy. This assess-
generations that reproduce this propagation. We define a matrix ment of variability is necessary for getting useful estimates of
Ak to adjust P̄ 0 and P̄ 1+ so that the propagation in P k matches large cascade probability because large cascades are rare, and
the empirical propagation for each generation up to generation probability estimates for rare events have the potential to be so
8. For generation 9 and above, the empirical propagation for wildly variable that they are useless.
each generation is too noisy to use individually and we combine The Markov chain only models the statistics of successive
those generations to obtain a constant transition matrix. That transitions in the observed data. Also, there is an inherent limi-
is, P 0 = P̄ 0 A0 , P 1 = P̄ 1+ A1 , . . . , P 8 = P̄ 1+ A8 , P 9+ = tation of not being able to account for transitions and states not
P̄ 1+ A9+ . Then the transition matrices for all the generations are present in the observed data. That is, the common transitions
P 0 , P 1 , P 2 , P 3 , P 4 , P 5 , P 6 , P 7 , P 8 , P 9+ , P 9+ , P 9+ , . . .. and states and some of the rarer transitions and states will be
ZHOU et al.: MARKOVIAN INFLUENCE GRAPH FORMED FROM UTILITY LINE OUTAGE DATA TO MITIGATE LARGE CASCADES 3233

present in the data and will be represented in the Markov model, k given that the cascade is propagating, that is
while the rarer transitions and states not present in the data
P[Xk = si ] πk [i]
will be neglected. However, the Markov chain can produce, in dk [i] = = , i = 1, . . ., |S|
addition to the observed cascades, combinations of the observed P[Xk = s0 ] 1 − πk [0]
transitions that are different than and much more extensive than Then the quasi-stationary distribution is d∞ = limk→∞ dk .
the observed cascades. The Markov chain approximates the Diagonal entries of Q̄1+ corresponding to P̄ 1+ are all zero
statistics of cascading rather than reproducing only the observed and all other entries are positive. According to the Perron-
cascades. Frobenius theorem [46], Q̄1+ has a unique maximum modu-
We exploit the asymptotic properties of the Markov chain to lus eigenvalue μ, which is real, positive and simple with left
calculate the transmission lines most involved in the propagation eigenvector v  . By normalizing v  , we make v  a probability
of larger cascades, and we show with the Markov chain that vector. We write w for the corresponding right eigenvector.
upgrading these lines can significantly reduce the probability of Moreover, 0 < μ < 1 and μ is strictly greater than the modulus
large cascades. Since a large cascade of line outages with many of the other eigenvalues of Q̄1+ . Suppose the cascade starts
generations is very likely to shed substantial load, mitigating with probability distribution π 0 (note that π0 [0] = 0). According
large cascades will also mitigate blackouts with large amounts to (5), the probability of being in state i at generation k is
of load shed. πk [i] = (π 0 P 0 P 1 · · ·P k−2 P k−1 )[i] = (π 0 P (k−1) )[i]. In par-
A Markov chain driven by real data incorporates all the causes, ticular, the probability that the cascade terminates by generation
mechanisms, and conditions of the cascading that occurred, k is πk [0] = π 0 P (k) [0] = π 0 P (k) e0 . Then for i = 1, . . . , |S|,
but does not distinguish particular causes of the interactions.
However, once the lines critical to large cascades have been πk+1 [i] (π 0 P (k) )[i] (π 0 P (k) )[i]
dk+1 [i] = = =
identified with the influence graph, the causes related to outage 1 − πk+1 [0] (k)
1 − π 0 P e0 π 0 P (k) (1 − e0 )
of those particular lines can be identified by analyzing event
logs and cause codes. Also, the overall impact on cascading of The first row of P k is always [1 0 · · · 0]. Since π0 [0] =
factors such as loading and weather can be studied by dividing 0, let π 0 = [0 π̄ 0 ]. Then π 0 P (k) (1 − e0 ) = π̄ 0 Q(k) 1 and
the data into low and high loading or good and bad weather and (π 0 P (k) )[i] = (π̄ 0 Q(k) )[i] for i = 1, . . . , |S|. And Q(k) =
k−1 k
forming influence graphs for each case. Q̄0 Q̄1+ m=0 (1 − αm ), so that d∞ = limk→∞ dk+1 is
While the Markov model is driven by historical data in this k−1 k
paper, the Markov model is not limited to historical data. The p̄0 Q(k) p̄0 Q̄0 Q̄1+ m=0 (1 − αm )
Markov model could be driven by simulated cascades or a
d∞ = lim
k→∞ p̄0 Q(k) 1
= lim
k→∞ p̄ Q̄ Q̄k−1
k
0 0 1+ (1 − αm )1 m=0
combination of simulated and historical cascades. Moreover, if
k−1 
the probabilities of specific cascading interactions between line p̄0 Q̄0 μ wv
= = v
outages are available, these probabilities could be combined into p̄0 Q̄0 μk−1 wv  1
the entries of the Markov transition matrices. The Markov chain (k−1)
is applied here to cascading transmission line outages, but the where Q̄ → μk−1 wv  as k → ∞. Therefore, the dominant
formulation would apply generally to process real or simulated left eigenvector of Q̄1+ is d∞ .
data for the cascading outage of components within or between For our data, the top three eigenvalues in modulus are μ =
networked infrastructures. 0.502 and −0.136 ± 0.122 i with corresponding moduli 0.502
We show how to estimate the Markov chain from detailed and 0.381.
outage data that is routinely collected by utilities. Being driven
by observed data has some significant advantages of realism. ACKNOWLEDGMENT
In particular, and in contrast with simulation approaches, no
The authors gratefully thank BPA for making the outage data
assumptions about the detailed mechanisms of cascading need to
public. The analysis and any conclusions are strictly the author‘s
made. Since the Markov chain driven by utility data has different
and not BPA’s.
assumptions than simulation, we regard the Markov chain and
simulation approaches as complementary. The Markov chain
driven by observed data offers another way to find critical lines REFERENCES
and to test proposed mitigations of cascading by predicting the [1] P. Hines, J. Apt, and S. Talukdar, “Large blackouts in North America:
effect of the mitigation on the probabilities of small, medium, Historical trends and policy implications,” Energy Policy, vol. 37, no. 12,
pp. 5249–5259, Dec. 2009.
and large cascades. [2] B. A. Carreras, D. E. Newman, and I. Dobson, “North American blackout
time series statistics and implications for blackout risk,” IEEE Trans.
Power Syst., vol. 31, no. 6, pp. 4406–4414, Nov. 2016.
APPENDIX [3] R. Baldick et al., “Initial review of methods for cascading failure analysis in
DERIVING THE QUASI-STATIONARY DISTRIBUTION d∞ electric power transmission systems,” in Proc. IEEE PES General Meeting,
Pittsburgh, PA, USA, Jul. 2008, pp. 1–8.
The quasi-stationary distribution can be derived in a standard [4] M. Papic et al., “Survey of tools for risk assessment of cascading outages,”
in Proc. IEEE PES General Meeting, Detroit, MI, USA, Jul. 2011, pp. 1–9.
way [44], [45]. Let dk be a vector with entry dk [i] which is the [5] J. Bialek et al., “Benchmarking and validation of cascading failure analysis
probability that a cascade is in nonempty state si at generation tools,” IEEE Trans. Power Syst., vol. 31, no. 6, pp. 4887–4900, Nov. 2016.
3234 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 35, NO. 4, JULY 2020

[6] E. Ciapessoni et al., “Benchmarking quasi-steady state cascading outage [30] M. Rahnamay-Naeini, Z. Wang, N. Ghani, A. Mammoli, and M. M. Hayat,
analysis methodologies,” in Proc. IEEE Int. Conf. Probabilistic Methods “Stochastic analysis of cascading-failure dynamics in power grids,” IEEE
Appl. Power Syst., Boise, ID, USA, Jun. 2018, pp. 1–6. Trans. Power Syst., vol. 29, no. 4, pp. 1767–1779, Jul. 2014.
[7] M. Papic and I. Dobson, “Comparing a transmission planning study of [31] M. Rahnamay-Naeini and M. M. Hayat, “Cascading failures in interdepen-
cascading with historical line outage data,” in Proc. IEEE Int. Conf. Prob- dent infrastructures: An interdependent Markov-chain approach,” IEEE
abilistic Methods Appl. Power Syst., Beijing, China, Oct. 2016, pp. 1–7. Trans. Smart Grid, vol. 7, no. 4, pp. 1997–2006, Jul. 2016.
[8] C. Asavathiratham, S. Roy, B. Lesieutre, and G. Verghese, “The influence [32] Z. Wang et al., “Impacts of operators’ behavior on reliability of power
model,” IEEE Control Syst. Mag., vol. 21, no. 6, pp. 52–64, Dec. 2001. grids during cascading failures,” IEEE Trans. Power Syst., vol. 33, no. 6,
[9] M. Rahnamay-Naeini, “Designing cascade-resilient interdependent net- pp. 6013–6024, Nov. 2018.
works by optimum allocation of interdependencies,” in Proc. Int. Conf. [33] U. Nakarmi, M. Rahnamay-Naeini, M. J. Hossain, and M. A. Hasnat,
Comput. Netw. Commun., Kauai, HI, USA, Feb. 2016, pp. 1–7. “Interaction graphs for reliability analysis of power grids: A survey,” Feb.
[10] P. Hines, I. Dobson, E. Cotilla-Sanchez, and M. Eppstein, “‘Dual graph’ 2020, arXiv:1911.00475 [[Link]-ph].
and ‘random chemistry’ methods for cascading failure analysis,” in [34] I. Dobson, “Estimating the propagation and extent of cascading line
Proc. 46th Hawaii Intl. Conf. Syst. Sci., Maui, HI, USA, Jan. 2013, outages from utility data with a branching process,” IEEE Trans. Power
pp. 2141–2150. Syst., vol. 27, no. 4, pp. 2146–2155, Nov. 2012.
[11] J. Qi, K. Sun, and S. Mei, “An interaction model for simulation and [35] Bonneville power administration transmission services operations & relia-
mitigation of cascading failures,” IEEE Trans. Power Syst., vol. 30, no. 2, bility website. [Online]. Available: [Link]
pp. 804–819, Mar. 2015. Operations/Outages
[12] P. Hines, I. Dobson, and P. Rezaei, “Cascading power outages propagate [36] I. Dobson, B. A. Carreras, D. E. Newman, and J. M. Reynolds-Barredo,
locally in an influence graph that is not the actual grid topology,” IEEE “Obtaining statistics of cascading line outages spreading in an electric
Trans. Power Syst., vol. 32, no. 2, pp. 958–967, Mar. 2017. transmission network from standard utility data,” IEEE Trans. Power Syst.,
[13] K. Zhou, I. Dobson, P. Hines, and Z. Wang, “Can an influence graph vol. 31, no. 6, pp. 4831–4841, Nov. 2016.
driven by outage data determine transmission line upgrades that mitigate [37] A. C. Davison and D. V. Hinkley, Bootstrap Methods and Their Applica-
cascading blackouts?” in Proc. IEEE Int. Conf. Probabilistic Methods tion, vol. 1. Cambridge, U.K.: Cambridge Univ. Press, 1997.
Appl. Power Syst., Boise, ID, USA, Jun. 2018, pp. 1–6. [38] M. Vaiman et al., “Risk assessment of cascading outages: Methodologies
[14] J. Qi, J. Wang, and K. Sun, “Efficient estimation of component interactions and challenges,” IEEE Trans. Power Syst., vol. 27, no. 2, pp. 631–641,
for cascading failure analysis by EM algorithm,” IEEE Trans. Power Syst., May 2012.
vol. 33, no. 3, pp. 3153–3161, May 2018. [39] I. Dobson, “Finding a Zipf distribution and cascading propagation metric
[15] X. Zhang et al., “Identification of key transmission lines in power grid in utility line outage data,” 2018, arXiv:1808.08434 [[Link]-ph].
using modified k-core decomposition,” in Proc. 3rd Int. Conf. Electric [40] S. D. Guikema, “Formulating informative, data-based priors for failure
Power Energy Convers. Syst., Istanbul, Turkey, Oct. 2013, pp. 1–6. probability estimation in reliability analysis,” Reliability Eng. Syst. Saf.,
[16] H. M. Merrill and J. W. Feltes, “Cascading blackouts: Stress, vulnerability, vol. 92, no. 4, pp. 490–502, Apr. 2007.
and criticality,” 2016. [Online]. Available: [Link] [41] B. P. Carlin and T. A. Louis, Bayesian Methods for Data Analysis. Boca
[17] Z. Ma, C. Shen, F. Liu, and S. Mei, “Fast screening of vulnerable trans- Raton, FL, USA: CRC Press, 2008.
mission lines in power grids: A pagerank-based approach,” IEEE Trans. [42] E. T. Jaynes, “Bayesian methods: General background, ” in Maximum
Smart Grid, vol. 10, no. 2, pp. 1982–1991, Mar. 2019. Entropy and Bayesian Methods in Applied Statistics. Cambridge, U.K.:
[18] Y. Yang, T. Nishikawa, and A. E. Motter, “Vulnerability and cosuscepti- Cambridge Univ. Press, 1986.
bility determine the size of network cascades,” Phys. Rev. Lett., vol. 118, [43] S.-C. Fang, J. R. Rajasekera, and H.-S. J. Tsao, Entropy Optimization and
no. 4, 2017, Art. no. 048301. Mathematical Programming. Berlin, Germany: Springer, 2012.
[19] B. A. Carreras, D. E. Newman, and I. Dobson, “Determining the vulnera- [44] J. N. Darroch and E. Seneta, “On quasi-stationary distributions in absorb-
bilities of the power transmission system,” in Proc. 45th Hawaii Int. Conf. ing discrete-time finite Markov chains,” J. Appl. Probab., vol. 2, no. 1,
Sys. Sci., Maui, HI, USA, Jan. 2012, pp. 2044–2053. pp. 88–100, Jun. 1965.
[20] U. Nakarmi, M. Rahnamay-Naeini, and H. Khamfroush, “Critical com- [45] E. V. Doorn and P. Pollett, “Quasi-stationary distributions for discrete-state
ponent analysis in cascading failures for power grids using community models,” Eur. J. Oper. Res., vol. 230, pp. 1–14, 2013.
structures in interaction graphs,” IEEE Trans. Netw. Sci. Eng., to be [46] W. J. Stewart, Introduction to the Numerical Solution of Markov Chains.
published. Princeton, NJ, USA: Princeton Univ. Press, 1994.
[21] W. Ju, K. Sun, and J. Qi, “Multi-layer interaction graph for analysis and
mitigation of cascading outages,” IEEE Trans. Emerg. Sel. Topics Circuits
Syst., vol. 7, no. 2, pp. 239–249, Jun. 2017.
[22] C. Chen, W. Ju, K. Sun, and S. Ma, “Mitigation of cascading outages
using a dynamic interaction graph-based optimal power flow model,” IEEE
Access, vol. 7, pp. 168 637–168 648, 2019.
[23] X. Wei, J. Zhao, T. Huang, and E. Bompard, “A novel cascading faults
graph based transmission network vulnerability assessment method,” IEEE
Trans. Power Syst., vol. 33, no. 3, pp. 2995–3000, May 2018.
[24] K. Sun, Y. Hou, W. Sun, and J. Qi, Power System Control Under Cascading
Failures: Understanding, Mitigation, and System Restoration. Hoboken, Kai Zhou (Student Member, IEEE) received the B.S. degree in electrical
NJ, USA: Wiley, 2019. engineering from China Agricultural University, Beijing, China, in 2014, and the
[25] A. Wang, Y. Luo, G. Tu, and P. Liu, “Vulnerability assessment scheme M.S. degree in electrical engineering from Tianjin University, Tianjin, China,
for power system transmission networks based on the fault chain theory,” in 2017. He is currently working toward the Ph.D. degree with Iowa State
IEEE Trans. Power Syst., vol. 26, no. 1, pp. 442–450, Feb. 2011. University, Ames, IA, USA. His current research interests include cascading
[26] C. Luo, J. Yang, Y. Sun, J. Yan, and H. He, “Identify critical branches failures and data analytics.
with cascading failure chain statistics and hypertext-induced topic search
algorithm,” in Proc. IEEE PES General Meeting, Chicago, IL, USA, 2017,
pp. 1–5.
[27] L. Li, H. Wu, and Y. Song, “Temporal difference learning based criti-
cal component identifying method with cascading failure data in power
systems,” in Proc. IEEE PES General Meeting, Portland, OR, USA,
Aug. 2018, pp. 1–5.
[28] L. Li, H. Wu, Y. Song, and Y. Liu, “A state-failure-network method to
identify critical components in power systems,” Electric Power Syst. Res.,
vol. 181, 2020, Art. no. 106192. Ian Dobson (Fellow, IEEE) received the B.A. degree in mathematics from
[29] Z. Wang, A. Scaglione, and R. J. Thomas, “A Markov-transition model for Cambridge University, Cambridge, U.K., and the Ph.D. degree in electrical
cascading failures in power grids,” in Proc. 45th Hawaii Int. Conf. Syst. engineering from Cornell University, Ithaca, NY, USA. He is currently the
Sci., Maui, HI, USA, Jan. 2012, pp. 2115–2124. Sandbulte Professor of Engineering with Iowa State University, Ames, IA, USA.
ZHOU et al.: MARKOVIAN INFLUENCE GRAPH FORMED FROM UTILITY LINE OUTAGE DATA TO MITIGATE LARGE CASCADES 3235

Zhaoyu Wang (Member, IEEE) received the B.S. and M.S. degrees in electrical Alexander Roitershtein received the Ph.D. degree in applied mathematics from
engineering from Shanghai Jiaotong University, Shanghai, China, in 2009 and Technion, Israel Institute of Technology, Haifa, Israel, in 2004. He is a Research
2012, respectively, and the M.S. and Ph.D. degrees in electrical and computer Associate with the Department of Statistics, Texas A&M University, College
engineering from the Georgia Institute of Technology, Atlanta, GA, USA, in Station, TX, USA. His research interests include random walk models and their
2012 and 2015, respectively. He is the Harpole-Pentair Assistant Professor with applications, population dynamics, complex networks, microbiome ecology,
Iowa State University, Ames, IA, USA. His research interests include power applications of branching processes to biology, stochastic processes in random
distribution systems and microgrids, particularly on their data analytics and environment, and general theory of Markov chains.
optimization. He is the Principal Investigator for a multitude of projects focused
on these topics and funded by the National Science Foundation, the Department
of Energy, National Laboratories, PSERC, and Iowa Energy Center. He is the Arka P. Ghosh received the [Link]. and [Link]. degrees from Indian Statistical
Secretary of IEEE Power and Energy Society (PES) Award Subcommittee, a Co- Institute, New Delhi, India and the Ph.D. degree in statistics from the University
Vice Chair of PES Distribution System Operation and Planning Subcommittee, of North Carolina at Chapel Hill, Chapel Hill, NC, USA and joined Iowa State
and a Vice Chair of PES Task Force on Advances in Natural Disaster Mitigation University in 2005. He is a Professor of Statistics with Iowa State University,
Methods. He is an Editor for the IEEE TRANSACTIONS ON POWER SYSTEMS, Ames, IA, USA and holds courtesy appointments in Departments of Mathemat-
IEEE TRANSACTIONS ON SMART GRID, IEEE PES LETTERS, and IEEE OPEN ics as well as in Industrial Manufacturing and Systems Engineering. His research
ACCESS JOURNAL OF POWER AND ENERGY, and an Associate Editor for IET interests are primarily in applied probability and statistics, including stochastic
Smart Grid. modeling, Queueing theory, random graphs, network modeling.