IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL.

69, 2021 2725

Grid-Graph Signal Processing (Grid-GSP): A Graph

Signal Processing Framework for the Power Grid
Raksha Ramakrishna and Anna Scaglione , Fellow, IEEE

Abstract—The underlying theme of this paper is to explore the questions pertaining to sampling, interpolation, denoising and
various facets of power systems data through the lens of graph compression and questions that hinge on the underlying structure
signal processing (GSP), laying down the foundations of the Grid- of the voltage phasors graph signal arise.
GSP framework. Grid-GSP provides an interpretation for the
spatio-temporal properties of voltage phasor measurements, by The overarching goal of this paper is to develop GSP based
showing how the well-known power systems modeling supports models for power systems from first principles by building upon
a generative low-pass graph filter model for the state variables, the existing system-level knowledge of power systems to create
namely the voltage phasors. Using the model we formalize the a solid foundation to analyze power-grid measurements using
empirical observation that voltage phasor measurement data lie in tools from GSP. This is named the Grid-GSP framework. By
a low-dimensional subspace and tie their spatio-temporal structure
to generator voltage dynamics. The Grid-GSP generative model is identifying the correct graph shift operators (GSO), we extend
then successfully employed to investigate the problems, pertaining well-known results in GSP to power system data without losing
to the grid, of data sampling and interpolation, network inference, the associated physical interpretation.
detection of anomalies and data compression. Numerical results The core idea is to rewrite the differential algebraic equations
on a large synthetic grid that mimics the real-grid of the state of (DAE) [7], in a way often done in transient stability analysis of
Texas, ACTIVSg2000, and on real-world measurements from ISO-
New England verify the efficacy of applying Grid-GSP methods to power systems, to reveal that the inherent structure in voltage
electric grid data. phasors can be explained using a linear low-pass graph filter
as a generative model, whose inputs are the generator voltages.
Index Terms—Graph signal processing, phasor measurement
units, false data injection attack, optimal placement of PMU,
This input signal is the generators’ response to electric load
sampling and recovery, PMU data compression, network inference. in the grid. Through this model the paper shows also that the
temporal dynamics of the input signal, i.e. the generator voltages,
I. INTRODUCTION can be explained using a non-linear GSP model defined via
another GSO derived from the generator-only Kron-reduced
HE power grid is one of the foremost examples of a
T large-scale man-made network. The nodes of the associ-
ated graph are the grid buses and its edges are its transmission
network. This is done utilizing the well-known classical swing
equations [8], [9]. This spatio-temporal generative model sup-
ports the empirical observation that voltage data obtained using
lines. It is therefore natural to see measurements from the power
PMUs tend to be confined to a much smaller dimension com-
grid as graph signals [3] and model power grid measurements
pared to the size of the data record in both space and time [10],
using tools from the theory of graph signal processing (GSP)
[11]. Many papers have leveraged the empirical observation
whose goal is to extend fundamental insights that come from
of the low-rank of phasor data for the interpolation of miss-
the frequency analysis for time series to the domain of signals
ing data [11], correcting bad data [12] and to detect faulty
indexed by graphs [3]–[5]. One of the factors that motivate the
events [10], [13]–[15]. Importantly, our framework explicitly
development of GSP for the power grid is the abundance of
puts forth the structure of this low-dimensional subspace using
high-quality data that can be acquired using phasor measurement
our GSP-based generative model, directly tying this subspace to
units (PMU), the sensors producing estimates of the voltage
the graph Fourier domain of the GSO.
and current phasors [6]. With that, classical signal processing
A. Literature Review
Manuscript received May 30, 2020; revised November 7, 2020 and January
31, 2021; accepted April 3, 2021. Date of publication April 23, 2021; date of We review prior works by dividing the most relevant literature
current version May 21, 2021. The associate editor coordinating the review of related to this paper into three categories: 1) a general survey of
this manuscript and approving it for publication was Dr. Soummya Kar. This works that use concepts from graph theory and GSP in power
work was supported in part by the Director, Office of Electricity Delivery and
Energy Reliability, Cybersecurity for Energy Delivery Systems program, of the systems in the areas of sensor placement, interpolation and
U.S. Department of Energy, under Contract DOE0000780. Preliminary work network inference, 2) False Data Injection (FDI) attack detection
was presented in [1] and [2]. (Corresponding author: Raksha Ramakrishna.) and 3) literature pertaining to compression of PMU data.
Raksha Ramakrishna is with the School of Electrical Engineering and Com-
puter Science (EECS), KTH Royal Institute of Technology, Stockholm, Sweden Graph theory for power systems: Several papers have used
(e-mail: [email protected]). insights from spectral and algebraic graph theory. A few appli-
Anna Scaglione is with the School of Electrical Computer and Energy cations include optimal placement [16], [17] and generating sta-
Engineering (ECEE), Arizona State University (ASU), Tempe, AZ 85287-5706
USA (e-mail: [email protected]). tistically accurate topologies [18]. Grid topology identification
Digital Object Identifier 10.1109/TSP.2021.3075145 is a network inference problem and has been studied by several
1053-587X © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
2726 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 69, 2021

works such as in [19]–[23]. GSP concepts have been leveraged B. Contributions

in [24], [25] to detect FDI attacks. Prior work in [26] dealt with
The aim of this paper is to establish the framework of Grid-
performance limits on fault localization with inadequate number
GSP and elucidate properties of power grid signals using tools
of PMUs and connected it with graph signal sampling theory and
from GSP. In particular, we:
optimal placement of PMUs for best possible resolution of fault
1) Establish that PMU voltage measurements from the power
localization in this under-sampled regime.
grid are result of an excitation to a low-pass graph filter
The Kron-reduced network among the generator buses and
whose graph shift operator (GSO) is defined using a
the associated properties are used in [27], [28] to detect low-
function of the system admittance matrix. This was partly
frequency oscillations as well as the resulting islanding patterns.
explored in our previous work in [1], [2] and used for blind
In [29], the authors have shed light on the relationship that exists
community detection.
between graph Laplacian and modes in power systems. Recently,
2) Study the spatio-temporal structure of the excitation that,
a comprehensive review of graph-theoretical concepts in power
at a fast time scale, is dominated by the generators dy-
systems was presented in [7].
namics. It is shown that this excitation can be modeled
Additionally, there have been several papers adopting graphi-
as an auto-regressive graph filter [51] (GF-AR (2)) for
cal models for state estimation [30], topology estimation [31]
the input signals from the generator internal bus. The
and optimal power flow [32]. While the modeling approach
spatial properties in quasi-steady state are captured by
is valid, the graphical models capture correlation whereas our
defining another GSO using the ‘generator-only’ or Kron-
method models the underlying cause for that correlation struc-
reduced [9] network for the generator buses.
ture thereby opening the door for statistical and non-statistical
These models set the foundations to revisit known GSP
approaches.
based algorithms for sampling and reconstruction, interpola-
Note that with the exception of [24], [25], no other papers
tion and denoising and network inference in the context of
make the connection with GSP and even in the aforementioned
signal-processing PMU data. We harness the GFT for feature
work, GSP is used in an empirical manner. On the other hand
extraction, to detect anomalies (specifically, FDI attacks [2])
our preliminary work in [1], [2] established a case for GSP in a
and to derive a lossy PMU voltage data compression algorithm,
more rigorous manner.
leveraging the sparsity of the GFT signal. By elucidating all the
Detection of FDI attacks: While anomaly detection can be
steps in modeling power systems data from graph construction,
broadly applied to identify various events, a large body of
signal model, identification of the low-pass structure that is
prior research has focused on FDI attacks that can bypass
responsible for low-dimensional representation to signal denois-
classical bad data detection (BDD) mechanisms [33]–[36] and
ing, network inference and anomaly detection, we illustrate how
trigger incorrect decisions or hide line overflows and contingen-
one can similarly develop GSP based models in other application
cies [37]–[39]). FDI attacks to PMUs can, for instance, exploit
domains especially data that can be modeled as the output of
the vulnerability of GPS signals to spoofing attacks [40], [41].
low-pass graph signals [52].
Recent papers on PMU data integrity have proposed leverag-
ing the low rank spatio-temporal nature of PMU data not only
to help with erasures but also to strengthen conventional BDD C. Paper Organization
mechanisms [42]. In [38] the authors have suggested an FDI Section II reviews concepts from GSP focusing on complex-
attack strategy that can pass the aforementioned BDD approach valued graph signals and applicable more broadly to bandpass
in [42] by generating false samples that approximately preserve signals whose signal models rely on complex envelopes or
the original subspace structure. phasors. It also reviews measurements and parameters pertaining
Compression of PMU data: Due to their relatively high sam- to the grid. Section III lays the foundation for Grid-GSP mapping
pling rate and their wide deployment, PMU data have have called the physical laws to a spatio-temporal generative model for
for compression. The performance of several off-the-shelf loss- voltage signals. Through these lens, in Section IV, the paper
less encoding techniques applied to PMU data were investigated revisits algorithms and tools from GSP for PMU data pertaining
in [43]. In [44], a lossless compression called slack referenced to sampling and reconstruction along with optimal placement
encoding (SRE) method of PMU voltage measurements is in- of PMUs, interpolation of missing samples and network infer-
troduced, by identifying a slack-bus and differentially encoding ence. Section V highlights applications of Grid-GSP to detect
the difference between slack-bus measurements in time and all FDI attacks and for sequential lossy voltage data compression.
other buses. Similarly, in [45], phasor angle data is encoded The algorithms and methods are tested numerically in Sec-
in a lossless manner by preprocessing data using techniques tion VI. Section VII summarizes conclusions and future research
from [44] and then using Golomb-Rice entropy encoding. directions.
The idea of slow-variation with time in PMU data such as Notation: Boldfaced lowercase letters are used for vectors,
phase angles is used in [46] to firstly transform data into fre- x and uppercase for matrices, A. Transpose is x , A and
quency domain (in time) and then using a ‘reverse water-filling’ conjugate transpose is AH . [x]M is the new vector that has
technique to encode frequency components in the difference elements of the vector indexed by the set M. The operation
measurements. Many lossy compression techniques utilize the {.}, {.} denote the real and imaginary parts of the argument.
low-rank structure inherent in voltage phasor data [47]–[49]. Pseudo-inverse of a matrix is A† . The operation diag(x)
Several other wavelet based compression algorithms exist in the creates a diagonal matrix with elements from a vector x and
literature as well [50]. Diag(A) is the extraction of diagonal values of a matrix A.

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
RAMAKRISHNA AND SCAGLIONE: GRID-GRAPH SIGNAL PROCESSING (GRID-GSP): A GRAPH SIGNAL PROCESSING FRAMEWORK 2727

II. PRELIMINARIES After normalizing the eigenvectors such that ui 1 = 1 ∀i, it

is clear that S1 (ui ) > S1 (uj ) ⇒ |λi | > |λj |. Hence, the as-
A. Graph Signal Processing (GSP) in a Nutshell
cending order of eigenvalues corresponds to increase in fre-
Consider an undirected graph G = (N , E) with nodes i ∈ N quency, |λ0 | = 0 ≤ |λ1 | ≤ |λ2 | . . . |λ|N | |. This ordering is not
and edges (i, j) ∈ E. A graph signal x ∈ C|N | is a vector whose unique since two distinct complex eigenvalues can have the same
ith entry [x]i is associated to node i ∈ N . The set of nodes magnitude.
connected to node i is called the neighborhood of i and denoted The Graph Fourier Transform (GFT) basis is the complex
as Ni . GSP generalizes the notion of discrete time shift for a orthogonal basis U in (4). Hence, the GFT of a graph signal x
time series by introducing the concept of graph shift operator denoted by x̃ and the inverse GFT are given by x̃ = U x and
(GSO): x = Ux̃ respectively where [x̃]m is the frequency component
Definition 1: A graph shift operator (GSO) is a linear neigh- that corresponds to the m-th eigenvalue λm .2 Also, we can define
borhood operator, so that each entry of the shifted graph signal the graph-frequency response of the graph filter, h̃, by writing
is a linear combination of the graph signal neighbors’ values [5].
The linear combinations can use complex-valued weights 
K
H (S) = hk Sk = Udiag(h̃)U , (6)
sij ∈ C, with sij = 0, (i, j) ∈/ E and the GSO can be defined k=0
as matrix multiplication, with a matrix S ∈ C|N |×|N | . Although K
not the only option, one common choice for the GSO, is that of 
[h̃]i = H(λi ), H(λ) := hk λ k (7)
the graph weighted Laplacian,1 i.e: k=0

k∈Ni si,k , i = j hk ∈ C, i = 1, 2, . . . |N |. The frequency response of the filter
[S]i,j = (1)
−si,j , i = j is given by elements in h̃. Subsequently, the input and output of
a graph filter in graph-frequency domain are related as
In this work, we focus on complex symmetric GSOs, S = S
as is applicable to the weighted graph Laplacian for the power v = H(S)x → ṽ = diag(h̃)x̃, (8)
grid. Having defined the notion of shift, one can introduce the
notion of shift-invariance: which is analogous to convolution theorem for time-domain
Definition 2: Given a GSO S a shift invariant operator H signals. Naturally, this leads to the extension of notions such
acting on a graph signal is such that: as low-pass, high-pass and band-pass filters and signals that are
at the heart of sampling and interpolation schemes.
H : x → v, ⇐⇒ H : Sx → Sv. (2)
Linear shift-invariant operators must be matrix polynomials B. GSP for Time Series of Graph Signals
of the GSO S [53]. Therefore, a linear shift-invariant graph filter
is a linear operator and can be defined as: So far, only the nodal index for the graph signal v was consid-
ered. However, one can also encounter graph signal processes

K−1
i.e. temporal variations in a graph signal {v t }t≥0 . Since we
v = H(S)x, H (S) = hk S k (3)
k=0 are interested in the temporal characterization of voltage graph
Additionally, linear shift-invariant graph filters satisfy the con- signals, we revise GSP concepts that are applied to time series of
dition: SH(S) = H(S)S. graph signals [57], [58] in this subsection. Then, we utilize these
Consider the following eigenvalue decomposition of the com- concepts while modeling the temporal dynamics at generator
plex symmetric GSO S, given by Theorem 4.4.13 in [54] for buses in Section III-B.
diagonalizable complex symmetric matrices: In order to characterize graph signal process {v t }t≥0 , a joint
time-vertex domain is considered in the literature by defining
S = UΛU , U U = UU = I. (4) filters whose response is shift invariant with respect to the time
Here Λ is the diagonal matrix with eigenvalues series shift operator z −1 and an appropriately chosen GSO [59].
λ0 , λ1 , . . . λ|N |−1 on the principal diagonal and U are complex To study the same, map the time series of graph signal v t in both
orthogonal eigenvectors. An equivalent concept of frequency the graph frequency (GF) and z−domain by the application of
domain in GSP is defined using eigenvalues and eigenvectors z-transform to the GFT of the graph signal process:
of the GSO. 
+∞
Graph frequencies are the eigenvalues of the GSO and the V (z) = v t z −t , V (z) = UT V (z), (9)
t=0
order of frequencies is based on the total variation (TV) crite-
rion [5], [55] defined using the discrete p Dirichlet form Sp (x)
with p = 1 as in [56] as:
2 It is worth noting that the graph shift operator and Fourier transforms do
S1(x) = Sx 1 ⇒ S1 (ui ) = Sui 1 = |λi | ui 1 (5) not have in general important properties that are found in their conventional
counterparts for time series. One notable fact is that the spectrum of Sx does
not have the same amplitude as the spectrum of x. In fact, the GSO effect is
1 In truth, the Laplacian should be considered a graph differential operator as closer to that of a derivative, since each of the GFT coefficients is rescaled by
opposed to a shift operator, but we use the conventional name nonetheless in the the corresponding frequency. For complex symmetric non-Hermitian operators,
rest of the paper to be consistent with the literature. unfortunately Parseval theorem is also not valid.

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
2728 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 69, 2021

We refer to V (z) as the z-GFT. A graph temporal filter’s [60] 2 × 2 branch admittance matrix that relates current and voltage
impulse response Ht (S) and output v t are injections at the from and to ends of a transmission line [61].
From the branch admittance matrix of the network, applying
K
 t
 Kirchhoff’s law, one can relate the current and voltage phasors
Ht (S) = hk,t Sk , v t = Ht−τ (S)xτ , (10)
for the entire network, introducing a system admittance matrix,
k=0 τ =0
Y ∈ C|N | [61] thus obtaining the network version of Ohm’s law
respectively. Graph filter output v t in the z-domain is: (see (18)). The matrix Y is defined as:
+∞
 
V (z) = H(S⊗z)X(z), where H(S⊗z) := Ht (S)z −t (11) k∈Ni yi,k , i = j
[Y ]i,j = (16)
t=0 −yi,j , i = j
when the input is xt with z-transform X(z) and Ht (S) are where yi,j is the admittance of the branch between buses i and
matrix polynomials of the GSO operator: j if (i, j) ∈ E. The system admittance matrix Y is a complex

K 
K symmetric matrix and it is equivalent to the complex-valued
Ht (S) = hk,t Sk ↔ H(S ⊗ z) = Hk (z)Sk . (12) graph Laplacian matrix associated with the power grid. Next,
k=0 k=0
we will partition the nodes or buses into generator and non-
Here Hk (z) is the z-transform of the filter coefficients hk,t . We generators, so that:
can define also the following impulse response in the GF domain: 

K Y gg Y g
[h̃t ]i = Ht (λi ), Ht (λ) := hk,t λk Y = , (17)
(13) Yg Y 
k=0

and the graph-temporal joint transfer function in the z and GF where Y gg is the generator buses-only network, Y g includes
domain as: the portion connecting generators and loads and Y  corre-
 
+∞ K sponds to the section of the grid connecting the loads buses
[h̃(z)]i = H(λi , z), H(λ, z) = hk,t λk z −t (14) among themselves. The shunt (fixed admittance to ground at a
t=0 k=0
bus) elements at all generator buses are denoted by y gsh ∈ C|NG |
With that, we obtain following input-output relationship: and at all load buses by y sh ∈ C|NL | .
Ṽ (z) = diag(h̃(z))X̃(z), (15) The state of the system, from which all other physical quanti-
ties of interest can be derived, are the voltage phasors at each bus.
by applying GFT to z-domain in (11). In the following we assume that a PMU installed on node/bus
In this work, we focus on a class of graph-temporal filters i ∈ N provides a noisy measurement of voltage and current
called GF-ARMA (q, r) filter [51], [60]. The input-output re- phasors at time t where v(t, i) = |v(t, i)|ejθ(t,i) . With some
lation in both time and z-GFT domain are described below, abuse of notation, we will refer to the PMU data as v(t, i) as
respectively: well. Let the vector of voltage phasors collected at time t be
v t −A1 (S)v t−1 · · ·−Aq (S)v t−q=B0 (S)xt+ · · ·+Br (S)xt−r , v t ∈ C|N | . After v t is partitioned into voltages at generator and
  non-generator buses, let igt ∈ C|NG | be the generator current and
diag(ã(z))Ṽ (z) = diag b̃(z) X̃(z), it ∈ C|NL | the load current. Ohm’s law for a network is:3
  
  y gsh v gt igt
where a (z) = 1 − qt=1 a  t z −t and 
b(z) = rt=0 
bt z −t are the Y + diag v t = i t , where v t = ,i t =
z-transform of the graph frequency responses of the graph filter y sh v t it
taps {At (S)}qt=1 , {Bt (S)}rt=0 for the GF-ARMA (q, r) filter. (18)
Particularly, the GF-AR (2) filter is used in Section III-B to
describe generator temporal dynamics. To describe the operating conditions of the system we introduce
a few more quantities. In power systems transient dynamic
C. Measurements and Parameters of the Electric Grid analysis the impact of generating units is modeled as an internal
bus characterized by a generator impedance (or admittance)
The electric grid network can be represented by an undi- y g ∈ C|NG | for g ∈ NG connected to an ideal voltage source
rected graph G = (N , E) where nodes are buses and its edges called internal voltage [61]; we denote its value at time t by
are its transmission lines. The vertex set is a union between E(t, i) = |E(t, i)|eδ(t,i) , i ∈ NG and the corresponding vector
set of generator, NG and non-generator/ load buses, NL , i ∈ as et ∈ C|NG | so that [et ]i = E(t, i). The current at generator
{NG ∪ NL } = N and the edge set (i, j) ∈ E depicts electrical bus in (18), igt , is obtained as the multiplication of generator
connections. To obtain Ohm’s law for a network of transmission admittance and the difference in voltage at the internal bus and
lines, one starts from the telegrapher equations for a single
line to obtain the ABCD parameters that relate input-output
currents and voltages in the Fourier domain. The equations are 3 Note that the admittances values are frequency responses evaluated at 60 Hz

then rearranged and the so-called π-model is attained, which (for the US) and the voltage and current signals are the corresponding envelopes
at the same frequency; Hence the assumption is the voltage and currents are
is an equivalent circuit containing a series impedance element narrowband and the convolution can be approximated by gain and phase rotation
and parallel susceptance elements. The π-model leads to the equal to the Fourier response at 60 Hz.

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
RAMAKRISHNA AND SCAGLIONE: GRID-GRAPH SIGNAL PROCESSING (GRID-GSP): A GRAPH SIGNAL PROCESSING FRAMEWORK 2729

the generator bus [7] : The Grid-GSP generative model for voltage phasor measure-
ments v t is given by (23). The linear shift-invariant graph filter
igt = diag(y g ) (et − v gt ) (19)
is H(S) = S−1 .
As mentioned in Section. I, the generators respond to electric Remark 1: Shift-invariance of H(S) = S−1 can be directly
load in the grid. In order to model the generators response, a verified from (2) i.e. H(S)S = SH(S) = I. Since H(S) is shift
commonly used approximation is that at the load buses  ∈ NL , invariant, it can be expressed as a matrix polynomial in S
admittances are slowly varying over time [7]. We denote them (Theorem 1 in [53]). Also, H(S) = S−1 can be written as in
as y  (t) ∈ CN . (3) where coefficients hk can be determined by the application
of Cayley-Hamilton theorem for inverse matrices [62].
III. GRAPH SIGNAL PROCESSING FOR THE GRID H(S) is approximately a low-pass graph filter [52] due to
Having described the relevant GSP concepts and introduced the inversion of GSO since the graph frequency response of the
grid quantities and parameters of interest, we are ready to filter can be written from (7) as diag(h̃) = Λ−1 . This implies
introduce the Grid-GSP framework.4 Firstly, we define the GSO that as the graph frequency decreases, the magnitude of the filter
for the grid, then support the definition by introducing the response declines. More importantly, since generic power grids
graph-filter model for voltage phasors, and finally characterize tend to be organized as communities system admittance matrix
the temporal dynamics. All of the above yields a GSP generative Y tends to be sparse [63]. Therefore the GSO S has a high
model for the voltage phasor measurements as a low-pass GSP condition number and the graph frequency response of H(S) is
model, as detailed next. such that it tapers off after a certain λk .
To visualize this more explicitly, consider ΛK to be the
A. Grid Graph Generative Model diagonal matrix with entries λi , i ∈ K = {1, . . . , k}. Define a
low-pass filter with k frequency components and consequently
Grid-GSP for voltage phasors data relies on the following the voltage phasor measurements as
definitions: 
Definition 3: The grid graph shift operator (GSO) is a   λ−1 , i ∈ K
Hk (S)  U diag h̃k U , h̃k = i
complex symmetric matrix equal to a diagonal perturbation of i 0, else
the system admittance matrix with generator admittance values, 
 diag(y g )et
y g + y gsh v t ≈ Hk (S) + ηt , (24)
S  Y + diag (20) it
y sh
From the definition of the GSO it follows that: where Hk (S) will represent the principal subspace of the volt-
Definition 4: The grid Graph Fourier Transform (GFT) age phasors whose dimensionality is the number of graph-
basis for voltage phasors is the orthogonal matrix U given by frequencies |K|. Therefore (24) defines the low-dimensional
the eigenvalue decomposition of the GSO in Definition 3: generative model for quasi-steady state voltage phasor mea-
surements. The error term η t now also captures modeling
S = UΛU , |λmin | > 0 (21) approximation.’
Here, the GSO S is a complex-symmetric matrix To provide insights on the temporal dynamics of the voltage
that has the same support as the electric-grid graph Laplacian phasors, we need to capture the structure of the excitation term.
as Y with the diagonal addition of generator admittances. As a matter of fact, et and it , have different dynamics, as
Note that unlike the graph Laplacian, this GSO is invertible, discussed in the subsequent subsections.
|λmin | > 0. Even when shunt elements y gsh , y sh are ignored as
conventionally done to solve power-flow problems in power B. A GSP Model for Generator Dynamics: et
systems, a diagonal term with the generator admittances y g The excitation term corresponding to generator currents has
that is added to the principal diagonal of Y , makes the GSO S elements as [et ]i coming from each generator i ∈ G. We illus-
invertible (see (20)). trate a non-linear dynamical model for the generators internal
With the GSO S is defined as in (20), one can rewrite (18) voltages, namely et ∈ CNG utilizing a GF-AR(2) graph tempo-
and substitute for igt from (19): ral filter from Section II-B. The model is inspired by the classical
  swing equations [8], [9] that describes the coupled dynamics
y gsh diag(y g ) (et − v gt )
Y +diag v t = (22) of the generators phase angles, δi (t), i ∈ G and the resulting
y sh it variation in frequency, ωi (t)  δ˙i − ω0 where ω0 = 2πf0 with
From now on with slight abuse of notation we denote v t as f0 being the grid frequency (50 or 60 Hz).
voltage phasor measurements that are noisy therefore we add Our model, relies on two steps. First, we model the dynamics
measurement noise ηt which yields the following equation, of a signal xt obtained through the following non-linear trans-
 formation of the internal generator voltages:
diag(y g )et 1
v t = H(S) + ηt (23) xt  (diag(m)) 2 ln(et )
it
1 1
δ t = (diag(m))− 2 {xt }, |e|t = (diag(m))− 2 {xt }
4 Our preliminary GSP modeling effort can be found in [1] (25)

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
2730 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 69, 2021

where the vector m entries are the so-called generators masses, are typically ignored. The swing equation for the generators
δ t are the generators angles that appear in the swing equations angles [64] are a key tool for power systems dynamical analysis:
and |e|t are internal generator voltage magnitudes. Second, like
diag (m) δ̈ + diag (d) δ̇ = w − Y red δ, (32)
in the swing equations, to describe the generators interactions,
we resort to a Kron-reduction [8], [9] of the network, in which where m are the generators masses, introduced previously, d
generators are all adjacent. To define this generator-only network are the damping coefficients of generators (often neglected)
and the corresponding GSO, consider the following admittance and w − Y red δ is the imbalance between the electrical and
matrix, Yall that describes the network topology consisting of mechanical power that triggers the change in generator angular
1
the generator internal buses, generator buses and non-generator velocity and acceleration. Note that {x} = (diag(m)) 2 δ.
buses like done in [9]: We can manipulate (32) to prove the following:
⎡ ⎤ Proposition 1: Let w be such that {w} =
diag (y g +y gsh ) − diag (y g +y gsh ) 0 1
(diag(m))− 2 w. Using the approximation:
Y all = ⎣  ⎦
− diag (y g +y gsh ) 0 Sci diag (d) diag (m)−1 ≈ χI,
 
i.e. the homogeneous simplification as in [64], [65], the dynam-
Sci  S +diag 0 y  
ics of {xt } can be justified with an GF-AR (2) model with
In order to model Y all , it is assumed that the loads are varying GSO Sred :
very slowly in time i.e. y  (t) ≈ y  ∀t. Then, let us denote by {xt } − A1 (Sred ){xt−1 } − A2 (Sred ){xt−2 } ≈ {w},
Sh(A, B) the Schur complement of block B of matrix A. We
compute the Schur complement of block Sci of the matrix Y all A1 (Sred ) := (2−χ)I − Sred , A2 (Sred ) := (χ−1)I (33)
which is nothing but Kron reduction. The Schur complement of Proof: Simple algebra on (32) allows us to recast the equa-
the Sci in Y all has two contributions: tions in the following form:
Sh(Y all , Sci ) = jY red + E red (26) 1
{ẍ} + χ{ẋ} = diag (m)− 2 w − Sred {x}. (34)
where E red is a real diagonal dominated matrix, and the imag-
Assuming that the sampling rate is fast enough, and normal-
inary part Y red has the structure of a graph Laplacian. The
izing it to 1, the finite difference approximations for the deriva-
proposed dynamical model for the graph signal xt relies on the
tives are ẋ ≈ xt − xt−1 , ẍ ≈ xt+1 − 2xt + xt−1 and can be
following definition for the GSO of the Kron-reduced generator-
used to obtain AR (2) GF equations for the samples {xt } in
only graph:
(33). 
Definition 5: A GSO is defined for the Kron-reduced gener-
The model that we introduce is simply extending the GF-
ator only network as
AR (2) model to capture both the real and imaginary part of
1 1
Sred = (diag(m))− 2 Y red (diag(m))− 2 ∈ R|NG | (27) xt i.e. internal generator voltage magnitudes|e|t and angles δ t
respectively and suggesting to search the 2|NG | parameters to
with the following eigenvalue decomposition, fit the model with ã1 , ã2 rather than exploring a general MIMO
Sred = Ured Λred U
red (28) filter response. For simplicity of representation, we write the
dynamical equation for xt in the GF domain,
and the orthonormal GFT basis being Ured .
We introduce the GSP based dynamical model for the x̃t = diag (ã1 ) x̃t−1 + diag (ã2 ) x̃t−2 + w̃t (35)
complex-valued generator internal voltages et via graph tem- such that the impulse response of the filter at graph-frequency
poral filter GF-AR (2) as follows, λred,i is defined by [ã1 ]i , [ã2 ]i . Note from (13) that [ã1 ]i , [ã2 ]i
can be written as polynomials in graph frequency λred,i ,
GSP-based dynamics for generator internal voltages
K
1 −1 K
2 −1
 
1 [ã1 ]i = ak,1 λkred,i , [ã2 ]i = ak,2 λkred,i , (36)
et = exp (diag(m))− 2 xt (29) k=0 k=0

which characterizes the poles of the AR system using graph

where xt is a GF-AR (2) process, i.e. the z-GFT X̃(z)
frequencies. From (33), a first order polynomial may suffice to
satisfies the following:
characterize [ã1 ]i , [ã2 ]i in most cases.
diag (ã(z)) X̃(z) = W̃ (z) (30)
C. Load Dynamics: it
ã(z) = 1 − ã1 z −1 − ã2 z −2 (31)
There are several papers in the literature that deal with load
forecasting and modeling [66]. We adopt a simple AR-2 model
The GSP based dynamical model takes inspiration from swing per node or load bus to describe the dynamics of the load,
equation for generator angles and we empirically choose to
it = diag (b1 ) it−1 + diag (b2 ) it−2 + t (37)
model generator internal voltage magnitude |e|t also using a
GF-AR (2) model although in most power system models, the where parameters b1 , b2 are estimated load data time series. The
dynamics of the amplitudes of the generators internal voltages block diagram in Fig. 1 summarize our modeling efforts.

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
RAMAKRISHNA AND SCAGLIONE: GRID-GRAPH SIGNAL PROCESSING (GRID-GSP): A GRAPH SIGNAL PROCESSING FRAMEWORK 2731

where Sred,M is the Schur complement of the block SMc Mc in

the GSO S i.e.,
Sred,M = Sh(S,SMc Mc) = SMM −SMMc S−1 
Mc Mc SMMc


Lemma 1 translates to an interesting self-similarity/fractional
property for voltage graph signals in that the down-sampled
version v M is still a low-pass graph signal. The self-similarity
is due to SH(S) = I in (39). In summary, for voltage graph
signals,
 
v = S−1 i, v M = S−1 −1
red,M iM −SMMc SMc Mc iMc (42)
In the power grid, this property has been illustrated empirically
in several papers [12], [67] that highlight low-dimensionality of
Fig. 1. Block diagram showing generative model for voltage phasor measure-
ments.
measurements from a subset of buses. Although the reduced-
graph is denser compared to the original graph, it still helps
to infer faults or events that occurred in a subset of nodes
The unique nature of voltage phasor measurements allows us where sensors are not installed as long as correct placement
to describe a similar model for any subset of measurements on strategies are devised i.e. that of choosing the subset M. Work
a graph. This is discussed next. in [26] explored the optimal placement for fault localization in
the under-sampled regime and also made connections with GSP
D. Low-Pass Property of Down-Sampled Voltage theory.
Graph Signal
Let v M (time index t is ignored for simplicity) be the down- IV. REVISITING ALGORITHMS FROM GSP FOR PMU DATA
sampled voltage graph signal where M ⊂ N is the set of node In this section we study some of the implications Grid-
indices at which measurements are available. It can be shown that GSP has while understanding sampling, optimal placement of
any down-sampled graph signal with arbitrary graph frequency measurement devices in power systems, interpolation of miss-
response is low-pass in the reduced-graph frequency domain. It ing samples and network inference. The underlying generative
suggests that one can utilize all the methods for low-pass graph model responsible for low-rank nature of data that has been
signals onto down-sampled versions of the graph signal as well. established in the previous section helps explaining the success
This is summarized in Lemma 1 below. that many past works, such as [12], [68], [69], have attained in
Lemma 1: Let v M be any graph signal down-sampled in the recovering missing PMU data using matrix completion methods.
vertex-domain with |M| samples. Let the GSO defined with The low-pass nature of the voltage graph signals discussed in
respect to the full graph S be invertible. Then, with the GSO Section III provides the theoretical underpinning that support
defined with respect to the reduced-graph of M vertices as the arguments made in the literature.
Sred,M , graph signal v M is the output of low-pass graph filter
H(Sred,M )  S−1red,M A. Sampling and Recovery of Grid-Graph Signals

v M = H (Sred,M ) ϕ (38) From the approximation in (24) we see that voltage graph sig-
nals have graph frequency content that drops as λk grows. This
where the GSO for the reduced-graph is given by Kron-reduction characteristic renders the signal approximately band-limited
of S, Sred,M = Sh(S, SMc Mc ). in the GFT domain [70] which means that there is a cut-off
Proof: Consider a graph signal v with arbitrary graph fre- frequency λk such that frequency content corresponding to λk+1
quency response with respect to GSO S, and higher is negligible. Let the GFT basis corresponding to
the first dominant k graph frequecies be UK . The bandlimiting
v = H (S) x = S−1 (SH (S) x) (39)
operator is, BK = UK U K ∈C
|N |×K
and the low frequency
The GSO S is rewritten in a 2 × 2 block form component of v t is:

SMM SMMc B K v t = UK U
K vt (43)
S= , (40)
S
MMc SM c M c Similarly, a vertex limiting operator (with |M|) vertices is
DM = P M P  M where P M has columns that are coordinate
and S−1 can be written using inverse formula for block matri-
vectors such that each column chooses a vertex/node. When the
ces. When graph signal v is down-sampled, only M rows are
voltage measurements on the electrical network are from a few
considered on both sides of (39). Thus we have,
nodes, i ∈ M at time t, it can be written as [v t ]M = P 
M v t . For
ϕ
   reconstruction, results in [70] dictate the necessary condition be
that |M| ≥ |K|. In the presence of modeling error relative to
v M = S−1
red,M I|M|
−1
−SMM SMc Mc
c (SH (S) x) (41)
the perfect band-limited definition, optimal sampling pattern

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
2732 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 69, 2021

i.e. the best placement for PMUs on the grid to minimize where the two regularizing terms measure the variation in the
the worst-case reconstruction error is closely tied to the grid graph and time domain and cg , ct are the corresponding regu-
topology and the model mismatch relative to a strictly band- larization constants. Importantly, one can use the GSO of the
limited graph signal [71]. An optimal placement strategy of reduced graph, Sred,M if we only have access to a subset of
PMUs that minimizes the worst-case reconstruction error in measurements on the grid, M and employ the same formulation
the presence of model mismatch due to imperfect band-limited as in (46) for interpolation of missing samples.
nature of the voltage graph signal, also known as the E-optimal
design [71], is sought by maximizing the smallest singular value, C. Network Inference as Graph Laplacian Learning
σmin (D M UK ), i.e. choose rows of UK such that they are as
uncorrelated as possible and the resulting matrix has the highest The problem of estimation of GSO S from voltage phasor
condition number [70], [71]. Consider then the spatial sampling measurements can be cast as a solving a problem similar to graph
mask D M = diag(1M ) that selects M locations. Laplacian learning [74] which seeks the GSO that minimizes
1) Sampling: The optimal placement of M PMUs maxi- the total variation of the observed voltage phasors. If current
mizes σmin (D M UK ) which amounts to choosing the rows of measurements it are available, then another regularization term
UK with the smallest possible coherence (as close as possible Sv t − it 22 can be added such that Ohm’s law is satisfied.
to being orthogonal). In [70] and references therein, a greedy Therefore, estimation of GSO can be accomplished by solving
method is employed to find M rows from UK so that the least the following problem:
singular value is maximized. 
T
2
Power systems topologies exhibit naturally a community min Sv t 1 + γ S − diag (Diag (S)) F
S t=1
structure that is reflected in the system admittance matrix Y [1]
due to population density or clusters of loads. It is known that T
 2
a method to determine k communities in a graph is to minimize + Sv t − it 2 (47)
the Ratio Cut [72] and spectral clustering performs a relaxed t=1
Ratio Cut minimization via k−means algorithm on rows of the subject to [S]i,j = [S]j,i , [S]i,j = [S]j,i , i = j, (48)
eigenvectors UK [73]. Thus, choosing rows of UK to be uncor-
related is intuitively putting PMUs in different graph-clusters Tr (S) = α |N | , Tr (S) = β |N | (49)
or communities. This fact was also discussed in [26] in the
context of sensor placements for fault localization. The PMUs Additional constraints on the GSO can be imposed based on the
sampling rate in time exceeds the needs for reconstructions in properties of complex-symmetry (see (48)), sparse off-diagonal
a quasi-steady state conditions by a significant margin and it is entries via the term S − diag(Diag(S)) 2F and dominant
designed to help detect sharp transients in the system. diagonal values (see (49)). Also, S tends to have larger imaginary
2) Reconstruction: Voltage data samples are obtained down- values than real especially on the diagonal. and α, β > 1 control
sampling in space after the optimal placement of PMUs. At time the amplitude of real and imaginary values on the diagonal. As
t when |M| samples, [v t ]M are available, the following model before, the problem above can be recast with down-sampled
applies voltage graph signals to infer the Kron-reduced GSO Sred,M
with the approximation that the term SMMc S−1 Mc Mc iMc in (41)
[v t ]M ≈ P 
M UK ṽ t (44) is treated as additive Gaussian noise. Simulation results for
network inference can be found in Section VI.
where ṽ t is the GFT of graph signal v t Therefore, reconstruction
in spatial domain is done via GFT basis as
 † V. APPLICATIONS OF GRID-GSP
v̂ t = UK P  M UK [v t ]M (45) The goal of this section is to showcase the benefits of casting
problems in the Grid-GSP framework through two exemplary
B. Interpolation of Missing Samples applications, namely anomaly detection and data compression.
When voltage measurements are missing or corrupted, de- The common thread between them is the use of the Grid-GFT
noising and interpolation of such data can be cast as a graph as a tool to extract informative features from PMU data.
signal recovery problem by regularizing the total variation, (TV).
Overall, the problem resembles time-vertex graph signal re- A. Detection of FDI Attacks on PMU Measurements
covery [59]. Let V = [v 1 v 2 . . . v T ] represent the voltage
phasor measurements matrix collected over T time instants. Let This application is based on our preliminary work in [2].
PΩ (V̂) be the set of available measurements that have samples Note that, even though we cast the problem as that of FDI
in entries of set Ω and are noisy, attacks detection, the idea can be easily extended to unveil
sudden changes due to physical events (like fault-currents, or
2

T topology changes) that similarly excite high GF content. We
min PΩ (V̂−V) F +cg Sv t 1
V t=1 assume that we have access to PMU measurements of voltage
and current from the buses they are installed on. Let A be the

T
+ ct v t − v t−1 2
(46) set of available measurements where PMUs are installed and U
2
t=2 be set of unavailable ones. A measurement model can be written

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
RAMAKRISHNA AND SCAGLIONE: GRID-GRAPH SIGNAL PROCESSING (GRID-GSP): A GRAPH SIGNAL PROCESSING FRAMEWORK 2733

using ‘state’ to be the voltage as sparse support recovery problem:

    2
îA Y AA Y AU v A min Π⊥HHk (S) (z−Hδv)2 subject to δv t 1 ≤μ (55)
= +εt (50) δv t
v̂ A I|A| 0 vU
         Constraint on the 1 norm is used to incorporate the prior
zt H v knowledge that the attacker has access to a few measurement
buses, C  N . Note that the performance of the algorithm is
The attacker follows the strategy of FDI attack to manipulate
also dependent on the number of graph-frequency components
both current and voltage on the set of malicious buses, i ∈ C ⊂ A
i.e. k considered.
by introducing a perturbation
B. Compression of PMU Measurements
δv Tt = δv TC 0T|P|+|U | , such that Y PC δv C = 0 (51)
The proposed compression algorithm leverages both (23) and
where P is the set of honest nodes. This requires special con- (35). The measure of distortion we use is the mean-squared error
ditions and placement, since Y PC is tall. Nonetheless, since the (MSE):
system admittance matrix Y is generally sparse [63], Y PC does

T
2
not have full column-rank for a sufficient number of attackers d(v, v̂)  (|N | T )−1 v t − v̂ t 2 (56)
C even when all the measurements are available with A = N . t=0
Our detection problem entails deciding between the hypotheses where T denotes the time instant at which samples are stopped
of attack H1 and no attack H0 . To this end, we can leverage collecting. Since we have a temporal dynamical model for the
the low-dimensional generative model for the voltage graph evolution of voltage in time, we use differential encoding [75]
signal that comes from (24), which imposes additional constraint to quantize the residuals in both generator and load dynamics,
on the perturbation along with that in (51). In short, for the w̃t and t respectively. The voltage at time t is:
attacker to be successful and undetected, she needs to have
knowledge of system parameters and the graph filter with k Algorithm 1: Encoding Algorithm for Compression.
frequency components Hk (S). However, since the attacker does
not have all this knowledge, a typical FDI attack as studied in Input: x̃0 , x̃1 , i0 , i1 , {v t }Tt=2
literature is launched using (51). Using the generative model 1: for t = 2 : T do
2: States x̃ ˆ t , ît from (59) and (60) respectively.
in (24), we know that under normal operating conditions in
quasi-steady state, the received data z t under the no-attack and 3: Voltage estimate:
⎡  ⎤
attack hypotheses H0 , H1 respectively have the structure: − 12 ˆt
diag(y g ) exp diag(m) U x̃
⎧ v̂ 0t = H(S)⎣
red
⎦
⎨H0 : HHk (S) (diag(y g )et ) i  +εt
 
⎪ ît
t
zt =     (57)
⎪H : HH (S)
⎩ 1 k (diag(y g )et ) it +Hδv t +εt
4: Compute GFT of modeling error:
(52) ξ̃ t = U (v t − v̂ 0t )
ˆ ˆ
Therefore, we project z t onto the subspace orthogonal to 5: Quantize: ξ̃ t = Q{ξ̃ t }, v̂ t = v̂ 0t + Uξ̃ t
columnspace of HHk (S) to get a test statistic, d(z). The 6: Update states,
projector is: 
ˆ t ← U
x̃
1
red diag(m) ln(êt ), ît ← [Sv̂ t ]NL
2

Π⊥HHk (S)  I − (HHk (S)) (HHk (S))† (53)

where êt = (diag(y g ))−1 [Sv̂ t ]NG (58)
and under the no attack hypothesis H0 , energy in the orthogonal
7: end for
subspace is less than when there is an attack, H1 . This can be ˆ
Output: {ξ̃ t }Tt=2
converted to the following test,
  2 H1
d(z t )  Π⊥HHk (S) z t 2  τ (54) 
H0
 1  
diag (y g ) exp diag(m)− 2 Ured x̃0t + w̃t
v t = H(S)
where τ is a threshold that can be chosen based on an empirical i,0
t + t
receiver operator characteristics (ROC) curve. Note that, since
where x̃0t = diag (a1 ) x̃t−1 + diag (a2 )x̃t−2 (59)
Hk (S) is a low pass filter, the projector Π⊥HHk (S) in (53) is  
filtering high graph frequencies and the detection measures the i,0
t = diag (b1 ) ît−1 + diag (b2 ) ît−2 (60)
energy on such frequencies as a signature for anomalies.
Thus, v t can be approximated as:
Isolation of compromised buses or estimate of δv can also be  
 1
undertaken with a similar logic. Firstly, using the assumptions diag (y g) exp diag(m)− 2 Ured x̃0t
in the previous section we can solve the following regression v t≈ H(S) +ξ t (61)
problem to recover δv t , formulating a LASSO relaxation of the i,0
t

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
2734 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 69, 2021

Algorithm 2: Decoding Algorithm for Reconstruction.

ˆ
Input: x̃0 , x̃1 , i0 , i1 , {ξ̃ t }Tt=2
1: for t = 2 : T do
2: States x̃ ˆ t , ît from (59) and (60) respectively.
3: Reconstruction of voltage from (57)
ˆ
v̂ t = v̂ 0t + Uξ̃ t
4: Update states x̃ ˆ t , ît from (58)
5: end for
Output: {v̂ t }Tt=2

Note that, the vector ξ t GFT, U ξ t has energy mostly in lower

frequency components and is therefore an appropriate term Fig. 2. Support of the GSO S of the ACTIVSg2000 network.
to quantize using an optimal rate allocation. Specifically, we
allocate bits to each component by setting a desired level of
total distortion, applying the reverse water-filling result [76]
which is optimum for a random vector U ξ t whose entries
are circularly symmetric complex independent Gaussian random
variables and then quantize the components accordingly.5 Then,
we use the quantized vector U ξ t to update the state i.e. to
estimate x̃t and it . Algorithms 1 and 2 describe the encoding
and decoding algorithms respectively. Note that the proposed
scheme of compression is sequential unlike others in literature.
Several corrections can be made as data is collected in time such Fig. 3. Magnitude of Graph Fourier Transform (GFT) for voltage graph signal,
as the update of parameters ã1 , ã2 , b1 , b2 . |v˜t | (left) and input to generator-only network |x˜t | (right) plotted with respect
to normalized graph frequency and shown in log scale.

VI. NUMERICAL RESULTS

The numerical results in this section are mostly obtained using
data from the synthetic ACTIVSg2000 case [77], a realistic
model emulating the ERCOT system, which includes 2,000
buses-with 432 generators and the rest non-generator buses.
The ACTIVSg2000 case data include a realistic PMU data time
series, in which 392 generators are dispatched to meet variable
load demand. The sampling rate, as for real PMUs, is 30 samples
per second. As all the system related parameters are known,
it is easier to verify the proposed modeling strategy through
the ACTIVSg2000 PMU data set. Fig. 2 shows the support
of the graph Laplacian or the Y matrix when ordered into Fig. 4. Magnitude of GFT for spatially down-sampled voltage graph signal,
generator and non-generator buses. The block-diagonal structure |ṽ M | plotted with respect to normalized graph frequency for different placement
strategies that show the effect on the low-pass nature. Steeper attenuation means
is notable, and is the result of the population distribution in the loss of spatial resolution.
state of Texas, which is concentrated in 8 metropolitan areas.
Grid-GSP model: In Fig. 3, magnitude of GFT of voltage
graph signal v t and the input xt are plotted for a single time
instant with respect to their corresponding normalized graph exponent in the input, x̃t = U red xt with the generator GSO
frequencies |λi |/ maxi |λi | and shown in log-scale. From the Sred , is plotted with respect to the graph frequencies in Fig. 3.
linear decay, it is evident that the magnitude of GFT coefficients The decay in GFT coefficients with respect to frequency is less
|ṽ t | corresponding to lower frequencies are more significant pronounced confirming that graph signal xt is not necessarily
as compared to higher frequencies. Similarly, the GFT of the low-pass and in general depends on the topology of the generator
only network.
5 The covariance matrix is not diagonal and ideally one would first whiten the
In Fig. 4, magnitude of GFT of the down-sampled voltage
vector U ξt and then quantize the individual components with bit-allocation graph signal, ṽ M = U red,M v M for |M| = 867 and |M| = 261
akin to reverse water-filling. Since the statistics of U ξt are time-varying, with two different down-sampling strategies: with PMUs placed
one has to perform the whitening transform at each time instant which is at buses in few communities within the GSO S and the other
a cumbersome operation. Therefore we make the assumption of a diagonal
covariance matrix while sacrificing the benefit of modeling the underlying being optimal placement for graph signal reconstruction. The
correlations among the random variables. placement strategy has an effect on the low-pass nature of the

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
RAMAKRISHNA AND SCAGLIONE: GRID-GRAPH SIGNAL PROCESSING (GRID-GSP): A GRAPH SIGNAL PROCESSING FRAMEWORK 2735

Fig. 5. AR model fit to the GFT of xt . Component corresponding to smallest

graph frequency, λred,1 , [x̃t ]1 shown.
Fig. 8. Optimal placement of |M| = 100 PMUs. |K| = 100. The strategy
places PMUs in different graph communities.

Fig. 6. AR model for load bus, j = 1312, i.e. bus with highest absolute value
of load.

Fig. 9. Reconstruction performance after optimal placement [70] of |M|

PMUs. Number of frequencies used: |K| = |M|. For random placement, |K| =
100 used.

temporal response depending on what GFT frequency mode is

excited with Fourier spectra that are more or less concentrated
towards low frequencies depending on the GFT mode. Hence,
temporal dynamics can inform about what is happening in space
(i.e. the trends are coupled).
Revising GSP tools: sampling and optimal placement Fig. 8
Fig. 7. 2−D frequency response (in both graph and time domains) of x̃t . It
shows coupling between temporal and spatial trends.
shows the placement of |M| = 100 PMUs super-imposed on
the support of the ordered Grid-GSO, S when |K| = 100 graph
frequency components are considered. Note the distribution of
down-sampled signal. The steeper attenuation of GFT magni- PMUs to different communities as well as on the generator buses
tude with placement strategy being community-wise is a result as they belong to different graph communities. Fig. 9 exhibits
of loss in spatial-resolution. the performance of the GSP based reconstruction method on
To highlight the temporal variation in the GFT domain of optimally placed |M| PMUs that provide down-sampled mea-
input exponent, x̃t , a short time-series of the real and imaginary surements, v M . The number of graph frequencies considered
parts along with the fit of the AR model are shown in Fig. 5. for reconstruction are |K| = |M|. Even with just 5% of mea-
As expected, the AR model fits well. Fig. 6 shows the similar surements (100 PMUs), the reconstruction error is extremely
AR-2 model fit to the load current at a bus that had the highest low. For random placement, |M| PMUs are chosen at random
absolute value of load. and |K| = 100 graph frequencies are used for reconstruction.
To emphasize the temporal nature of the input, the 2- The trial of random placement is repeated 1,000 times and the
dimensional frequency response (in both graph and time do- most frequently occuring error (estimate of mode of the error
mains) is plotted for the input x̃t in Fig. 7. The figure provides distribution) is plotted. As expected, the reconstruction error for
evidence of the coupling between the graph frequencies and random placement is orders of magnitude higher than optimal
time series Fourier power spectrum, and the variability of the placement.

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
2736 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 69, 2021

Fig. 10. The map of PMUs placed in ISO-NE test case 3 [78] (left) and the support of estimated GSO via (47) (right) shown. Note that the community structure
corresponds to groups of PMUs in the actual system as highlighted in the figure for a few clusters.

Fig. 11. Interpolation of missing measurements for an ISO-NE case [78] using
GSO based regularization. Note the contiguous missing of samples and our
ability to interpolate. The relative noise level used is, (|M|T )σ 2 / V 2F =
10−4 Normalized MSE for this run is 6.22 × 10−4 .

Fig. 12. Comparison of AM-FIHT algorithm in [79] (r = 10, n1 = 3, β = 0)

and the proposed GSO based interpolation with 50% of missing measurements
To illustrate that the proposed modeling holds and algorithms in the ISO-NE dataset [78].
work well also for real PMU data, in the next numerical exper-
iments we used a real-world dataset of measurements from 35
tested the accelerated multichannel fast iterative hard threshold-
PMUs placed in ISO New-England grid (ISO-NE) [78]. The data
ing (AM-FIHT) algorithm proposed in [79], which regularizes
corresponds to a period of 180 seconds when a large generator
the reconstruction task assuming that the Hankel matrix formed
near Ln:2 and Ln:4 introduces oscillations in the system. We
with the columns of V, i.e. H(V), has low rank r,
decimated in time the PMU signals down to sampling frequency
2
1 sample/s. min PΩ (V̂ − V) F subject to rank (H(V)) = r (62)
Network inference: As the underlying GSO is unknown, it is V

estimated via (47) with the goal of recovering the underlying The plot comparing the two methods is shown in Fig. 12. As
reduced-GSO. Since admittance values are not given, we only seen, the GSO based method outperforms the AM-FIHT for
compare the support of the estimated GSO with the community this dataset, indicating that the regularization using the GSO
of PMUs in the network. Fig. 10 shows the support of the esti- is more effective at capturing the low-rank nature of the data,
mated GSO and compares it with the map of PMUs highlighting compared to seeking an arbitrary low rank structure in the the
a few clusters of correspondence. From Fig. 10 we see that Hankel matrix of the data.
the block-diagonal nature of the estimated GSO captures the Detection of FDI attacks: Fig. 13 shows the magnitude of
community structure in the map. the projection of the received measurement z on the orthogonal
Interpolation of missing measurements: Once the GSO is subspace Π⊥HHk (S) . From Fig. 13 it is evident that when there
estimated, we consider the interpolation problem in (46) for the is no attack, the magnitude of the projected component is orders
same ISO-NE dataset. We delete data at random and add noise. of magnitude lower than when the measurements are under the
We solve the problem in (46) to recover missing measurements. FDI attack. This validates the idea of using high GFT frequency
In Fig. 11 we compare the original, corrupted and recovered activity as an indicator of anomalies. Fig. 14 shows the empirical
measurements. Corrupted measurements have missing samples receiver operator characteristics (ROC) curve highlighting the
not just at random but also contiguous in time. The normalized detection performance of the proposed FDI attack detection
MSE, V − V̂ 2F / V 2F is the metric used to gauge the recon- scheme. The detection performance remains good, even when
struction performance. As a comparison, on the same data, we very few buses are attacked. We compare the performance of

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
RAMAKRISHNA AND SCAGLIONE: GRID-GRAPH SIGNAL PROCESSING (GRID-GSP): A GRAPH SIGNAL PROCESSING FRAMEWORK 2737

Fig. 13. Components of projection of received measurement z onto the

orthogonal subspace, Π⊥HHk (S) z, |A| = 500, |K| = 200, |C| = 250.

Fig. 16. Empirical ROC curve for methods proposed here and by Drayer &
Routtenberg [25] when all voltage measurements are available, |A| = 2, 000.
A percent of the measurements, |C|/|A|×100 are malicious. The relative noise
level is 10−2 .

Fig. 14. Empirical ROC curve for different |C| with different percentage
of malicious measurements, |C|/|A| × 100 with |A| = 500 (out of 2,000)
available measurements.

Fig. 17. Empirical rate distortion (RD) curve for the proposed compression
method compared with singular value thresholding and quantization.

to noise and also significantly more sensitive in detecting the

attack vectors, even when only few buses are attacked.
Fig. 15 shows the reconstruction of magnitude of the attack
vector δv when 500 measurements are available and number of
attacked buses |C| = 50.
Compression based results: For voltage data compression,
we compared with two schemes: scalar quantization and sin-
gular value thresholding (SVT) from [49]. Fig. 17 plots the
Fig. 15. Reconstruction of attack vector. empirical rate-distortion (RD) curve and shows the comparison
between all 3 schemes. As expected, scalar quantization does
the proposed FDI attack detection with that of the method poorly compared to the other schemes. The SVT scheme simply
in [25] when the full state i.e. when all voltage measurements uses few of the largest singular vectors for data reconstruction.
are available, |A| = 2, 000. The underlying principle to detect Considering that it is indicative of voltage graph signal lying
the attack in [25] is to look at the magnitude of graph fre- in a low-dimensional subspace, it is not surprising that the SVT
quency components at higher graph frequencies which is similar scheme does well. However, the SVT curve rate-distortion curve
in principle to the detection test we undertake. They use the eventually saturates. Note that the performance of the proposed
real and imaginary parts of the system admittance matrix as method are comparable to those of the SVT. However, the latter
2 GSOs, {Y } and {Y } respectively. Their test statistic is is a batch method, while the proposed method is sequential,
comprised of four components that are the frequency response which has important implication for the online communications
of high-pass filtered real and imaginary voltage measurements of PMU data.
(see Algorithm.2 in [25]). Fig. 16 shows the empirical ROC
curves that compare the performance of the proposed method
and the one in [25] when all voltage measurements are available VII. CONCLUSION
and are noisy. The relative noise level used is 10−2 . As evident In this paper, we proposed the framework of Grid-GSP for the
from the curves, the proposed method performs better than the power grid that highlights the inherent spatio-temporal struc-
method in [25]. This is because our test statistic is more robust ture in the voltage phasors by employing concepts from GSP.

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
2738 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 69, 2021

Grid-GSP revisits the concepts of sampling and reconstruction, [18] Z. Wang, A. Scaglione, and R. Thomas, “Generating statistically correct
interpolation, network inference and applications, to detection random topologies for testing smart grid communication and control
networks,” IEEE Trans. Power Syst., vol. 1, no. 1, pp. 28–39, Jun. 2010.
of FDI attacks and a lossy sequential data compression, were [19] X. Li and H. Poor Vincent and A. Scaglione, “Blind topology identification
introduced using the lens of GSP. The resulting algorithms were for power systems,” in Proc. IEEE Int. Conf. Smart Grid Commun., 2013,
tested on data from both synthetic and real-world datasets. The pp. 91–96.
[20] S. Grotas, Y. Yakoby, I. Gera, and T. Routtenberg, “Power systems topol-
paper opens the door to leverage the GSP foundations for all ogy and state estimation by graph blind source separation,” IEEE Trans.
types of grid data analytical tasks. Signal Process., vol. 67, no. 8, pp. 2036–2051, Apr. 2019.
[21] Z. Xiang, K. Huang, W. Deng, and C. Yang, “Blind topology identification
for smart grid based on dictionary learning,” in Proc. IEEE Symp. Ser.
ACKNOWLEDGMENT Comput. Intell., 2019, pp. 1319–1326.
[22] D. Deka, M. Chertkov, and S. Backhaus, “Joint estimation of topology and
The authors would like to thank the anonymous reviewers and injection statistics in distribution grids with missing nodes,” IEEE Control
Netw. Syst., vol. 7, no. 3, pp. 1391–1403, Sep. 2020.
the editor for their comments to improve the quality of the paper. [23] S. Talukdar, D. Deka, H. Doddi, D. Materassi, M. Chertkov, and
The views expressed in the material are those of the authors and M. V. Salapaka, “Physics informed topology learning in networks of linear
do not necessarily reflect those of the sponsors. dynamical systems,” Automatica, vol. 112, 2020, Art. no. 108705.
[24] E. Drayer and T. Routtenberg, “Detection of false data injection attacks in
power systems with graph fourier transform,” in Proc. IEEE Global Conf.
Signal Inf. Process., 2018, pp. 890–894.
REFERENCES [25] E. Drayer and T. Routtenberg, “Detection of false data injection attacks in
[1] R. Ramakrishna and A. Scaglione, “On modeling voltage phasor measure- smart grids based on graph signal processing,” IEEE Syst. J., vol. 14, no.
ments as graph signals,” in Proc. IEEE Data Sci. Workshop, Jun. 2019, 2, pp. 1886–1896, Aug. 2019.
pp. 275–279. [26] M. Jamei et al., “Phasor measurement units optimal placement and perfor-
[2] R. Ramakrishna and A. Scaglione, “Detection of false data injection attack mance limits for fault localization,” IEEE J. Sel. Areas Commun., vol. 38,
using graph signal processing for the power grid,” in Proc. IEEE Global no. 1, pp. 180–192, Jan. 2020.
Conf. Signal Inf. Process., 2019, pp. 1–5. [27] L. Fan, “Interarea oscillations revisited,” IEEE Trans. Power Syst., vol. 32,
[3] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst, no. 2, pp. 1585–1586, Mar. 2017.
“The emerging field of signal processing on graphs: Extending high- [28] T. Huang, N. M. Freris, P. R. Kumar, and L. Xie, “Localization of forced
dimensional data analysis to networks and other irregular domains,” IEEE oscillations in the power grid under resonance conditions,” in Proc. 52nd
Signal Process. Mag., vol. 30, no. 3, pp. 83–98, May 2013. Annu. Conf. Inf. Sci. Syst., 2018, pp. 1–5.
[4] N. Tremblay, P. Gonçalves, and P. Borgnat, “Design of graph filters [29] L. Guo, C. Zhao, and S. H. Low, “Graph laplacian spectrum and primary
and filterbanks,” in Cooperative Graph Signal Process., Elsevier, 2018, frequency regulation,” in Proc. IEEE Conf. Decis. Control, 2018, pp. 158–
pp. 299–324. 165.
[5] A. Sandryhaila and J. M. F. Moura, “Discrete signal processing on [30] Y. Weng, R. Negi, and M. D. Ilić, “Graphical model for state estimation in
graphs: Frequency analysis,” IEEE Trans. Signal Process., vol. 62, no. 12, electric power systems,” in Proc. IEEE Int. Conf. Smart Grid Commun.,
pp. 3042–3054, Jun. 2014. 2013, pp. 103–108.
[6] A. Phadke and J. Thorp, “History and applications of phasor measure- [31] D. Deka, M. Chertkov, and S. Backhaus, “Topology estimation using
ments,” in Proc. IEEE PES Power Syst. Conf. Expo., 2006, pp. 331–335. graphical models in multi-phase power distribution grids,” IEEE Trans.
[7] T. Ishizaki, A. Chakrabortty, and J.-I. Imura, “Graph-theoretic analysis of Power Syst., vol. 35, no. 3, pp. 1663–1673, May 2020.
power systems,” Proc. IEEE, vol. 106, no. 5, pp. 931–952, May 2018. [32] K. Dvijotham, M. Chertkov, P. Van Hentenryck, M. Vuffray, and S. Misra,
[8] F. Dörfler and F. Bullo, “Kron reduction of graphs with applications to “Graphical models for optimal power flow,” Constraints, vol. 22, no. 1,
electrical networks,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 60, pp. 24–49, 2017.
no. 1, pp. 150–163, Jan. 2013. [33] R. B. Bobba, K. M. Rogers, Q. Wang, H. Khurana, K. Nahrstedt, and
[9] F. Dorfler and F. Bullo, “Spectral analysis of synchronization in a lossless T. J. Overbye, “Detecting false data injection attacks on DC state estima-
structure-preserving power network model,” in Proc. 1st IEEE Int. Conf. tion,” in Preprints First Workshop Secure Control Syst., CPSWEEK, vol.
Smart Grid Commun., 2010, pp. 179–184. 2010, 2010.
[10] L. Xie, Y. Chen, and P. R. Kumar, “Dimensionality reduction of syn- [34] G. Dan and H. Sandberg, “Stealth attacks and protection schemes for state
chrophasor data for early event detection: Linearized analysis,” IEEE estimators in power systems,” in Proc. 1st IEEE Int. Conf. Smart Grid
Trans. Power Syst., vol. 29, no. 6, pp. 2784–2794, Nov. 2014. Commun., 2010, pp. 214–219.
[11] P. Gao, M. Wang, S. G. Ghiocel, J. H. Chow, B. Fardanesh, and [35] O. Kosut, L. Jia, R. J. Thomas, and L. Tong, “Malicious data attacks on
G. Stefopoulos, “Missing data recovery by exploiting low-dimensionality the smart grid,” IEEE Trans. Smart Grid, vol. 2, no. 4, pp. 645–658, Dec.
in power system synchrophasor measurements,” IEEE Trans. Power Syst., 2011.
vol. 31, no. 2, pp. 1006–1013, Mar. 2016. [36] G. Liang, J. Zhao, F. Luo, S. R. Weller, and Z. Y. Dong, “A review of false
[12] M. Wang, “Data quality management of synchrophasor data in power data injection attacks against modern power systems,” IEEE Trans. Smart
systems by exploiting low-dimensional models,” in Proc. 51st Annu. Conf. Grid, vol. 8, no. 4, pp. 1630–1638, Jul. 2017.
Inf. Sci. Syst., 2017, pp. 1–2. [37] I. Esnaola, S. M. Perlaza, H. V. Poor, and O. Kosut, “Maximum distor-
[13] M. Jamei et al., “Anomaly detection using optimally-placed PMU sen- tion attacks in electricity grids,” IEEE Trans. Smart Grid, vol. 7, no. 4,
sors in distribution grids,” IEEE Trans. Power Syst., vol. 33, no. 4, pp. 2007–2015, Jul. 2016.
pp. 3611–3622, Jul. 2018. [38] J. Zhang, Z. Chu, L. Sankar, and O. Kosut, “False data injection attacks
[14] W. Li, M. Wang, and J. H. Chow, “Real-time event identification through on phasor measurements that bypass low-rank decomposition,” in Proc.
low-dimensional subspace characterization of high-dimensional syn- IEEE Int. Conf. Smart Grid Commun., Oct. 2017, pp. 96–101.
chrophasor data,” IEEE Trans. Power Syst., vol. 33, no. 5, pp. 4937–4947, [39] Y. He, G. J. Mendis, and J. Wei, “Real-time detection of false data injection
Sep. 2018. attacks in smart grid: A deep learning-based intelligent mechanism,” IEEE
[15] J. Kim, L. Tong, and R. J. Thomas, “Subspace methods for data attack on Trans. Smart Grid, vol. 8, no. 5, pp. 2505–2516, Sep. 2017.
state estimation: A data driven approach,” IEEE Trans. Signal Process., [40] D. P. Shepard, T. E. Humphreys, and A. A. Fansler, “Evaluation of
vol. 63, no. 5, pp. 1102–1114, Mar. 2015. the vulnerability of phasor measurement units to gps spoofing attacks,”
[16] P. Du, Z. Huang, R. Diao, B. Lee, and K. K. Anderson, “PMU placement Int. J. Crit. Infrastructure Protection, vol. 5, no. 3-4, pp. 146–153,
for enhancing dynamic observability of a power grid,” in Proc. IEEE Conf. 2012.
Innov. Technol. Efficient Reliable Electricity Supply, Sep. 2010, pp. 15–21. [41] L. Heng, J. J. Makela, A. D. Dominguez-Garcia, R. B. Bobba, W. H.
[17] A. Pal, G. A. Sanchez-Ayala, V. A. Centeno, and J. S. Thorp, “A PMU Sanders, and G. X. Gao, “Reliable GPS-based timing for power systems:
placement scheme ensuring real-time monitoring of critical buses of the A multi-layered multi-receiver architecture,” in Proc. Power Energy Conf.
network,” IEEE Trans. Power Del., vol. 29, no. 2, pp. 510–517, Apr. 2014. Illinois, 2014, pp. 1–7.

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.
RAMAKRISHNA AND SCAGLIONE: GRID-GRAPH SIGNAL PROCESSING (GRID-GSP): A GRAPH SIGNAL PROCESSING FRAMEWORK 2739

[42] P. Gao et al., “Identification of successive “unobservable” cyber data [68] N. Dahal, R. L. King, and V. Madani, “Online dimension reduction of
attacks in power systems through matrix decomposition,” IEEE Trans. synchrophasor data,” in Proc. IEEE PES Transmiss. Distrib. Conf. Expo.,
Signal Process., vol. 64, no. 21, pp. 5557–5570, Nov. 2016. 2012, pp. 1–7.
[43] P. Top and J. Breneman, “Compressing phasor measurement data,” in Proc. [69] M. Liao, D. Shi, Z. Yu, Z. Yi, Z. Wang, and Y. Xiang, “An alternating
IEEE Power Energy Soc. Gen. Meeting, Jul. 2013, pp. 1–4. direction method of multipliers based approach for PMU data recovery,”
[44] R. Klump, P. Agarwal, J. E. Tate, and H. Khurana, “Lossless compression IEEE Trans. Smart Grid, vol. 10, no. 4, pp. 4554–4565, Jul. 2019.
of synchronized phasor measurements,” in Proc. IEEE PES Gen. Meeting, [70] M. Tsitsvero, S. Barbarossa, and P. Di Lorenzo, “Signals on graphs:
Jul. 2010, pp. 1–7. Uncertainty principle and sampling,” IEEE Trans. Signal Process., vol. 64,
[45] J. E. Tate, “Preprocessing and golomb-rice encoding for lossless com- no. 18, pp. 4845–4860, Sep. 2016.
pression of phasor angle data,” IEEE Trans. Smart Grid, vol. 7, no. 2, [71] A. Anis, A. Gadde, and A. Ortega, “Efficient sampling set selection for
pp. 718–729, Mar. 2016. bandlimited graph signals using graph spectral proxies,” IEEE Trans.
[46] S. Kirti, Z. Wang, A. Scaglione, and R. Thomas, “On the communication Signal Process., vol. 64, no. 14, pp. 3775–3789, Jul. 2016
architecture for wide-area real-time monitoring in power networks,” in [72] L. Hagen and A. B. Kahng, “New spectral methods for ratio cut partitioning
Proc. 40th Annu. Hawaii Int. Conf. Syst. Sci., 2007, pp. 119–119. and clustering,” IEEE Trans. Comput.-Aided Design Integr. Circuits Syst.,
[47] P. H. Gadde, M. Biswal, S. Brahma, and H. Cao, “Efficient compres- vol. 11, no. 9, pp. 1074–1085, Sep. 1992.
sion of PMU data in WAMS,” IEEE Trans. Smart Grid, vol. 7, no. 5, [73] U. V. Luxburg, “A tutorial on spectral clustering,” Statist. Comput., vol. 17,
pp. 2406–2413, Sep. 2016. no. 4, pp. 395–416, 2007.
[48] Y. Ge, A. J. Flueck, D. K. Kim, J. B. Ahn, J. D. Lee, and D. Y. Kwon, [74] X. Dong, D. Thanou, M. Rabbat, and P. Frossard, “Learning graphs from
“Power system real-time event detection and associated data archival data: A signal representation perspective,” IEEE Signal Process. Mag.,
reduction based on synchrophasors,” IEEE Trans. Smart Grid, vol. 6, no. 4, vol. 36, no. 3, pp. 44–63, May 2019.
pp. 2088–2097, Jul. 2015. [75] W. Weber, “Differential encoding for multiple amplitude and phase shift
[49] J. C. S. de Souza, T. M. L. Assis, and B. C. Pal, “Data compression in smart keying systems,” IEEE Trans. Commun., vol. 26, no. 3, pp. 385–391, Mar.
distribution systems via singular value decomposition,” IEEE Trans. Smart 1978.
Grid, vol. 8, no. 1, pp. 275–284, Jan. 2017. [76] T. M. Cover and J. Thomas, Elements of Information Theory. Hoboken,
[50] R. Mehra, V. Patel, F. Kazi, N. M. Singh, and S. R. Wagh, “Modes NJ, USA: Wiley, 1991.
preserving wavelet based multi-scale PCA algorithm for compression of [77] A. B. Birchfield, T. Xu, K. M. Gegner, K. S. Shetye, and T. J. Overbye,
smart grid data,” in Proc. Int. Conf. Adv. Comput., Commun. Informat., “Grid structural characteristics as validation criteria for synthetic net-
Aug 2013, pp. 817–821. works,” IEEE Trans. Power Syst., vol. 32, no. 4, pp. 3258–3265, Jul. 2017.
[51] E. Isufi, A. Loukas, A. Simonetto, and G. Leus, “Autoregressive moving [78] S. Maslennikov et al., “A test cases library for methods locating the sources
average graph filtering,” IEEE Trans. Signal Process., vol. 65, no. 2, of sustained oscillations,” in Proc. IEEE PES Gen. Meeting, Boston, MA,
pp. 274–288, Jan. 2017. 2016, pp. 1–5.
[52] R. Ramakrishna, H. T. Wai, and A. Scaglione, “A user guide to low-pass [79] S. Zhang, Y. Hao, M. Wang, and J. H. Chow, “Multi-channel hankel matrix
graph signal processing and its applications: Tools and applications,” IEEE completion through nonconvex optimization,” IEEE J. Sel. Topics Signal
Signal Process. Mag., vol. 37, no. 6, pp. 74–85, Nov. 2020. Process., vol. 12, no. 4, pp. 617–632, Aug. 2018.
[53] A. Sandryhaila and J. M. Moura, “Discrete signal processing on graphs,”
IEEE Trans. Signal Process., vol. 61, no. 7, pp. 1644–1656, Apr. 2013.
[54] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge, U.K.: Cam-
bridge Univ. Press, 1990. Raksha Ramakrishna received the B.E degree in
[55] S. Mallat, A Wavelet Tour of Signal Processing: The Sparse Way. New electronics and communications engineering from
York, NY, USA: Academic Press, 2009. the Rashtreeya Vidyalaya College of Engineering,
[56] R. Singh, A. Chakraborty, and B. Manoj, “Graph Fourier transform based Bangalore, India, in 2014, and the M.S. and Ph.D.
on direcetd laplacian,” in Proc. Int. Conf. Signal Process. Commun., 2016, degrees in electrical engineering from Arizona State
pp. 1–5. University, in 2017 and 2020, respectively. She is
[57] E. Isufi, G. Leus, and P. Banelli, “2-dimensional finite impulse response currently a Postdoctoral Researcher with the Division
graph-temporal filters,” in Proc. IEEE Global Conf. Signal Inf. Process., of Network and System Technology, KTH Royal
2016, pp. 405–409. Institute of Technology, Stockholm, Sweden. Her
[58] E. Isufi, A. Loukas, A. Simonetto, and G. Leus, “Filtering random graph research interests include the domain of statistical
processes over random time-varying graphs,” IEEE Trans. Signal Process., signal processing, data analytics for power systems,
vol. 65, no. 16, pp. 4406–4421, Aug. 2017. and recently in security and privacy in federated machine learning systems.
[59] F. Grassi, A. Loukas, N. Perraudin, and B. Ricaud, “A time-vertex signal
processing framework: Scalable processing and meaningful representa-
tions for time-series on graphs,” IEEE Trans. Signal Process., vol. 66, Anna Scaglione (Fellow, IEEE) received the M.Sc.
no. 3, pp. 817–829, Feb. 2018. degree in 1995 and the Ph.D. degree in 1999, She is
[60] E. Isufi, A. Loukas, A. Simonetto, and G. Leus, “Separable autoregressive currently a Professor with the School of Electrical
moving average graph-temporal filters,” in Proc. 24th Eur. Signal Process. Computer and Energy Engineering, Arizona State
Conf., 2016, pp. 200–204. University, Tempe, AZ, USA. She was a Professor of
[61] J. D. Glover, M. S. Sarma, and T. J. Overbye, Power System Analysis and electrical engineering with University of California,
Design. Stamford, CT, US: Cengage Learning, 2008. Davis, USA, during 2010–2014, an Associate Pro-
[62] H. P. Decell, Jr, “An application of the cayley-hamilton theorem to gener- fessor with University College Davis, Ireland, during
alized matrix inversion,” SIAM Rev., vol. 7, no. 4, pp. 526–528, 1965. 2008–2010 and with Cornell University, Ithaca, NY,
[63] N. Sato and W. F. Tinney, “Techniques for exploiting the sparsity or the USA, during 2006–2008, and Assistant Professor
network admittance matrix,” IEEE Trans. Power App. Syst., vol. 82, no. 69, with Cornell University during 2001–2006 and with
pp. 944–950, Dec. 1963. The University of New Mexico, Albuquerque, NM, USA, during 2000–2001.
[64] P. Huynh, H. Zhu, Q. Chen, and A. E. Elbanna, “Data-driven estimation Her expertise is in the broad area of statistical signal processing for communi-
of frequency response from ambient synchrophasor measurements,” IEEE cation, electric power systems and networks. Her research focuses on various
Trans. Power Syst., vol. 33, no. 6, pp. 6590–6599, Nov.2018. applications of signal processing in network and data science, which include
[65] F. Paganini and E. Mallada, “Global analysis of synchronization perfor- intelligent infrastructure for energy delivery and information systems. She was
mance for power systems: Bridging the theory-practice gap,” IEEE Trans. elected an IEEE fellow in 2011. During 2019–2020, she was a SPS Distinguished
Autom. Control, vol. 65, no. 7, pp. 3007–3022, Jul. 2020. Lecturer. She was the recipient of the 2000 IEEE Signal Processing Transactions
[66] F. Gao, J. S. Thorp, A. Pal, and S. Gao, “Dynamic state prediction based Best Paper Award and more recently was honored for the 2013, IEEE Donald
on auto-regressive (AR) model using PMU data,” in Proc. IEEE Power G.Fink Prize Paper Award for the Best Review Paper in that year in the
Energy Conf. Illinois, 2012, pp. 1–5. IEEE publications, the 2013 IEEE Signal Processing Society Young Author
[67] M. Wang et al., “A low-rank matrix approach for the analysis of large Best Paper Award (Lin Li) and several best conference paper awards, and the
amounts of power system synchrophasor data,” in Proc. 48th Hawaii Int. 2020 Technical Achievement Award from the IEEE Communication Society
Conf. System Sci., Jan 2015, pp. 2637–2644. Technical Committee on Smart Grid Communications.

Authorized licensed use limited to: East Carolina University. Downloaded on June 18,2021 at 00:45:40 UTC from IEEE Xplore. Restrictions apply.