1

Grid-Graph Signal Processing (Grid-GSP):

A Graph Signal Processing Framework
for the Power Grid
Raksha Ramakrishna and Anna Scaglione

Abstract—The underlying theme of this paper is to explore the extend well-known results in GSP to power system data
various facets of power systems data through the lens of graph without losing the associated physical interpretation.
signal processing (GSP), laying down the foundations of the Grid- The core idea is to rewrite the differential algebraic equa-
arXiv:2103.06068v1 [eess.SP] 10 Mar 2021

GSP framework. Grid-GSP provides an interpretation for the

spatio-temporal properties of voltage phasor measurements, by tions (DAE) [7], in a way often done in transient stability
showing how the well-known power systems modeling supports analysis of power systems, to reveal that the inherent structure
a generative low-pass graph filter model for the state variables, in voltage phasors can be explained using a linear low-
namely the voltage phasors. Using the model we formalize the em- pass graph filter as a generative model, whose inputs are
pirical observation that voltage phasor measurement data lie in a the generator voltages. This input signal is the generators’
low-dimensional subspace and tie their spatio-temporal structure
to generator voltage dynamics. The Grid-GSP generative model is response to electric load in the grid. Through this model the
then successfully employed to investigate the problems pertaining paper shows also that the temporal dynamics of the input
to the grid of data sampling and interpolation, network inference, signal, i.e. the generator voltages, can be explained using
detection of anomalies and data compression. Numerical results a non-linear GSP model defined via another GSO derived
on a large synthetic grid that mimics the real-grid of the state from the generator-only Kron-reduced network. This is done
of Texas, ACTIVSg2000, and on real-world measurements from
ISO-New England verify the efficacy of applying Grid-GSP utilizing the well-known classical swing equations [8], [9].
methods to electric grid data. This spatio-temporal generative model supports the empirical
observation that voltage data obtained using PMUs tend to
be confined to a much smaller dimension compared to the
I. I NTRODUCTION size of the data record in both space and time [10], [11].
Many papers have leveraged the empirical observation of the
The power grid is one of the foremost examples of a large- low-rank of phasor data for the interpolation of missing data
scale man-made network. The nodes of the associated graph [11], correcting bad data [12] and to detect faulty events [10],
are the grid buses and its edges are its transmission lines. [13]–[15]. Importantly, our framework explicitly puts forth the
It is therefore natural to see measurements from the power structure of this low-dimensional subspace using our GSP-
grid as graph signals [3] and model power grid measurements based generative model, directly tying this subspace to the
using tools from the theory of graph signal processing (GSP) graph Fourier domain of the GSO.
whose goal is to extend fundamental insights that come
from the frequency analysis for time series to the domain
of signals indexed by graphs [3]–[5]. One of the factors that A. Literature review
motivate the development of GSP for the power grid is the We review prior works by dividing the most relevant liter-
abundance of high-quality data that can be acquired using ature into three categories: 1) a general survey of works that
phasor measurement units (PMU)s, the sensors producing use concepts from graph theory and GSP in power systems
estimates of the voltage and current phasors [6]. With that, in the areas of sensor placement, interpolation and network
classical signal processing questions pertaining to sampling, inference, 2) FDI attack detection and 3) literature pertaining
interpolation, denoising and compression and questions that to data compression for PMU data.
hinge on the underlying structure of the voltage phasors graph Graph theory for power systems: Several papers have used
signal arise. insights from spectral and algebraic graph theory. A few
The overarching goal of this paper is to develop GSP based applications include optimal placement [16], [17] and gen-
models for power systems from first principles by building erating statistically accurate topologies [18]. Grid topology
upon the existing system-level knowledge of power systems to identification is a network inference problem and has been
create a solid foundation to analyze power-grid measurements studied by several works such as in [19]–[23]. GSP concepts
using tools from GSP. We call this the Grid-GSP framework. have been leveraged in [24], [25] to detect FDI attacks. Prior
By identifying the correct graph shift operators (GSO), we work in [26] dealt with performance limits on fault localization
with inadequate number of PMUs and connected it with graph
Work done at the School of Electrical, Computer and Energy Engineering
(ECEE), Arizona State University, Tempe, AZ. This research was supported signal sampling theory and optimal placement of PMUs for
in part by the Director, Office of Electricity Delivery and Energy Reliability, best possible resolution of fault localization in this under-
Cybersecurity for Energy Delivery Systems program, of the U.S. Department sampled regime.
of Energy, under contract DOE0000780. Any opinions, and findings expressed
in this material are those of the authors and do not necessarily reflect those The Kron-reduced network among the generator buses and
of the sponsors Preliminary work was presented in [1], [2]. the associated properties are used in [27], [28] to detect
2

low-frequency oscillations as well as the resulting islanding B. Contributions

patterns. In [29], the authors have shed light on the relationship The aim of this paper is to establish the framework of Grid-
that exists between graph Laplacian and modes in power sys- GSP and elucidate properties of power grid signals using tools
tems. Recently, a comprehensive review of graph-theoretical from GSP. In particular, we:
concepts in power systems is presented in [7]. 1) Establish that PMU voltage measurements from the
Additionally, there have been several papers adopting graph- power grid are result of an excitation to a low-pass graph
ical models for state estimation [30], topology estimation [31] filter whose graph shift operator (GSO) is defined using
and optimal power flow [32]. While the modeling approach a function of the system admittance matrix. This was
is valid, the graphical models capture correlation whereas partly explored in our previous work in [1], [2] and used
our method models the underlying cause for that correlation for blind community detection.
structure thereby opening the door for statistical and non- 2) Study the spatio-temporal structure of the excitation that,
statistical approaches. at a fast time scale, is dominated by the generators
dynamics. It is shown that this excitation can be modeled
Note that with the exception of [24], [25], no other as an auto-regressive graph filter [51] (GF-AR (2))
papers make the connection with GSP and even in the for the input signals from the generator internal bus.
aforementioned work, GSP is used in an empirical manner. The spatial properties in quasi-steady state are captured
On the other hand, in our preliminary work in [1], [2], by defining another GSO using the ‘generator-only’ or
established a case for GSP in a more rigorous manner. Kron-reduced [9] network for the generator buses.
These models set the foundations to revisit known GSP based
algorithms for sampling and reconstruction, interpolation and
Detection of FDI attacks: While anomaly detection can denoising and network inference in the context of signal-
be broadly applied to identify various events, a large body processing PMU data. We harness the GFT for feature extrac-
of prior research has focused on False Data Injection (FDI) tion, to detect anomalies (specifically, FDI attacks [2]) and
attacks that can bypass classical bad data detection (BDD) to derive a lossy PMU voltage data compression algorithm,
mechanisms [33]–[36] and trigger incorrect decisions or hide leveraging the sparsity of the GFT signal. By elucidating
line overflows and contingencies [37]–[39]). FDI attacks to all the steps in modeling power systems data from graph
PMUs can, for instance, exploit the vulnerability of GPS construction, signal model, identification of the low-pass struc-
signals to spoofing attacks [40], [41]. ture that is responsible for low-dimensional representation to
Recent papers on PMU data integrity have proposed signal denoising, network inference and anomaly detection, we
leveraging the low rank spatio-temporal nature of PMU illustrate how one can similarly develop GSP based models in
data not only to help with erasures but also to strengthen other application domains especially data that can be modeled
conventional BDD mechanisms [42]. In [38] the authors as the output of low-pass graph signals [52].
have suggested an FDI attack strategy that can pass the
aforementioned BDD approach in [42] by generating false C. Paper organization
samples that approximately preserve the original subspace Section II reviews concepts from GSP focusing on complex-
structure. valued graph signals and applicable more broadly to bandpass
signals whose signal models rely on complex envelopes or
phasors. It also reviews measurements and parameters per-
Compression of PMU data: Due to their relatively high taining to the grid. Section III lays the foundation for Grid-
sampling rate and their wide deployment, PMU data have GSP mapping the physical laws to a spatio-temporal generative
have called for compression. The performance of several off- model for voltage signals. Through these lens, in Section IV,
the-shelf lossless encoding techniques applied to PMU data the paper revisits algorithms and tools from GSP for PMU
were investigated in [43]. In [44], a lossless compression data pertaining to sampling and reconstruction along with
called slack referenced encoding (SRE) method of PMU optimal placement of PMUs, interpolation of missing samples
voltage measurements is introduced, by identifying a slack- and network inference. Section V highlights applications of
bus and differentially encoding the difference between slack- Grid-GSP to detect FDI attacks and for sequential lossy
bus measurements in time and all other buses. Similarly, in voltage data compression. The algorithms and methods are
[45], phasor angle data is encoded in a lossless manner by tested numerically in Section VI. Section VII summarizes
preprocessing data using techniques from [44] and then using conclusions and future research directions.
Golomb-Rice entropy encoding. Notation: Boldfaced lowercase letters are used for vectors,
The idea of slow-variation with time in PMU data such as x and uppercase for matrices, A. Transpose is x> , A> and
phase angles is used in [46] to firstly transform data into conjugate transpose is AH . [x]M is the new vector that has
frequency domain (in time) and then using a ‘reverse water- elements of the vector indexed by the set M. The opera-
filling’ technique to encode frequency components in the tion <{.}, ={.} denote the real and imaginary parts of the
difference measurements. Many lossy compression techniques argument. Pseudo-inverse of a matrix is A† . The operation
utilize the low-rank structure inherent in voltage phasor data diag(x) creates a diagonal matrix with elements from a
[47]–[49]. Several other wavelet based compression algorithms vector x and Diag(A) is the extraction of diagonal values
exist in the literature as well [50]. of a matrix A.
 3

II. P RELIMINARIES After normalizing the eigenvectors such that kui k1 = 1 ∀i,
A. Graph Signal Processing (GSP) in a nutshell it is clear that S1 (ui ) > S1 (uj ) =⇒ |λi | > |λj |. Hence,
the ascending order of eigenvalues corresponds to increase in
Consider an undirected graph G = (N , E) with nodes i ∈ N
frequency, |λ0 | = 0 ≤ |λ1 | ≤ |λ2 | . . . |λ|N | |. This ordering is
and edges (i, j) ∈ E. A graph signal x ∈ C|N | is a vector
not unique since two distinct complex eigenvalues can have
whose ith entry [x]i is associated to node i ∈ N . The set of
the same magnitude.
nodes connected to node i is called the neighborhood of i and
The Graph Fourier Transform (GFT) basis is the complex
denoted as Ni . GSP generalizes the notion of discrete time
orthogonal basis U in (4). Hence, the GFT of a graph signal
shift for a time series by introducing the notion of graph shift
x, x̃ and the inverse GFT are given by x̃ = U> x and x =
operator (GSO):
Ux̃ respectively where [x̃]m is the frequency component that
Definition 1. A graph shift operator (GSO) is a linear corresponds to the m-th eigenvalue λm 2 . Also, we can define
neighborhood operator, so that each entry of the shifted graph the graph-frequency response of the graph filter, h̃, by writing
signal is a linear combination of the graph signal neighbors’ P
K
values [5]. H (S) = hk Sk = Udiag(h̃)U> , (6)
k=0
The linear combinations can use complex-valued weights K
X
sij ∈ C, with sij = 0, (i, j) ∈
/ E and, clearly, the GSO can be [h̃]i=H(λi ), H(λ) := hk λk (7)
defined as matrix multiplication, with a matrix S ∈ C|N |×|N | . k=0
Although not the only option, one common choice for the hk ∈ C, i = 1, 2, . . . |N |. The frequency response of the filter
GSO, is that of the graph weighted Laplacian1 , i.e: is given by elements in h̃. Subsequently, the input and output
(P of a graph filter in graph-frequency domain are related as
k∈Ni si,k , i = j
[S]i,j = (1) v = H(S)x → ṽ = diag(h̃)x̃, (8)
−si,j , i 6= j
which is analogous to convolution theorem for time-domain
In this work, we focus on complex symmetric GSOs, S = S>
signals. Naturally, this leads to the extension of notions such
as is applicable to the weighted graph Laplacian for the power
as low-pass, high-pass and band-pass filters and signals that
grid. Having defined the notion of shift, one can introduce the
are at the heart of sampling and interpolation schemes.
notion of shift-invariance:
Definition 2. Given a GSO S a shift invariant operator H B. GSP for time series of graph signals
acting on a graph signal is such that: So far, only the nodal index for the graph signal v was
H : x 7→ v, ⇐⇒ H : Sx 7→ Sv. (2) considered. However, one can also encounter graph signal
Linear shift-invariant operators must be matrix polynomials processes i.e. temporal variations in a graph signal {vt }t≥0 .
of the GSO S [53]. Therefore, a linear shift-invariant graph Since we are interested in the temporal characterization of
filter is a linear operator and can be defined as: voltage graph signals, we revise GSP concepts that are applied
to time series of graph signals [57], [58] in this subsection.
P
K−1
Then, we utilize these concepts while modeling the temporal
v = H(S)x, H (S) = hk Sk (3)
k=0 dynamics at generator buses in Section III-B.
Additionally, linear shift-invariant graph filters satisfy the In order to characterize graph signal process {vt }t≥0 , a joint
condition: SH (S) = H (S) S. time-vertex domain is considered in the literature by defining
Consider the following eigenvalue decomposition of the filters whose response is shift invariant with respect to the
complex symmetric GSO S, given by Theorem 4.4.13 in [54] time series shift operator z −1 and an appropriately chosen
for diagonalizable complex symmetric matrices: GSO [59]. To study the same, map the time series of graph
signal vt in both the graph frequency (GF) and z−domain by
S = UΛU> , U> U = UU> = I. (4) the application of z-transform to the GFT of the graph signal
Here Λ is the diagonal matrix with eigenvalues process:
P+∞
λ0 , λ1 , . . . λ|N |−1 on the principal diagonal and U are V (z) = t=0 vt z −t , Ve (z) = UT V (z), (9)
complex orthogonal eigenvectors. An equivalent concept of
frequency domain in GSP is defined using eigenvalues and We refer to Ve (z) as the z-GFT. A graph temporal filter’s [60]
eigenvectors of the GSO. impulse response Ht (S) and output vt are
Graph frequencies are the eigenvalues of the GSO and K
X t
X
the order of frequencies is based on the total variation (TV) Ht (S) = hk,t Sk , vt = Ht−τ (S)xτ , (10)
criterion [5], [55] defined using the discrete p Dirichlet form k=0 τ =0
Sp(x) with p = 1 as in [56] as: 2 It is worth noting that the graph shift operator and Fourier transforms do
not have in general important properties that are found in their conventional
S1(x) = kSxk1 =⇒ S1 (ui ) = kSui k1 = |λi |kui k1 (5) 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 closer to that of a derivative, since each of the GFT coefficients is rescaled by
as opposed to a shift operator, but we use the conventional name nonetheless the corresponding frequency. For complex symmetric, rather than Hermitian
in the rest of the paper to be consistent with the literature. operators, unfortunately also Parseval theorem is not valid.
4

respectively. Graph filter output vt in the z-domain is: where yi,j is the admittance of the branch between buses i and
+∞
X j if (i, j) ∈ E. The system admittance matrix Y is a complex
V (z) = H(S⊗z)X(z), where H(S⊗z) := Ht (S)z −t (11) symmetric matrix and it is equivalent to the complex-valued
t=0 graph Laplacian matrix associated with the power grid. Next,
when the input is xt with z-transform X(z) and Ht (S) are we will partition the nodes or buses into generator and non-
matrix polynomials of the GSO operator: generators, so that:
 
PK P
K Ygg Yg`
Ht (S) = hk,t Sk ↔ H(S⊗z) = Hk (z)Sk . (12) Y = , (17)
k=0 Yg`> Y``
k=0
Here Hk (z) is the z-transform of the filter hk,t . We can define where Ygg is the generator buses-only network, Yg` includes
also the following impulse response in the GF domain: the portion connecting generators and loads and Y`` corre-
PK sponds to the section of the grid connecting the loads buses
[h̃t ]i = Ht (λi ), Ht (λ) := k=0 hk,t λk (13)
among themselves. The shunt (fixed admittance to ground
and the graph-temporal joint transfer function in the z and GF at a bus) elements at all generator buses are denoted by
g
domain as: ysh ∈ C|NG | and at all load buses by ysh `
∈ C|NL | .
P+∞ PK The state of the system, from which all other physical
[h̃(z)]i = H(λi , z), H(λ, z) = t=0 k=0 hk,t λk z −t (14)
quantities of interest can be derived, are the voltage phasors
With that, we obtain following input-output relationship: at each bus. In the following we assume that a PMU installed
 
on node/bus i ∈ N provides a noisy measurement of voltage
Ṽ (z) = diag h̃(z) X̃(z), (15)
and current phasors at time t where v(t, i) = |v(t, i)|ejθ(t,i) .
by applying GFT to z-domain in (11). With some abuse of notation, we will refer to the PMU data
In this work, we focus on a class of graph-temporal filters as v(t, i) as well. Let the vector of voltage phasors collected
called GF-ARMA (q, r) filter [51], [60] . The input-output at time t be vt ∈ C|N | . After vt is partitioned into voltages
relation in both time and z-GFT domain are described below, at generator and non-generator buses, let igt ∈ C|NG | be the
respectively: generator current and i`t ∈ C|NL | the load current. Ohm’s law
vt −A1 (S)vt−1 · · ·−Aq (S)vt−q=B0 (S)xt+ · · ·+Br (S)xt−r , for a network is3 :
    g   g  g
ysh vt i
diag (ã(z)) Ṽ (z) = diag b̃(z) X̃(z), Y +diag ` vt = it , where vt = ` ,it = t` (18)
ysh vt it
P Pr
where a e (z) = 1 − qt=1 a e t z −t and e
b(z) = t=0 e bt z −t are To describe the operating conditions of the system we in-
the z-transform of the graph frequency responses of the graph troduce a few more quantities. In power systems transient
filter taps {At (S)}qt=1 , {Bt (S)}rt=0 for the GF-ARMA (q, r) dynamic analysis the impact of generating units is modeled
filter. Particularly, the GF-AR (2) filter is used in Section III-B as an internal bus characterized by a generator impedance
to describe generator temporal dynamics. (or admittance) yg ∈ C|NG | for g ∈ NG connected to an
C. Measurements and parameters of the electric grid ideal voltage source called internal voltage [61]; we denote
The electric grid network can be represented by an undi- its value at time t by E(t, i) = |E(t, i)|eδ( t,i) , i ∈ NG and the
rected graph G = (N , E) where nodes are buses and its corresponding vector as et ∈ C|NG | so that [et ]i = E(t, i).
edges are its transmission lines. The vertex set is a union The current at generator bus in (18), igt , is obtained as the
between set of generator, NG and non-generator/ load buses, multiplication of generator admittance and the difference in
NL , i ∈ {NG ∪ NL } = N and the edge set (i, j) ∈ E depicts voltage at the internal bus and the generator bus [7] :
electrical connections. To obtain Ohm’s law for a network of igt = diag(yg ) (et − vtg ) (19)
transmission lines, one starts from the telegrapher equations
for a single line to obtain the so-called ABCD parameters As mentioned in Section. I, the generators respond to electric
that relate input-output currents and voltages in the Fourier load in the grid. In order to model the generators response, a
domain. The equations are then rearranged and the so-called commonly used approximation is that at the load buses ` ∈
π-model is attained, which is an equivalent circuit containing NL are slowly varying admittances [7]. We denote them as
a series impedance element and parallel susceptance elements. y` (t) ∈ CN .
The π-model leads to the 2 × 2 branch admittance matrix that
relates current and voltage injections at the from and to ends III. G RAPH S IGNAL P ROCESSING FOR THE GRID
of a transmission line [61]. From the branch admittance matrix
Having described the relevant GSP concepts and introduced
of the network, applying Kirchhoff’s law, one can relate the
grid quantities and parameters of interest, we are ready to
current and voltage phasors for the entire network, introducing
introduce the Grid-GSP framework4 . Firstly, we define the
a system admittance matrix, Y ∈ C|N | [61] thus obtaining
the network version of Ohm’s law (see (18)). The matrix Y 3 Note that the admittances values are frequency responses evaluated at

is defined as: 60Hz (for the US) the voltage and current signals are the corresponding
(P envelopes at the same frequency; hence the assumption is the voltage and
k∈Ni yi,k , i = j currents are narrowband and the convolution can be approximated by gain
[Y ]i,j = (16) and phase rotation equal to the Fourier response at 60 Hz.
−yi,j , i 6= j 4 Our preliminary GSP modeling effort can be found in [1]
 5

GSO for the grid, then support the definition by introducing the power grids tend to be organized as communities system
graph-filter model that justifies it, and finally characterize its admittance matrix Y tends to be sparse [63]. Therefore the
temporal dynamics. All of the above yields a GSP generative GSO S has a high condition number and the graph frequency
model for the voltage phasor measurements as a low-pass GSP response of H(S) is such that it tapers off after a certain λk .
model, as detailed next. To visualize this more explicitly, consider ΛK to be the
diagonal matrix with entries λi , i ∈ K = {1, . . . , k}. Define a
A. Grid graph generative model low-pass filter with k frequency components and consequently
the voltage phasor measurements as
Grid-GSP for voltage phasors data relies on the following (
definitions:   h i λ−1 , i ∈ K
>
Hk (S) , U diag h̃k U , h̃k = i
Definition 3. The graph shift operator (GSO) is a complex i 0, else
 
symmetric matrix equal to a diagonal perturbation of the diag(yg )et
vt ≈ Hk (S) + ηt , (24)
system admittance matrix with generator admittance values, i`t
 g 
yg + ysh where Hk (S) will represent the principal subspace of the
S , Y + diag ` (20)
ysh voltage phasors whose dimensionality is the number of graph-
From the definition of the GSO it follows that: frequencies |K|. Therefore (24) defines the low-dimensional
generative model for quasi-steady state voltage phasor mea-
Definition 4. The grid Graph Fourier Transform (GFT) surements. The error term ηt now also captures modeling
basis for voltage phasors is the orthogonal matrix U given approximation.
by the eigenvalue decomposition of the GSO in Definition 3: To provide insights on the temporal dynamics of the voltage
S = UΛU> , |λmin | > 0 (21) phasors, we need to capture the structure of the excitation
term. As a matter of fact, et and i`t , have different dynamics,
Here, the GSO S is a complex-symmetric matrix that has as discussed in the subsequent subsections.
the same support as the electric-grid graph Laplacian as Y
with the diagonal addition of generator admittances. Note that B. A GSP model for generator dynamics: et
unlike the graph Laplacian, this GSO is invertible, |λmin | > 0.
Even when shunt elements ysh g `
, ysh are ignored as convention- The excitation term corresponding to generator currents has
ally done to solve power-flow problems in power systems, a elements as [et ]i coming from each generator i ∈ G. We
diagonal term with the generator admittances yg that is added illustrate a non-linear dynamical model for the generators
to the principal diagonal of Y , makes the GSO S invertible internal voltages, namely et ∈ CNG utilizing a GF-AR(2)
(see (20)). graph temporal filter from Section II-B. The model is inspired
With the GSO S is defined as in (20), one can rewrite (18) by the classical swing equations [8], [9] that describes the
and substitute for igt from (19) : coupled dynamics of the generators phase angles, δi (t), i ∈ G
  g    and the resulting variation in frequency, ωi (t) , δ˙i −ω0 where
ysh diag(yg ) (et − vtg ) ω0 = 2πf0 with f0 being the grid frequency (50 or 60 Hz).
Y +diag ` v t = (22)
ysh i`t Our model, relies on two steps. First, we model the dynam-
From now on with slight abuse of notation we denote vt as ics of a signal xt obtained through the following non-linear
voltage phasor measurements that are noisy therefore we add transformation of the internal generator voltages:
measurement noise ηt which yields the following equation, 1
xt , (diag(m)) 2 ln(et ) (25)
  − 21 − 12
diag(yg )et δt = (diag(m)) ={xt }, |e|t = (diag(m)) <{xt }
vt = H(S) + ηt (23)
i`t
where the vector m entries are the so-called generators
The Grid-GSP generative model for voltage phasor measure- masses, δt are the generators angles that appear in the swing
ments vt is given by (23). The linear shift-invariant graph filter equations and |e|t are internal generator voltage magnitudes.
is H(S) = S−1 . Second, like in the swing equations, to describe the generators
interactions, we resort to a Kron-reduction [8], [9] of the
Remark 1. Shift-invariance of H(S) = S−1 can be directly network, in which generators are all adjacent. To define this
verified from (2) i.e. H(S)S = SH(S) = I. Since H(S) is generator-only network and the corresponding GSO, consider
shift invariant, it can be expressed as a matrix polynomial in the following admittance matrix, Yall that describes the net-
S (Theorem 1 in [53]). Also, H(S) = S−1 can be written as in work topology consisting of the generator internal buses,
(3) where coefficients hk can be determined by the application generator buses and non-generator buses like done in [9]:
of Cayley-Hamilton theorem for inverse matrices [62].  g  g 
diag (yg +ysh ) − diag (yg +ysh ) 0
H(S) is approximately a low-pass graph filter [52] due to Yall =  g >
− diag (yg +ysh ) 0 Sci
the inversion of GSO since the graph frequency
  response of  >


Sci , S +diag 0 y`
the filter can be written from (7) as diag h̃ = Λ−1 . This
implies that as the graph frequency decreases, the magnitude In order to model Yall , it is assumed that the loads are varying
of the filter response declines. More importantly, since generic very slowly in time i.e. y` (t) ≈ y` ∀t. Then, let us denote
6

by Sh(A, B) the Schur complement of block B of matrix i.e. the homogeneous simplification as in [64], [65], the
A. We compute the Schur complement of block Sci of the dynamics of ={xt } can be justified with an GF-AR (2) model
matrix Yall which is nothing but Kron reduction. The Schur with GSO Sred :
complement of the Sci in Yall has two contributions:
={xt } − A1 (Sred )={xt−1 } − A2 (Sred )={xt−2 } ≈ ={w},
Sh(Yall , Sci ) = jYred + Ered (26) A1 (Sred ) := (2−χ)I − Sred , A2 (Sred ) := (χ−1)I (33)
Proof. Simple algebra on (32) allows us to recast the equations
where Ered is a real diagonal dominated matrix, and the
in the following form:
imaginary part Yred has the structure of a graph Laplacian.
The proposed dynamical model for the graph signal xt relies − 12
={ẍ} + χ={ẋ} = diag (m) w − Sred ={x}. (34)
on the following definition for the GSO of the Kron-reduced
generator-only graph: Assuming that the sampling rate is fast enough, and nor-
malizing it to 1, the finite difference approximations for the
Definition 5. A GSO is defined for the Kron-reduced generator derivatives are ẋ ≈ xt − xt−1 , ẍ ≈ xt+1 − 2xt + xt−1 and
only network as can be used to obtain AR (2) GF equations for the samples
1 1 ={xt } in (33).
Sred = (diag(m))− 2 Yred (diag(m))− 2 ∈ R|NG | (27)
The model that we introduce is simply extending the GF-
with the following eigenvalue decomposition, AR (2) model to capture both the real and imaginary part of
xt i.e. internal generator voltage magnitudes|e|t and angles
Sred = Ured Λred U>
red (28) δt respectively and suggesting to search the 2|NG | parameters
to fit the model with ã1 , ã2 rather than exploring a general
and the orthonormal GFT basis being Ured . MIMO filter response. For simplicity of representation, we
write the dynamical equation for xt in the GF domain,
We introduce the GSP based dynamical model for the
complex-valued generator internal voltages et via graph tem- x̃t = diag (ã1 ) x̃t−1 + diag (ã2 ) x̃t−2 + w̃t (35)
poral filter GF-AR (2) as follows,
such that the impulse response of the filter at graph-frequency
λred,i is defined by [ã1 ]i , [ã2 ]i . Note from (13) that [ã1 ]i , [ã2 ]i
GSP-based dynamics for generator internal voltages
 1

xt
et = exp (diag(m))− 2 xt (29) generator
L dynamics L − 12
w̃t x̃t Ured diag (m)
where xt is a GF-AR (2) process, i.e. the z-GFT X̃(z) N N
z −1 z −1 exp (.)
satisfies the following:
et
diag (ã(z)) X̃(z) = W̃ (z) (30) diag (ã1 ) diag (ã2 )  T
diag (yg ) 0
ã(z) = 1 − ã1 z −1 − ã2 z −2 (31)  
load current
L dynamicsL 0 L
t i`t
IL
The GSP based dynamical model takes inspiration from N N
swing equation for generator angles and we empirically choose z −1 z −1 H(S)
to model generator internal voltage magnitude |e|t also using a L
GF-AR (2) model although in most power system models, the diag (b1 ) diag (b2 ) ηt vt
dynamics of the amplitudes of the generators internal voltages voltage phasor
are typically ignored. The swing equation for the generators Fig. 1: Block diagram showing generative model for voltage
angles [64] are a key tool for power systems dynamical phasor measurements.
analysis:
can be written as polynomials in graph frequency λred,i ,
diag (m) δ̈ + diag (d) δ̇ = w − Yred δ, (32) PK1 −1 PK2 −1
[ã1 ]i = k=0 ak,1 λkred,i , [ã2 ]i = k=0 ak,2 λkred,i , (36)
where m are the generators masses, introduced previously, d which characterizes the poles of the AR system using graph
are the damping coefficients of generators (often neglected) frequencies. From (33), a first order polynomial may suffice
and w − Yred δ is the imbalance between the electrical and to characterize [ã1 ]i , [ã2 ]i in most cases.
mechanical power that triggers the change in generator angular
1
velocity and acceleration. Note that ={x} = (diag (m)) 2 δ. C. Load dynamics: i`t
We can manipulate (32) to prove the following:
There are several papers in the literature that deal with load
1
Proposition 1. Let w be such that ={w} = (diag(m))− 2 w. forecasting and modeling [66]. We adopt a simple AR-2 model
Using the approximation: per node or load bus to describe the dynamics of the load,

diag (d) diag (m)

−1
≈ χI, i`t = diag (b1 ) i`t−1 + diag (b2 ) i`t−2 + t (37)
 7

where parameters b1 , b2 are estimated load data time series. where sensors are not installed as long as correct placement
The block diagram in Fig.1 summarize our modeling efforts. strategies are devised i.e. that of choosing the subset M. Work
The unique nature of voltage phasor measurements allows in [26] explored the optimal placement for fault localization
us to describe a similar model for any subset of measurements in the under-sampled regime and also made connections with
on a graph. This is discussed next. GSP theory.

D. Low-pass property of down-sampled voltage graph signal IV. R EVISITING ALGORITHMS FROM GSP FOR PMU DATA
Let vM (time index t is ignored for simplicity) be the In this section we study some of the implications Grid-
down-sampled voltage graph signal where M ⊂ N is the GSP has while understanding sampling, optimal placement of
set of node indices at which measurements are available. measurement devices in power systems, interpolation of miss-
It can be shown that any down-sampled graph signal with ing samples and network inference. The underlying generative
arbitrary graph frequency response is low-pass in the reduced- model responsible for low-rank nature of data that has been
graph frequency domain. It suggests that one can utilize all established in the previous section helps explaining the success
the methods for low-pass graph signals onto down-sampled that many past works, such as [12], [69], [70], have attained
versions of the graph signal as well. This is summarized in in recovering missing PMU data using matrix completion
lemma 1 below. methods. The low-pass nature of the voltage graph signals
discussed in Section III provides the theoretical underpinning
Lemma 1. Let vM be any graph signal down-sampled in the that support the arguments made in the literature.
vertex-domain with |M| samples. Let the GSO defined with
respect to the full graph S be invertible. Then, with the GSO A. Sampling and recovery of grid-graph signals
defined with respect to the reduced-graph of M vertices as From the approximation in (24) we see that voltage graph
Sred,M , graph signal vM is the output of low-pass graph filter signals have graph frequency content that drops as λk grows.
H (Sred,M ) , S−1red,M
This characteristic renders the signal approximately band-
limited in the GFT domain [71] which means that there is
vM = H (Sred,M ) ϕ (38) a cut-off frequency λk such that frequency content corre-
where the GSO for the reduced-graph is given by Kron- sponding to λk+1 and higher is negligible. Let the GFT basis
reduction of S, Sred,M = Sh (S, SMc Mc ). corresponding to the first dominant k graph frequecies be UK .
The bandlimiting operator is, BK = UK U> K ∈ C
|N |×K
and
Proof. Consider a graph signal v with arbitrary graph fre- the low frequency component of vt is:
quency response with respect to GSO S,
BK vt = UK U>
K vt (43)
v = H (S) x = S−1 (SH (S) x) (39)
Similarly, a vertex limiting operator (with |M|) vertices is
The GSO S is rewritten in a 2 × 2 block form D M = PM PM >
where PM has columns that are coordinate
 
S SMMc vectors such that each column chooses a vertex/node. When
S = >MM , (40) the voltage measurements on the electrical network are from
SMMc SMc Mc
a few nodes, i ∈ M at time t, it can be written as [vt ]M =
and S−1 can be written using inverse formula for block PM >
vt . For reconstruction, results in [71] dictate the necessary
matrices. When graph signal v is down-sampled, only M rows condition be that |M| ≥ |K|. In the presence of modeling
are considered on both sides of (39). Thus we have, error relative to the perfect band-limited definition, optimal
ϕ sampling pattern i.e. the best placement for PMUs on the grid
z }|  {
vM = S−1
red,M I|M| −SMMc S−1
M M
c c (SH (S) x) (41) to minimize the worst-case reconstruction error is closely tied
to the grid topology and the model mismatch relative to a
where Sred,M is the Schur complement of the block SMc Mc strictly band-limited graph signal [72]. An optimal placement
in the GSO S i.e., strategy of PMUs that minimizes the worst-case reconstruction
Sred,M = Sh(S,SMc Mc) = SMM −SMMc S−1 > error in the presence of model mismatch due to imperfect
Mc Mc SMMc
band-limited nature of the voltage graph signal, also known
Lemma 1 translates to an interesting self- as the E-optimal design [72], is sought by maximizing the
similarity/fractional property for voltage graph signals smallest singular value, σmin (D M UK ), i.e. choose rows of
in that the down-sampled version vM is still a low-pass UK such that they are as uncorrelated as possible and the
graph signal. The self-similarity is due to H (S) = I. In resulting matrix has the highest condition number [71], [72].
summary, for voltage graph signals, Consider then the spatial sampling mask D M = diag(1M )

that selects M locations.
v = S−1 i, vM = S−1 −1
red,M iM −SMMc SMc Mc iMc (42)
1) Sampling: The optimal placement of M PMUs maxi-
In the power grid, this property has been illustrated empirically mizes σmin (D M UK ) which amounts to choosing the rows of
in several papers [12], [68] that highlight low-dimensionality UK with the smallest possible coherence (as close as possible
of measurements from a subset of buses. Although the to being orthogonal). In [71] and references therein, a greedy
reduced-graph is denser compared to the original graph, it still method is employed to find M rows from UK so that the least
helps to infer faults or events that occurred in a subset of nodes singular value is maximized.
8

Power systems topologies exhibit naturally a community satisfied. Therefore, estimation of GSO can be accomplished
structure that is reflected in the system admittance matrix Y by solving the following problem:
[1] due to population density or clusters of loads. It is known PT 2
that a method to determine k communities in a graph is to
min
S t=1 kSvt k1 + γkS − diag (Diag (S))kF

minimize the Ratio Cut [73] and spectral clustering performs T

X
a relaxed Ratio Cut minimization via k−means algorithm + kSvt − it k22 (47)
on rows of the eigenvectors UK [74]. Thus, choosing rows t=1
of UK to be uncorrelated is intuitively putting PMUs in subject to <[S]i,j = <[S]j,i , =[S]i,j = =[S]j,i , i 6= j, (48)
different graph-clusters or communities. This fact was also <Tr (S) = α|N |, =Tr (S) = β|N | (49)
discussed in [26] in the context of sensor placements for fault
localization. The PMUs sampling rate in time exceeds the Additional constraints on the GSO can be imposed based on
needs for reconstructions in a quasi-steady state conditions the properties of complex-symmetry (see (48)), sparse off-
by a significant margin and it is designed to help detect sharp diagonal entries via the term kS − diag (Diag (S))k2F and
transients in the system. dominant diagonal values (see (49)). Also, S tends to have
2) Reconstruction: Voltage data samples are obtained larger imaginary values than real especially on the diagonal.
down-sampling in space after the optimal placement of PMUs and α, β > 1 control the amplitude of real and imaginary
and also uniformly down-sampling in time. At time t when values on the diagonal. As before, the problem above can
|M| samples, [vt ]M are available, the following model applies be recast with down-sampled voltage graph signals to infer
the Kron-reduced GSO Sred,M with the approximation that
>
[vt ]M ≈ PM UK ṽt (44) the term SMMc S−1 Mc Mc iMc in (41) is treated as additive
Gaussian noise. Simulation results for network inference can
where ṽt is the GFT of graph signal vt Therefore, reconstruc- be found in Section VI.
tion in spatial domain is done via GFT basis as
>

† V. A PPLICATIONS OF G RID -GSP
v̂t = UK PM UK [vt ]M (45)
The goal of this section is to showcase the benefits of casting
Reconstruction in temporal domain can be done independently problems in the Grid-GSP framework through two exemplary
by up-sampling, i.e. via the windowed inverse Fourier trans- applications, namely anomaly detection and data compression.
form of the up-sampled signal created from uniformly time- The common thread between them is the use of the Grid-GFT
decimated data. as a tool to extract informative features from PMU data.

B. Interpolation of missing samples

A. Detection of FDI attacks on PMU measurements
When voltage measurements are missing or corrupted, de- This application is based on our preliminary work in [2].
noising and interpolation of such data can be cast as a graph Note that, even though we cast the problem as that of FDI
signal recovery problem by regularizing the total variation, attacks detection, the idea can be easily extended to unveil
(TV). Overall, the problem resembles time-vertex
 graph signal sudden changes due to physical events (like fault-currents, or
recovery [59]. Let V = v1 v2 . . . vT represent the topology changes) that similarly excite high GF content. We
voltage phasor measurements
  matrix collected over T time assume that we have access to PMU measurements of voltage
instants. Let PΩ V̂ be the set of available measurements and current from the buses they are installed on. Let A be the
that have samples in entries of set Ω and are noisy, set of available measurements where PMUs are installed and
  PT U be set of unavailable ones. A measurement model can be
min kPΩ V̂−V k2F +cg t=1 kSvt k1 written using ‘state’ to be the voltage as
V
PT     
+ ct t=2 kvt − vt−1 k22 (46) îA YAA YAU vA
= +ε (50)
v̂A I|A| 0 vU
where the two regularizing terms measure the variation in | {z } | {z } | {z }
zt H v
the graph and time domain and cg , ct are the corresponding
regularization constants. Importantly, one can use the GSO The attacker follows the strategy of FDI attack to manipulate
of the reduced graph, Sred,M if we only have access to a both current and voltage on the set of malicious buses, i ∈
subset of measurements on the grid, M and employ the same C ⊂ A by introducing a perturbation
formulation as in (46) for interpolation of missing samples. h i
δvtT = δvCT 0T|P|+|U | , such that YPC δvC = 0 (51)
C. Network inference as graph Laplacian learning where P is the set of honest nodes. This requires special
The problem of estimation of GSO S from voltage phasor conditions and placement, since YPC is tall. Nonetheless, since
measurements can be cast as a solving a problem similar the system admittance matrix Y is generally sparse [63], YPC
to graph Laplacian learning [75] which seeks the GSO that does not have full column-rank for a sufficient number of
minimizes the total variation of the observed voltage phasors. attackers C even when all the measurements are available with
If current measurements it are available, then another regular- A = N . Our detection problem entails deciding between the
ization term kSvt − it k22 can be added such that Ohm’s law is hypotheses of attack H1 and no attack H0 . To this end, we
 9

can leverage the low-dimensional generative model for the Algorithm 1 Encoding algorithm for compression
voltage graph signal that comes from (24), which imposes T
Input: x̃0 , x̃1 , i`0 , i`1 , {vt }t=2
additional constraint on the perturbation along with that in 1: for t = 2 : T do
(51). In short, for the attacker to be successful and undetected, 2: ˆ t , î`t from (59) and (60) respectively.
States x̃
she needs to have knowledge of system parameters and the 3: Voltage estimate:
graph filter with k frequency components Hk (S). However, " n o#
1
since the attacker does not have all this knowledge, a typical 0 diag(yg ) exp diag(m)− 2 Ured x̃ ˆt
FDI attack as studied in literature is launched using (51). Using v̂t = H(S) (57)
î`t
the generative model in (24), we know that under normal

operating conditions in quasi-steady state, the received data zt 4: Compute GFT of modeling n o error: ξ̃t = U> vt − v̂t0
under the no-attack and attack hypotheses H0 , H1 respectively ˆ ˆ
5: Quantize: ξ̃t = Q ξ̃t , v̂t = v̂t0 + Uξ̃t
have the structure: Update states,
 h
6:

H0 :HHk (S) (diag(yg )et )>
> i >
i`t +εt ˆ t ← U>
x̃
1
2 `
red diag(m) ln(êt ), ît ← [Sv̂t ]NL (58)
zt = h
i>
H :HH (S)
 > >
where êt = (diag(yg ))
−1
[Sv̂t ]NG
1 k (diag(yg )et ) i`t +Hδvt +εt
(52)
7:end for
n oT
Therefore, we project zt onto the subspace orthogonal to ˆ
Output: ξ̃t
columnspace of HHk (S) to get a test statistic, d(z). The t=2
projector is: " n 1
o#
diag (yg ) exp diag(m)− 2 Ured x̃0t + w̃t
Π⊥HHk (S) , I − [HHk (S)] [HHk (S)]
†
(53) vt = H(S)
i`,0
t + t
and under the no attack hypothesis H0 , energy in the orthog- where x̃0t = diag (a1 ) x̃t−1 + diag (a2 )x̃t−2 (59)
onal subspace is less than when there is an attack, H1 . This
i`,0
t = diag (b1 ) î`t−1 + diag (b2 ) î`t−2 (60)
can be converted to the following test,
H1 Thus, vt can be approximated as:
d(zt ) , kΠ⊥HHk (S) zt k22 R τ (54) " n 1
o# !
H0 diag (yg) exp diag(m)− 2 Ured x̃0t
vt≈ H(S) +ξt (61)
where τ is a threshold that can be chosen based on an empirical i`,0
t
receiver operator characteristics (ROC) curve. Note that, since
Note that, the vector ξt GFT, U> ξt has energy mostly in
Hk (S) is a low pass filter, the projector Π⊥HHk (S) in (53)
lower frequency components and is therefore an appropriate
is filtering high graph frequencies and the detection measures
term to quantize using an optimal rate allocation. Specifically,
the energy on such frequencies as a signature for anomalies.
Isolation of compromised buses or estimate of δv can also we allocate bits to each component by setting a desired level
undertaken with a similar logic. Firstly, using the assumptions of total distortion, applying the reverse water-filling result [77]
in the previous section we can solve the following regression which is optimum for a random vector U> ξt whose entries are
problem to recover δvt , formulating a LASSO relaxation of circularly symmetric complex independent Gaussian random
the sparse support recovery problem: variables and then quantize the components accordingly.5
Then, we use the quantized vector U> ξt to update the state
min kΠ⊥HHk (S) (z−Hδv)k22 subject to kδvt k1 ≤ µ (55) i.e. to estimate x̃t and i`t . Algorithms 1 and 2 describe the
δvt
encoding and decoding algorithms respectively. Note that the
Constraint on the `1 norm is used to incorporate the prior proposed scheme of compression is sequential unlike others in
knowledge that the attacker has access to a few measurement literature. Several corrections can be made as data is collected
buses, C  N . Note that the performance of the algorithm is in time such as the update of parameters ã1 , ã2 , b1 , b2 .
also dependent on the number of graph-frequency components
i.e. k considered.
VI. N UMERICAL R ESULTS
B. Compression of PMU measurements The numerical results in this section are mostly obtained
The proposed compression algorithm leverages both (23) using data from the synthetic ACTIVSg2000 case [78], a
and (35). The measure of distortion we use is the mean- realistic model emulating the ERCOT system, which includes
squared error (MSE): 2, 000 buses-with 432 generators and the rest non-generator
PT buses. The ACTIVSg2000 case data include a realistic PMU
d(v, v̂) , (|N |T )−1 t=0 kvt − v̂t k22 (56)
5 The covariance matrix is not diagonal and ideally one would first whiten
where T denotes the time instant at which samples are stopped the vector U> ξt and then quantize the individual components with bit-
collecting. Since we have a temporal dynamical model for the allocation akin to reverse water-filling. Since the statistics of U> ξt are time-
evolution of voltage in time, we use differential encoding [76] varying, one has to perform the whitening transform at each time instant which
is a cumbersome operation. Therefore we make the assumption of a diagonal
to quantize the residuals in both generator and load dynamics, covariance matrix while sacrificing the benefit of modeling the underlying
w̃t and t respectively. The voltage at time t is: correlations among the random variables.
10

Algorithm 2 Decoding algorithm for reconstruction GFT: |ṽ| = |U> v| GFT: |x̃t | = |U>
red v|
n oT
ˆ 101
Input: x̃0 , x̃1 , i`0 , i`1 , ξ̃t 100
t=2
1: for t = 2 : T do 100
ˆ t , î`t from (59) and (60) respectively. 10−2
2: States x̃ 10−1
ˆ
3: Reconstruction of voltage from (57) v̂t = v̂t0 + Uξ̃t 10−4
ˆ ` 10−2
4: Update states x̃t , ît from (58) 10−6
10 −4
10−3
10 −2
10 −1
100 0 0.2 0.4 0.6 0.8 1
5: end for |λred,i |/ maxi |λred,i |
T |λi |/ maxi |λi |
Output: {v̂t }t=2
Fig. 3: Magnitude of Graph Fourier Transform (GFT) for
voltage graph signal, |v˜t | (left) and input to generator-only
Graph Shift Operator (GSO) S network |x̃t | (right) plotted with respect to normalized graph
0
generator frequency.
buses

500 PMU in grid communities Optimally-placed PMU

|M| = 867 |M| = 261

red,M vM |
101 101
1000
100 100

GFT: |v˜M | = |U>

−1
10 10−1
1500 −2
10 10−2
10−3
10−3
2000 10−4
0 500 1000 1500 2000 10−4 10−3 10−2 10−1 100 10−4 10−3 10−2 10−1 100
non-generator buses
|λred,M,i |/ maxi |λred,M,i | |λred,M,i |/ maxi |λred,M,i |
Fig. 2: Support of the GSO S of the network. Fig. 4: Magnitude of GFT for spatially downsampled voltage
data time series, in which 392 generators are dispatched to graph signal, |ṽM | plotted with respect to normalized graph
meet variable load demand. The sampling rate, as for real frequency.
PMUs, is 30 samples per second. As all the system related To highlight the temporal variation in the GFT domain
parameters are known, it is easier to verify the proposed of input exponent, x̃t , a short time-series of the real and
modeling strategy through the ACTIVSg2000 PMU data set. imaginary parts along with the fit of the AR model are shown
Fig. 2 shows the support of the graph Laplacian or the Y in Fig. 5. As expected, the AR model fits well. Fig. 6 shows
matrix when ordered into generator and non-generator buses. the similar AR-2 model fit to the load current at a bus that had
The block-diagonal structure is notable, and is the result of the highest absolute value of load. To emphasize the temporal
the population distribution in the state of Texas, which is nature of the input, the 2-dimensional frequency response (in
concentrated in 8 metropolitan areas. both graph and time domains) is plotted for the input x̃t in
Grid-GSP model: In Fig. 3, magnitude of GFT of voltage Fig. 7. The figure provides evidence of the coupling between
graph signal vt and the input xt are plotted for a single the graph frequencies and time series Fourier power spectrum,
time instant with respect to their corresponding normalized and the variability of the temporal response depending on what
graph frequencies |λi |/ maxi |λi | and shown in log-scale. From GFT frequency mode is excited with Fourier spectra that are
the linear decay, it is evident that the magnitude of GFT more or less concentrated towards low frequencies depending
coefficients |ṽt | corresponding to lower frequencies are more on the GFT mode. Hence, temporal dynamics can inform about
significant as compared to higher frequencies. Similarly, the what is happening in space (i.e. the trends are coupled).
GFT of the exponent in the input, x̃t = U> red xt with the Revising GSP tools: sampling and optimal placement
generator GSO Sred , is plotted with respect to the graph Fig. 8 shows the placement of |M| = 100 PMUs super-
frequencies in Fig. 3. The decay in GFT coefficients with h i
respect to frequency is less pronounced confirming that graph [x̃t ]1 ˆt
x̃
−25.2 1
signal xt is not necessarily low-pass and in general depends
<{[x̃t ]1 }

−25.25
on the topology of the generator only network. −25.3
−25.35
In Fig. 4, magnitude of GFT of the downsampled voltage
graph signal, ṽM = U> red,M vM for |M| = 867 and |M| =
20 40 60 80 100 120 140 160 180 200 220
−4.98
={[x̃t ]1 }

261 with two different down-sampling strategies: with PMUs −5

placed at buses in few communities within the GSO S and the −5.02
other being optimal placement for graph signal reconstruction. −5.04
20 40 60 80 100 120 140 160 180 200 220
The placement strategy has an effect on the low-pass nature
time t (in s)
of the downsampled signal. The steeper attenuation of GFT
magnitude with placement strategy being community-wise is Fig. 5: AR model fit to the GFT of xt . Component corre-
a result of loss in spatial-resolution. sponding to smallest graph frequency, λred,1 , [x̃t ]1
 11

  h i imposed on the support of the ordered Grid-GSO, S when

<{ i`t j } <{ î`t }
−1.83 j |K| = 100 graph frequency components are considered.
<{ i`t j }
−1.84 Note the distribution of PMUs to different communities as
 

−1.85
well as on the generator buses as they belong to different
−1.86
graph communities. Fig. 9 exhibits the performance of the
20 40 60 80 100 120 140 160 180 200 220 GSP based reconstruction method on optimally placed |M|
2.34 PMUs that provide down-sampled measurements, vM . The
={ i`t j }
 

2.32 number of graph frequencies considered for reconstruction

2.3 are |K| = |M|. Even with just 5% of measurements (100
20 40 60 80 100 120 140 160 180 200 220
PMUs), the reconstruction error is extremely low. For random
time t (in s) placement, |M| PMUs are chosen at random and |K| = 100
Fig. 6: AR model for load bus, j = 1312, i.e. bus with highest graph frequencies are used for reconstruction. The trial of
absolute value of load. random placement is repeated 1, 000 times and the most
frequently occuring error (estimate of mode of the error
distribution) is plotted. As expected, the reconstruction error
for random placement is orders of magnitude higher than
optimal placement.
To illustrate that the proposed modeling holds and algo-
rithms work well also for real PMU data, in the next numerical
experiments we used a real-world dataset of measurements
from 35 PMUs placed in ISO New-England grid (ISO-NE)
[79]. The data corresponds to a period of 180 seconds when a
large generator near Ln:2 and Ln:4 introduces oscillations in
the system. We decimated in time the PMU signals down to
sampling frequency 1 sample/s.
Fig. 7: 2−D frequency response for x̃t Network inference: As the underlying GSO is unknown, it is
Grid GSO S PMU position estimated via (47) with the goal of recovering the underlying
reduced-GSO. Since admittance values are not given, we only
Grid GSO S PMU position compare the support of the estimated GSO with the community
0 of PMUs in the network. Fig. 10 shows the support of the
estimated GSO and compares it with the map of PMUs
500 0 highlighting a few clusters of correspondence. From Fig. 10
we see that the block-diagonal nature of the estimated GSO
bus number

500 captures the community structure in the map.

1000 Interpolation of missing measurements: Once the GSO is
bus number

estimated, we consider the interpolation problem in (46) for

1000
1500 the same ISO-NE dataset. We delete data at random and
add noise. We solve the problem in (46) to recover missing
1500 measurements. In Fig.11 we compare the original, corrupted
2000
0 500 1000 1500 2000 and recovered measurements. Corrupted measurements have
Fig. 8: Optimal 2000 bus|M|
placement of number
= 100 PMUs. |K| = 100. missing samples not just at random but also contiguous in time.
0 500 1000 1500 2000 The normalized MSE, kV − V̂k2F /kVk2F is the metric used to
bus number
gauge the reconstruction performance. As a comparison, we
Optimal Placement Random Placement
tested on the same data the AM-FIHT algorithm proposed in
[80], which regularizes the reconstruction task assuming that
MSE: (|N |T )−1 Tt=0 kvt − v̂t k22

1010
the Hankel matrix formed with the columns of V, i.e. H(V),
107 has low rank r,
 
104 minkPΩ V̂ − V k2F subject to rank (H(V)) = r (62)
P

V
101 The plot comparing the two methods is shown in Fig. 12. As
seen, the GSO based method outperforms the AM-FIHT for
10−2
this dataset, indicating that the regularization using the GSO
10−5
0 200 400 600 800 1,000 1,200
is more effective at capturing the low-rank nature of the data,
PMUs |M|
compared to seeking an arbitrary low rank structure in the the
Hankel matrix of the data.
Fig. 9: Reconstruction performance after optimal placement
Detection of FDI attacks: Fig.13 shows the magnitude of
[71] of |M| PMUs. Number of frequencies used: |K| = |M|.
the projection of the received measurement z on the orthogo-
For random placement, |K| = 100 used.
nal subspace Π⊥HHk (S) . From Fig. 13 it is evident that when
12

Area 3: External
Ln:21
Sub:9
Area 1: Internal
Ln:20
Ln:1
Sub:1
Ln:1-11
Ln:19 Sub:11
Ln:17 Ln:25
Sub:8 Ln:23 Ln:24
Ln:18
Gen1 Sub:10
Ln:22
Area 2: External

Ln:15
Ln:11
Ln:16 Sub:7
Sub:5 Ln:12
Ln:10
Gen2
Ln:9
Ln:27
Ln:28 Sub:12
Ln:8 Sub:4 Ln:4
Ln:7 Ln:6
Sub:3 Ln:26 Sub:2
Ln:3
Ln:4-6 Ln:2
Ln:5
Ln:14

Ln:13 Sub:6

Gen1

Fig. 10: The map of PMUs placed in ISO-NE test case 3 [79] (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.

FDI Attack No attack

Π⊥HHk (Sgrid ) z i 300



200

Fig. 11: Interpolation of missing measurements for an ISO- 100



NE case using GSO based regularization. Note the contiguous

missing of samples and our ability to interpolate. The relative 0
200 400 600 800 1,000
noise level used is, (|M|T )σ 2 /kVk2F = 10−4 Normalized i
MSE for this run is 6.22 × 10−4 . Fig. 13: Components of projection of received measurement z
onto the orthogonal subspace, Π⊥HHk (S) z, |A| = 500, |K| =
AM-FIHT GSO-based denoising 200, |C| = 250.
10−1
/ kVk2F

10% 20% 30% 40%

Probability of detection, Pd

1.2
F

10−2
2
NMSE: V − V̂

1
10−3 0.8
0.6
10−4 −5 0.4
10 10−4 10−3 10−2 10−1
relative noise level: (|M|T )σ 2
/ kVk2F 0.2
Fig. 12: Comparison of AM-FIHT algorithm in [80] (r = 0
0 0.2 0.4 0.6 0.8 1
10, n1 = 3, β = 0) and the proposed GSO based interpolation Probability of false alarm Pf
with 50% of missing measurements in the ISO-NE dataset.
Fig. 14: Empirical ROC curve for different |C| when a per-
there is no attack, the magnitude of the projected component centage of them are malicious, |C|/|A| × 100 with |A| = 500
is orders of magnitude lower than when the measurements are (out of 2, 000) are available.
under the FDI attack. This validates the idea of using high GFT
frequency activity as an indicator of anomalies. Fig. 14 shows the method in [25] when the full state i.e. when all voltage
the empirical receiver operator characteristics (ROC) curve measurements are available, |A| = 2, 000. The underlying
highlighting the detection performance of the proposed FDI principle to detect the attack in [25] is to look at the magnitude
attack detection scheme. The detection performance remains of graph frequency components at higher graph frequencies
good, even when very few buses are attacked. We compare the which is similar in principle to the detection test we undertake.
performance of the proposed FDI attack detection with that of They use the real and imaginary parts of the system admittance
 13

Actual Estimated Proposed SVD Scalar quantization

10−1

− v̂t k2F
Attack vector |δv|A
0.8
10−3

t=1 kvt
0.6
10−5

PT
0.4

MSE T −1
10−7
0.2

0 10−9
100 200 300 400 500 0 2 4 6 8 10
bus number
rate (bits/sample)
Fig. 15: Reconstruction of attack vector.
Fig. 17: Empirical rate distortion (RD) curve for the proposed
compression method compared with singular value threshold-
Proposed method 20% Drayer & Routtenberg 20%
Proposed method 50% Drayer & Routtenberg 50%
ing and quantization.
1.2
while the proposed method is sequential, which has important
Probability of detection, Pd

1 implication for the online communications of PMU data.

0.8
VII. C ONCLUSIONS
0.6
In this paper, we proposed the framework of Grid-GSP for
0.4 the power grid that highlights the inherent spatio-temporal
0.2 structure in the voltage phasors by employing concepts from
GSP. Grid-GSP revisits the concepts of sampling and recon-
0
0 0.2 0.4 0.6 0.8 1 struction, interpolation, network inference and applications, to
Probability of false alarm Pf detection of FDI attacks and a lossy sequential data compres-
Fig. 16: Empirical ROC curve for methods proposed here and sion, were introduced using the lens of GSP. The resulting
by Drayer & Routtenberg [25] when all voltage measurements algorithms were tested on data from both synthetic and real-
are available, |A| = 2, 000. A percent of the measurements, world datasets. The paper opens the door to leverage the GSP
|C|/|A|×100 are malicious. The relative noise level is 10−2 . foundations for all types of grid data analytical tasks.
matrix as 2 GSOs, <{Y } and ={Y } respectively. Their R EFERENCES
test statistic is comprised of four components that are the
[1] R. Ramakrishna and A. Scaglione, “On Modeling Voltage Phasor Mea-
frequency response of high-pass filtered real and imaginary surements as Graph Signals,” in IEEE Data Science Workshop (DSW)
voltage measurements (see Algorithm.2 in [25]). Fig. 16 shows 2019, June 2019, pp. 275–279.
the empirical ROC curves that compare the performance of [2] R. Ramakrishna and A. Scaglione, “Detection of False Data Injection
Attack using Graph Signal Processing for the Power Grid,” in 2019 IEEE
the proposed method and the one in [25] when all voltage Global Conference on Signal and Information Processing (GlobalSIP).
measurements are available and are noisy. The relative noise IEEE, 2019.
level used is 10−2 . As evident from the curves, the proposed [3] D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Van-
dergheynst, “The emerging field of signal processing on graphs: Ex-
method performs better than the method in [25]. This is tending high-dimensional data analysis to networks and other irregular
because our test statistic is more robust to noise and also domains,” IEEE Signal Processing Magazine, 2013.
significantly more sensitive in detecting the attack vectors, [4] N. Tremblay, P. Gonçalves, and P. Borgnat, “Design of graph filters and
filterbanks,” in Cooperative and Graph Signal Processing. Elsevier,
even when only few buses are attacked. 2018, pp. 299–324.
[5] A. Sandryhaila and J. M. F. Moura, “Discrete Signal Processing on
Fig. 15 shows the reconstruction of magnitude of the attack Graphs: Frequency Analysis,” IEEE Transactions on Signal Processing,
vector δv when 500 measurements are available and number vol. 62, no. 12, June 2014.
of attacked buses |C| = 50. [6] A. Phadke and J. Thorp, “History and Applications of Phasor Measure-
ments,” in 2006 IEEE PES Power Systems Conference and Exposition.
Compression based results: For voltage data compression, IEEE, 2006, pp. 331–335.
we compared with two schemes: scalar quantization and [7] T. Ishizaki, A. Chakrabortty, and J.-I. Imura, “Graph-Theoretic Analysis
of Power Systems,” Proceedings of the IEEE, vol. 106, no. 5, pp. 931–
singular value thresholding (SVT) from [49]. Fig. 17 plots the 952, May 2018.
empirical rate-distortion (RD) curve and shows the comparison [8] F. Dörfler and F. Bullo, “Kron Reduction of Graphs With Applications
between all 3 schemes. As expected, scalar quantization does to Electrical Networks,” IEEE Transactions on Circuits and Systems-I:
Regular Papers, vol. 60, no. 1, 2013.
poorly compared to the other schemes. The SVT scheme [9] F. Dorfler and F. Bullo, “Spectral Analysis of Synchronization in a
simply uses few of the largest singular vectors for data Lossless Structure-Preserving Power Network Model,” in 2010 First
reconstruction. Considering that it is indicative of voltage IEEE International Conference on Smart Grid Communications, 2010.
[10] L. Xie, Y. Chen, and P. R. Kumar, “Dimensionality reduction of
graph signal lying in a low-dimensional subspace, it is not synchrophasor data for early event detection: Linearized analysis,” IEEE
surprising that the SVT scheme does well. However, the Trans. on Power Systems, vol. 29, no. 6, pp. 2784–2794, Nov 2014.
SVT curve rate-distortion curve eventually saturates. Note that [11] P. Gao, M. Wang, S. G. Ghiocel, J. H. Chow, B. Fardanesh, and G. Ste-
fopoulos, “Missing Data Recovery by Exploiting Low-dimensionality
the performance of the proposed method are comparable to in Power System Synchrophasor Measurements,” IEEE Transactions on
those of the SVT. However, the latter is a batch method, Power Systems, vol. 31, no. 2, pp. 1006–1013, 2016.
14

[12] M. Wang, “Data quality management of synchrophasor data in power [33] R. B. Bobba, K. M. Rogers, Q. Wang, H. Khurana, K. Nahrstedt,
systems by exploiting low-dimensional models,” in 2017 51st Annual and T. J. Overbye, “Detecting false data injection attacks on DC state
Conference on Information Sciences and Systems (CISS), 2017. estimation,” in Preprints of the First Workshop on Secure Control
[13] M. Jamei, A. Scaglione, C. Roberts, E. Stewart, S. Peisert, C. McParland, Systems, CPSWEEK, vol. 2010, 2010.
and A. McEachern, “Anomaly Detection Using Optimally-Placed PMU [34] G. Dan and H. Sandberg, “Stealth attacks and protection schemes for
Sensors in Distribution Grids,” IEEE Transactions on Power Systems, state estimators in power systems,” in Smart Grid Communications
vol. 33, no. 4, pp. 3611–3622, July 2018. (SmartGridComm), 2010 First IEEE International Conference on, 2010,
[14] W. Li, M. Wang, and J. H. Chow, “Real-Time Event Identifica- pp. 214–219.
tion Through Low-Dimensional Subspace Characterization of High- [35] O. Kosut, L. Jia, R. J. Thomas, and L. Tong, “Malicious data attacks
Dimensional Synchrophasor Data,” IEEE Transactions on Power Sys- on the smart grid,” IEEE Transactions on Smart Grid, vol. 2, no. 4, pp.
tems, vol. 33, no. 5, Sept 2018. 645–658, 2011.
[15] J. Kim, L. Tong, and R. J. Thomas, “Subspace Methods for Data Attack [36] G. Liang, J. Zhao, F. Luo, S. R. Weller, and Z. Y. Dong, “A review
on State Estimation: A Data Driven Approach,” IEEE Transactions on of false data injection attacks against modern power systems,” IEEE
Signal Processing, vol. 63, no. 5, pp. 1102–1114, March 2015. Transactions on Smart Grid, vol. 8, no. 4, pp. 1630–1638, July 2017.
[16] P. Du, Z. Huang, R. Diao, B. Lee, and K. K. Anderson, “PMU Placement [37] I. Esnaola, S. M. Perlaza, H. V. Poor, and O. Kosut, “Maximum
for enhancing dynamic observability of a power grid,” in 2010 IEEE Distortion Attacks in Electricity Grids,” IEEE Transactions on Smart
Conference on Innovative Technologies for an Efficient and Reliable Grid, vol. 7, no. 4, pp. 2007–2015, Jul. 2016.
Electricity Supply, Sept 2010, pp. 15–21. [38] J. Zhang, Z. Chu, L. Sankar, and O. Kosut, “False data injection
[17] A. Pal, G. A. Sanchez-Ayala, V. A. Centeno, and J. S. Thorp, “A PMU attacks on phasor measurements that bypass low-rank decomposition,”
Placement Scheme Ensuring Real-Time Monitoring of Critical Buses of in IEEE International Conference on Smart Grid Communications
the Network,” IEEE Transactions on Power Delivery, vol. 29, no. 2, pp. (SmartGridComm), Oct. 2017.
510–517, April 2014. [39] Y. He, G. J. Mendis, and J. Wei, “Real-time detection of false data
[18] Z. Wang, A. Scaglione, and R. Thomas, “Generating Statistically Correct injection attacks in smart grid: A deep learning-based intelligent mech-
Random Topologies for Testing Smart Grid Communication and Control anism,” IEEE Transactions on Smart Grid, vol. 8, no. 5, pp. 2505–2516,
Networks,” IEEE Transactions on Power Systems, vol. 1, no. 1, pp. 28– Sept 2017.
39, 2010. [40] D. P. Shepard, T. E. Humphreys, and A. A. Fansler, “Evaluation of
[19] Li, Xiao and Poor, H Vincent and Scaglione, Anna, “Blind topology the vulnerability of phasor measurement units to gps spoofing attacks,”
identification for power systems,” in 2013 IEEE International Confer- International Journal of Critical Infrastructure Protection, vol. 5, no.
ence on Smart Grid Communications (SmartGridComm). IEEE, 2013, 3-4, pp. 146–153, 2012.
pp. 91–96. [41] L. Heng, J. J. Makela, A. D. Dominguez-Garcia, R. B. Bobba, W. H.
[20] S. Grotas, Y. Yakoby, I. Gera, and T. Routtenberg, “Power Systems Sanders, and G. X. Gao, “Reliable GPS-based timing for power systems:
Topology and State Estimation by Graph Blind Source Separation,” A multi-layered multi-receiver architecture,” in Power and Energy
IEEE Transactions on Signal Processing, vol. 67, no. 8, pp. 2036–2051, Conference at Illinois (PECI), 2014, 2014, pp. 1–7.
2019. [42] P. Gao, M. Wang, J. H. Chow, S. G. Ghiocel, B. Fardanesh, G. Stefopou-
los, and M. P. Razanousky, “Identification of successive ”unobservable
[21] Z. Xiang, K. Huang, W. Deng, and C. Yang, “Blind Topology Identifi-
” cyber data attacks in power systems through matrix decomposition,”
cation for Smart Grid Based on Dictionary Learning,” in 2019 IEEE
IEEE Transactions on Signal Processing, vol. 64, no. 21, pp. 5557–5570,
Symposium Series on Computational Intelligence (SSCI), 2019, pp.
Nov 2016.
1319–1326.
[43] P. Top and J. Breneman, “Compressing phasor measurement data,” in
[22] D. Deka, M. Chertkov, and S. Backhaus, “Joint Estimation of Topology
2013 IEEE Power Energy Society General Meeting, July 2013, pp. 1–4.
and Injection Statistics in Distribution Grids With Missing Nodes,” IEEE
[44] R. Klump, P. Agarwal, J. E. Tate, and H. Khurana, “Lossless com-
Transactions on Control of Network Systems, vol. 7, no. 3, pp. 1391–
pression of synchronized phasor measurements,” in IEEE PES General
1403, 2020.
Meeting, July 2010, pp. 1–7.
[23] S. Talukdar, D. Deka, H. Doddi, D. Materassi, M. Chertkov, and M. V.
[45] J. E. Tate, “Preprocessing and Golomb-Rice encoding for lossless
Salapaka, “Physics informed topology learning in networks of linear
compression of phasor angle data,” IEEE Trans. on Smart Grid, vol. 7,
dynamical systems,” Automatica, vol. 112, 2020.
no. 2, pp. 718–729, March 2016.
[24] E. Drayer and T. Routtenberg, “Detection of False Data Injection Attacks [46] S. Kirti, Z. Wang, A. Scaglione, and R. Thomas, “On the Commu-
in Power Systems with Graph Fourier Transform,” in 2018 IEEE Global nication Architecture for Wide-Area Real-Time Monitoring in Power
Conference on Signal and Information Processing (GlobalSIP). IEEE, Networks,” in 2007 40th Annual Hawaii International Conference on
2018, 2018, pp. 890–894. System Sciences (HICSS’07), 2007.
[25] E. Drayer and T. Routtenberg, “Detection of False Data Injection Attacks [47] P. H. Gadde, M. Biswal, S. Brahma, and H. Cao, “Efficient compression
in Smart Grids Based on Graph Signal Processing,” IEEE Systems of PMU data in WAMS,” IEEE Trans. on Smart Grid, vol. 7, no. 5, pp.
Journal, August 2019. 2406–2413, Sept 2016.
[26] M. Jamei, R. Ramakrishna, T. Tesfay, R. Gentz, C. Roberts, [48] Y. Ge, A. J. Flueck, D. K. Kim, J. B. Ahn, J. D. Lee, and D. Y. Kwon,
A. Scaglione, and S. Peisert, “Phasor Measurement Units Optimal “Power system real-time event detection and associated data archival
Placement and Performance Limits for Fault Localization,” IEEE reduction based on synchrophasors,” IEEE Trans. on Smart Grid, vol. 6,
Journal on Selected Areas in Communications , 2019. no. 4, pp. 2088–2097, July 2015.
[27] L. Fan, “Interarea Oscillations Revisited,” IEEE Transactions on Power [49] J. C. S. de Souza, T. M. L. Assis, and B. C. Pal, “Data Compression in
Systems, vol. 32, no. 2, pp. 1585–1586, 2017. Smart Distribution Systems via Singular Value Decomposition,” IEEE
[28] T. Huang, N. M. Freris, P. R. Kumar, and L. Xie, “Localization of Forced Transactions on Smart Grid, vol. 8, no. 1, January 2017.
Oscillations in the Power Grid Under Resonance Conditions,” in 2018 [50] R. Mehra, V. Patel, F. Kazi, N. M. Singh, and S. R. Wagh, “Modes
52nd Annual Conference on Information Sciences and Systems (CISS), preserving wavelet based multi-scale PCA algorithm for compression
2018. of smart grid data,” in 2013 International Conference on Advances in
[29] L. Guo, C. Zhao, and S. H. Low, “Graph Laplacian Spectrum and Computing, Communications and Informatics (ICACCI), Aug 2013, pp.
Primary Frequency Regulation,” in 2018 IEEE Conference on Decision 817–821.
and Control (CDC), 2018, pp. 158–165. [51] E. Isufi, A. Loukas, A. Simonetto, and G. Leus, “Autoregressive Moving
[30] Y. Weng, R. Negi, and M. D. Ilić, “Graphical model for state estimation Average Graph Filtering,” IEEE Transactions on Signal Processing,
in electric power systems,” in 2013 IEEE International Conference on vol. 65, no. 2, pp. 274–288, 2017.
Smart Grid Communications (SmartGridComm). IEEE, 2013, pp. 103– [52] R. Ramakrishna, H. T. Wai, and A. Scaglione, “A User Guide to
108. Low-Pass Graph Signal Processing and Its Applications: Tools and
[31] D. Deka, M. Chertkov, and S. Backhaus, “Topology Estimation Using Applications,” IEEE Signal Processing Magazine, vol. 37, no. 6, pp.
Graphical Models in Multi-Phase Power Distribution Grids,” IEEE 74–85, November 2020.
Transactions on Power Systems, vol. 35, no. 3, pp. 1663–1673, May [53] A. Sandryhaila and J. M. Moura, “Discrete Signal Processing on
2020. Graphs,” IEEE Transactions on Signal Processing, vol. 61, no. 7, pp.
[32] K. Dvijotham, M. Chertkov, P. Van Hentenryck, M. Vuffray, and 1644–1656, 2013.
S. Misra, “Graphical models for optimal power flow,” Constraints, [54] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge University
vol. 22, no. 1, pp. 24–49, 2017. Press, 1990.
 15

[55] S. Mallat, A Wavelet Tour of Signal Processing: The Sparse Way. [80] S. Zhang, Y. Hao, M. Wang, and J. H. Chow, “Multi-Channel Hankel
Academic press, 2009. Matrix Completion through Nonconvex Optimization,” IEEE Journal of
[56] R. Singh, A. Chakraborty, and B. Manoj, “Graph Fourier transform Selected Topics in Signal Processing, vol. 12, no. 4, pp. 617–632, 2018.
based on direcetd Laplacian,” in 2016 International Conference on
Signal Processing and Communications (SPCOM), 2016.
[57] E. Isufi, G. Leus, and P. Banelli, “2-Dimensional finite impulse response
graph-temporal filters,” in 2016 IEEE Global Conference on Signal and
Information Processing (GlobalSIP), 2016.
[58] E. Isufi, A. Loukas, A. Simonetto, and G. Leus, “Filtering Random
Graph Processes Over Random Time-Varying Graphs,” IEEE Transac-
tions on Signal Processing, vol. 65, no. 16, August 2017.
[59] F. Grassi, A. Loukas, N. Perraudin, and B. Ricaud, “A Time-Vertex
Signal Processing Framework: Scalable Processing and Meaningful
Representations for Time-Series on Graphs,” IEEE Transactions on
Signal Processing, vol. 66, no. 3, pp. 817–829, February 2018.
[60] E. Isufi, A. Loukas, A. Simonetto, and G. Leus, “Separable autore-
gressive moving average graph-temporal filters,” in 2016 24th European
Signal Processing Conference (EUSIPCO). IEEE, 2016, pp. 200–204.
[61] J. D. Glover, M. S. Sarma, and T. J. Overbye, Power system analysis
and design. Cengage Learning, 2008.
[62] H. P. Decell, Jr, “An application of the cayley-hamilton theorem to
generalized matrix inversion,” SIAM Review, vol. 7, no. 4, pp. 526–528,
1965.
[63] N. Sato and W. F. Tinney, “Techniques for Exploiting the Sparsity or the
Network Admittance Matrix,” IEEE Transactions on Power Apparatus
and Systems, vol. 82, no. 69, pp. 944–950, 1963.
[64] P. Huynh, H. Zhu, Q. Chen, and A. E.Elbanna, “Data-Driven Estimation
of Frequency Response From Ambient Synchrophasor Measurements,”
IEEE Transactions on Power Systems, vol. 33, no. 6, 2018.
[65] F. Paganini and E. Mallada, “Global Analysis of Synchronization
Performance for Power Systems: Bridging the Theory-Practice Gap,”
IEEE Transactions on Automatic Control, vol. 65, no. 7, pp. 3007–3022,
2020.
[66] F. Gao, J. S. Thorp, A. Pal, and S. Gao, “Dynamic state prediction based
on Auto-Regressive (AR) Model using PMU data,” in 2012 IEEE Power
and Energy Conference at Illinois. IEEE, 2012, pp. 1–5.
[67] G. Golub and C. V. Loan, Matrix Computations. The Johns Hopkins
University Press, 1990.
[68] M. Wang, J. H. Chow, P. Gao, X. T. Jiang, Y. Xia, S. G. Ghiocel,
B. Fardanesh, G. Stefopolous, Y. Kokai, N. Saito, and M. Razanousky,
“A low-rank matrix approach for the analysis of large amounts of
power system synchrophasor data,” in 2015 48th Hawaii International
Conference on System Sciences, Jan 2015, pp. 2637–2644.
[69] N. Dahal, R. L. King, and V. Madani, “Online dimension reduction
of synchrophasor data,” in IEEE PES Transmission and Distribution
Conference and Exposition (T & D), 2012.
[70] M. Liao, D. Shi, Z. Yu, Z. Yi, Z. Wang, and Y. Xiang, “An Alternat-
ing Direction Method of Multipliers Based Approach for PMU Data
Recovery,” IEEE Transactions on Smart Grid, 2018.
[71] M. Tsitsvero, S. Barbarossa, and P. Di Lorenzo, “Signals on Graphs:
Uncertainty Principle and Sampling,” IEEE Transactions on Signal
Processing, vol. 64, no. 18, pp. 4845–4860, 2016.
[72] A. Anis, A. Gadde, and A. Ortega, “Efficient Sampling Set Selection
for Bandlimited Graph Signals Using Graph Spectral Proxies,” IEEE
Transactions on Signal Processing.
[73] L. Hagen and A. B. Kahng, “New Spectral Methods for Ratio Cut
Partitioning and Clustering,” IEEE Transactions on Computer-Aided
Design of Integrated Circuits and Systems, vol. 11, no. 9, pp. 1074–
1085, 1992.
[74] U. Von Luxburg, “A tutorial on spectral clustering,” Statistics and
computing, Springer, vol. 17, no. 4, pp. 395–416, 2007.
[75] X. Dong, D. Thanou, M. Rabbat, and P. Frossard, “Learning graphs
from data: A signal representation perspective,” IEEE Signal Processing
Magazine, vol. 36, no. 3, pp. 44–63, 2019.
[76] W. Weber, “Differential Encoding for Multiple Amplitude and Phase
Shift Keying Systems,” IEEE Transactions on Communication, vol. 26,
no. 3, 1978.
[77] T. M. Cover and J. Thomas, Elements of Information Theory. John
Wiley, 1991.
[78] A. B. Birchfield, T. Xu, K. M. Gegner, K. S. Shetye, and T. J. Overbye,
“Grid Structural Characteristics as Validation Criteria for Synthetic
Networks,” IEEE Transactions on Power Systems, vol. 32, no. 4, pp.
3258–3265, July 2017.
[79] S. Maslennikov, B. Wang, Q. Zhang, F. Ma, X. Luo, K. Sun, and
E. Litvinov, “A Test Cases Library for Methods Locating the Sources
of Sustained Oscillations,” in IEEE PES General Meeting, Boston, MA,
2016.