IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO.

15, AUGUST 1, 2019 4069

Real-Time Power System State Estimation and

Forecasting via Deep Unrolled Neural Networks
Liang Zhang , Student Member, IEEE, Gang Wang , Member, IEEE, and Georgios B. Giannakis , Fellow, IEEE

Abstract—Contemporary power grids are being challenged by grid becomes increasingly critical, not only for detection of sys-
rapid and sizeable voltage fluctuations that are caused by large- tem instabilities and protection [8], [28], but also for energy
scale deployment of renewable generators, electric vehicles, and management [8], [34].
demand response programs. In this context, monitoring the grid’s
operating conditions in real time becomes increasingly critical. Given the grid parameters and a set of measurements pro-
With the emergent large scale and nonconvexity, existing power vided by the supervisory control and data acquisition (SCADA)
system state estimation (PSSE) schemes become computationally system, PSSE aims to retrieve the unknown system state, that
expensive or often yield suboptimal performance. To bypass these is, complex voltages at all buses [30]. Commonly used state
hurdles, this paper advocates physics-inspired deep neural net- estimators include the weighted least-squares (WLS) and least-
works (DNNs) for real-time power system monitoring. By unrolling
an iterative solver that was originally developed using the exact absolute-value (LAV) ones, derived based on (weighted) 1 - or
ac model, a novel model-specific DNN is developed for real-time 2 -loss. To tackle the resultant nonconvex optimization, differ-
PSSE requiring only offline training and minimal tuning effort. To ent solvers have been proposed; see e.g., [28], [30]. However,
further enable system awareness, even ahead of the time horizon, those optimization-oriented PSSE schemes either require many
as well as to endow the DNN-based estimator with resilience, deep iterations or are computationally intensive, and they are further
recurrent neural networks (RNNs) are also pursued for power sys-
tem state forecasting. Deep RNNs leverage the long-term nonlinear challenged by growing dynamics and system size. These con-
dependencies present in the historical voltage time series to enable siderations motivate novel approaches for real-time large-scale
forecasting, and they are easy to implement. Numerical tests show- PSSE.
case improved performance of the proposed DNN-based estimation To that end, PSSE using plain feed-forward neural networks
and forecasting approaches compared with existing alternatives. (FNNs) was studied in [3], [19], [32], [33]. Once trained off-
In real load data experiments on the IEEE 118-bus benchmark
system, the novel model-specific DNN-based PSSE scheme outper- line using historical data and/or simulated samples, FNNs can
forms nearly by an order-of-magnitude its competing alternatives, be implemented for real-time PSSE, as the inference entails only
including the widely adopted Gauss–Newton PSSE solver. a few matrix-vector multiplications. Related approaches using
Index Terms—Power system state estimation, forecasting, least-
FNNs that ‘learn-to-optimize’ emerge in wireless communica-
absolute-value, proximal linear algorithm, recurrent neural net- tions [24], and outage detection [36]. Unfortunately, past ‘plain-
works, data validation. vanilla’ FNN-based PSSE schemes are model-agnostic, which
often require a non-trivial tuning effort, and yield suboptimal
I. INTRODUCTION performance. To devise NNs in a disciplined manner, recent
ECOGNIZED as the most significant engineering achieve- proposals in computer vision [10], [31] constructed deep (D)
R ment of the twentieth century, the North American power
grid is a complex cyber-physical system with transmission and
NNs by unfolding iterative solvers tailored to model-based op-
timization problems.
distribution infrastructure delivering electricity from generators In this work, we will pursue model-specific DNNs for PSSE
to consumers. Due to the growing deployment of distributed re- by unrolling existing iterative optimization-based PSSE solvers.
newable generators, electric vehicles, and demand response pro- On the other hand, PSSE by itself may be insufficient for sys-
grams, contemporary power grids are facing major challenges tem monitoring when states exhibit large variations (i.e., system
related to unprecedented levels of load peaks and voltage fluctu- dynamics) [28]. In addition, PSSE works (well) only if there
ations. In this context, real-time monitoring of the smart power are enough measurements achieving system observability, and
the grid topology along with the link parameters are precisely
Manuscript received February 25, 2019; revised May 25, 2019; accepted June known. To address these challenges, power system state fore-
20, 2019. Date of publication July 3, 2019; date of current version July 12, 2019. casting to aid PSSE [6], [22] is well motivated.
The associate editor coordinating the review of this manuscript and approv-
ing it for publication was Prof. Sotirios Chatzis. This work was supported by Power system state forecasting has so far been pursued via
the National Science Foundation under Grant 1508993, Grant 1509040, Grant (extended) Kalman filtering and moving horizon approaches
1514056, and Grant 1711471. This paper was presented in part at the IEEE in e.g., [5], [11], [17], and also through first-order vector auto-
Global Conference on Signal and Information Processing, Anaheim, CA, USA,
November 26–29, 2018. (Corresponding author: Gang Wang.) regressive (VAR) modeling [12]. Nonetheless, all the aforemen-
The authors are with the Digital Technology Center and the Depart- tioned state predictors, assume linear dynamics; yet in practice,
ment of Electrical and Computer Engineering, University of Minnesota, Min- the dependence of the current state on previous (estimated)
neapolis, MN 55455 USA (e-mail: [email protected]; [email protected];
[email protected]). one(s) is nonlinear and cannot be accurately characterized.
Digital Object Identifier 10.1109/TSP.2019.2926023 To render nonlinear estimators tractable, FNN-based state

1053-587X © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Saskatchewan. Downloaded on September 14,2021 at 03:51:12 UTC from IEEE Xplore. Restrictions apply.
4070 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 15, AUGUST 1, 2019

{Qfnn ,t }(n,n )∈Eto , {Pnn t t

 ,t }(n,n )∈E o , {Qnn ,t }(n,n )∈E o ]
t t

be the
measurement vector that collects all measured quantities at
time t, where sets Nto and Eto indicate the locations where the
corresponding nodal and line quantities are measured.
Per time instant t, PSSE aims to recover the system state vec-
r i r i
tor vt := [v1,t v1,t · · · vN,t vN,t ] ∈ R2N (in rectangular coor-
dinates) from generally noisy measurement zt . For brevity, the
subscript t of zt and vt will be omitted when discussing PSSE in
Fig. 1. Bus n is connected to bus n via line (n, n ). Sections II and III. Mathematically, the PSSE task can be posed
as follows.
prediction has been advocated with the transition mapping Given measurements z := {zm }M m=1 and matrices {Hm ∈
modeled by a single-hidden-layer FNN [6], [7]. Unfortunately, 2N ×2N M
R }m=1 obeying the following physical model
the number of FNN parameters grows linearly with the length
of the input sequences, discouraging FNNs from capturing zm = v Hm v + m , ∀m = 1, . . . , M (1)
long-term dependencies in voltage time series.
Our contribution towards real-time and accurate monitor- our goal is to recover v ∈ R2N , where {m }M m=1 account for
ing of the smart power grids is two-fold. First, we advocate the measurement noise and modeling inaccuracies.
model-specific DNNs for PSSE, by unrolling a recently pro- Adopting the LAV error criterion that is known to be robust
posed prox-linear SE solver [26], [27]. Toward this goal, we first to outliers, the ensuing LAV estimate is sought (see e.g., [30])
develop a reduced-complexity prox-linear solver. Subsequently, M
1   
we demonstrate how the unrolled prox-linear solver lends itself v̂ := arg min zm − v Hm v (2)
to building blocks (layers) in contemporary DNNs. In contrast v∈R2N M m=1
to ‘plain-vanilla’ FNNs, our novel prox-linear nets require min-
imal tuning efforts, and come naturally with ‘skip-connections,’ for which various solvers have been developed [2], [16]. In par-
a critical element in modern successful DNN architectures (see ticular, the recent prox-linear solver developed in [26] has well
[13]) that enables efficient training. Moreover, to enhance system documented merits, including provably fast (locally quadratic)
observability as well as enable system awareness ahead of time, convergence, as well as efficiency in coping with the non-
we advocate power system state forecasting via deep recurrent smoothness and nonconvexity in (2). Specifically, the prox-
NNs (RNNs). Deep RNNs enjoy a fixed number of parameters linear solver starts with some initial vector v0 , and iteratively
even with variable-length input sequences, and they are able to minimizes the regularized and ‘locally linearized’ (relaxed) cost
capture complex nonlinear dependencies present in time series in (2), to obtain iteratively (see also [4])
data. Finally, we present numerical tests using real load data on M
the IEEE 57- and 118-bus benchmark systems to corroborate the vi+1 = arg min z − Ji (2v − vi )1 + v − vi 22
v∈R2N 2μi
merits of the developed deep prox-linear nets and RNNs relative (3)
to existing alternatives. where i ∈ N is the iteration index, Ji := [vi Hm ]1≤m≤M is an
The remainder of this paper is organized as follows. Section II M × N matrix whose m-th row is vi Hm , and {μi > 0} is a
outlines the basics of PSSE. Section III introduces our novel pre-selected step-size sequence.
reduced-complexity prox-linear solver for PSSE, and advocates It is clear that the per-iteration subproblem (3) is a convex
the prox-linear net. Section IV deals with deep RNN for state quadratic program, which can be solved by means of standard
forecasting, as well as shows how state forecasting can aid in turn convex programming methods. One possible iterative solver of
DNN-based PSSE. Simulated tests are presented in Section V, (3) is based on the alternating direction method of multipliers
while the paper is concluded with the research outlook in (ADMM). Such an ADMM-based inner loop turns out to en-
Section VI. tail 2M + 2N auxiliary variables, thus requiring the update of
2M + 4N variables per iteration [26].
II. LEAST-ABSOLUTE-VALUE ESTIMATION Aiming at a reduced-complexity solver, in the next section
Consider a power network consisting of N buses that can be we will first recast (3) in a Lasso-type form, and subsequently
modeled as a graph G := {N , L}, where N := {1, 2, . . . , N } unroll the resultant double-loop prox-linear iterations (3) that
comprises all buses, and L := {(n, n )} ∈ N × N collects all constitute the key blocks of our DNN-based PSSE solver.
lines; see Figure 1. For each bus n ∈ N , let Vn := vnr + jvni
denote its corresponding complex voltage with magnitude |Vn |, III. THE PROX-LINEAR NET
and Pn (Qn ) denote the (re)active power injection. For each In this section, we will develop a DNN-based scheme to
f f
line (n, n ) ∈ L, let Pnn  (Qnn ) denote the (re)active power approximate the solution of (2) by unrolling the double-loop
t t
flow seen at the ‘forwarding’ end, and Pnn  (Qnn ) denote the prox-linear iterations. Upon defining the vector variable ui :=
(re)active power flow at the ‘terminal’ end. Ji (2v − vi ) − z, and plugging ui into (3), we arrive at
To perform PSSE, suppose Mt system variables are mea-
sured at time instant t. For a compact representation, let zt := M
f u∗i = arg min ui 1 + Bi (ui + z) − vi 22 (4)
[{|Vn,t |2 }n∈Nto , {Pn,t }n∈Nto , {Qn,t }n∈Nto , {Pnn  ,t }(n,n )∈E o ,
t
ui ∈RM 4μi

Authorized licensed use limited to: University of Saskatchewan. Downloaded on September 14,2021 at 03:51:12 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: REAL-TIME POWER SYSTEM STATE ESTIMATION AND FORECASTING VIA DEEP UNROLLED NEURAL NETWORKS 4071

Instead of solving the optimization problem with a (large)

Algorithm 1: Reduced-Complexity Prox-Linear Solver.
number of iterations, recent proposals (e.g., [10], [31]) advo-
Input: Data {(zm , Hm )}M m=1 , step sizes {μi }, η, and cated trainable DNNs constructed by unfolding and truncating
initialization v0 = 1, u00 = 0. those iterations, to obtain data-driven solutions. Specifically, ma-
1: for i = 0, 1, . . . , I do trices {Wik } and vectors {bki } are found by learning from past
2: Evaluate Wik , Ai , and bki according to (7). data via back-propagation rather than set according to (7). As
3: Initialize u0i . such, properly trained unrolled DNNs can achieve competitive
4: for k = 0, 1, . . . , K do performance even with few number of ‘iterations,’ i.e., layers;
5: Update uk+1 i using (6). see also [10], [29], [31]. In Section V, we will demonstrate that
6: end for our unrolled prox-linear net achieves a significant speedup over
7: Update vi+1 using (5). conventional optimization-based PSSE solvers as it only entails
8: end for few matrix-vector multiplications. In the following, we illustrate
how to unroll Alg. 1 to design DNNs for real-time PSSE.
Consider first unrolling the outer loop (3) up to, say the (I +
where Bi ∈ R2N ×M satisfying Bi Ji = I2N ×2N denotes the 1)-st iteration to obtain vI+1 . Leveraging the recursion (6), each
pseudo inverse of Ji . The pseudo inverse Bi exists because PSSE inner loop iteration i refines the initialization u0i = uK
i−1 to yield
requires M ≥ 2N to guarantee system observability. after K inner iterations uK i . Such an unrolling leads to a K(I +
Once the inner optimum variable u∗i is found, the next outer- 1)-layer structured DNN. Suppose that the sequence {vi }I+1 i=0
loop iterate vi+1 can be readily obtained as has converged, which means vI − vI+1  ≤  for some  > 0.
vi+1 = [Bi (u∗i + z) + vi ]/2 (5) It can then be deduced that vI+1 = BuI u∗I + BzI z with BuI :=
BI and BzI := BI .
following the definition of ui . Interestingly, (4) now reduces Our novel DNN architecture, that builds on the physics-based
to a Lasso problem [20], for which various celebrated solvers Alg. 1, is thus a hybrid combining plain-vanilla FNNs with the
have been put forward, including e.g., the iterative shrinkage and conventional iterative solver such as Alg. 1. We will hence-
thresholding algorithm (ISTA) [20]. forth term it ‘prox-linear net.’ For illustration, the prox-linear
Specifically, with k denoting the iteration index of the inner- net with K = 3 is depicted in Fig. 2. The first inner loop i = 0
loop, the ISTA for (4) proceeds across iterations k = 0, 1, . . . is highlighted in a dashed box, where u10 = Sη (A0 z + b0 ) be-
along the following recursion cause u00 = 0. As with [10], [29], [31], our prox-linear net can

ηM    k   be treated as a trainable regressor that predicts v from data
uk+1 = S k
η ui − Bi Bi u i + z − v i z, where the coefficients {bki }1≤k≤3 I k 1≤k≤3
0≤i≤I , {Ai }i=0 , {Wi }0≤i≤I ,
i
2μi
u z
 k k  BI , and BI are typically untied to enhance approximation ca-
= Sη Wi ui + Ai z + bki (6) pability and learning flexibility. Given historical and/or simu-
where η > 0 is a fixed step size with coefficients lated measurement-voltage training pairs {(zs , vs )}, these co-
efficients can be learned end-to-end using backpropagation [23],
ηM 
Wik := I − B Bi , ∀k ∈ N (7a) possibly employing the Huber loss [15] to endow the state esti-
2μi i mates with resilience to outliers.
ηM  Relative to the conventional FNN in Fig. 3, our proposed
Ai := − B Bi (7b)
2μi i prox-linear net features: i) ‘skip-connections’ (the bluish lines
in Fig. 2) that directly connect the input z to intermediate/output
ηM 
bki := B vi , ∀k ∈ N (7c) layers, and ii) a fixed number (M in this case) of hidden neurons
2μi i per layer. It has been shown both analytically and empirically
and Sη (·) is the so-termed soft thresholding operator that these skip-connections help avoid the so-termed ‘vanishing’
⎧ and ‘exploding’ gradient issues, therefore enabling successful
⎨ x − η, x > η and efficient training of DNNs [13], [14]. The ‘skip-connections’
Sη (x) := 0, −η ≤ x ≤ η (8)
⎩ is also a key enabler of the universal approximation capability
x + η, x < −η
of DNNs with a fixed number of hidden-neurons per-layer [18].
understood entry-wise when applied to a vector. With regards to The only hyper-parameters that must be tuned in our prox-
initialization, one can set u00 = 0 without loss of generality, and linear net are I, K, and η, which are also tuning parameters
u0i = u∗i−1 for i ≥ 1. required by the iterative optimization solver in Alg. 1. It is also
A reduced-complexity implementation of the prox-linear worth pointing out that other than the soft-thresholding nonlin-
PSSE solver is summarized in Alg. 1. With appropriate step sizes earity (a.k.a. activation function) used in Figs. 2 and 3, alter-
{μi } and η, the sequence {vi } generated by Alg. 1 converges to native functions such as the rectified linear unit (ReLU) can be
a stationary point of (2) [26]. In practice, Alg. 1 often requires a applied as well [10], [25]. We have observed in our simulated
large number K and I iterations to approximate the solution of tests that the prox-linear net with soft thresholding operators or
(2). Furthermore, the pseudo-inverse Bi has to be computed per ReLUs yield similar performance. To understand how different
outer-loop iteration. These challenges can discourages its use in network architectures affect the performance, ReLU activation
real-time applications. functions are used by default unless otherwise stated.

Authorized licensed use limited to: University of Saskatchewan. Downloaded on September 14,2021 at 03:51:12 UTC from IEEE Xplore. Restrictions apply.
4072 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 15, AUGUST 1, 2019

Fig. 2. Prox-linear net with K = 3 blocks.

Fig. 3. Plain-vanilla FNN which has the same per-layer number of hidden units as the prox-linear net.

the measurement function that summarizes equations in (1) at

time slot t + 1. To perform state forecasting, function φ must
be estimated or approximated – a task that we will accomplish
using RNN modeling, as we present next.
RNNs are NN models designed to learn from correlated time
series data. Relative to FNNs, RNNs are not only scalable to
long-memory inputs (regressors) that entail sequences of large
r, but are also capable of processing input sequences of variable
length [9]. Given the input sequence {vτ }tτ =t−r+1 , and an ini-
Fig. 4. Deep prox-linear net based real-time PSSE. tial state st−r , an RNN finds the hidden state1 vector sequence
{sτ }tτ =t−r+1 by repeating
The flow chart demonstrating the prox-linear net for real-time sτ = f (R0 vτ + Rss sτ −1 + r0 ) (11)
PSSE is depicted in Fig. 4, where the real-time inference stage
where f (·) is a nonlinear activation function (e.g., a ReLU or
is described in the left rounded rectangular box, while the off-
sigmoid unit), understood entry-wise when applied to a vector,
line training stage is on the left. Thanks to the wedding of the
whereas the coefficient matrices R0 , Rss , and the vector r0
physics in (1) with our DNN architecture design, the extensive
contain time-invariant weights.
numerical tests in Section V will confirm an impressive boost in
Deep RNNs are RNNs of multiple (≥3) processing layers,
performance of our prox-linear nets relative to competing FNN
which can learn compact representations of time series through
and Gauss-Newton based PSSE approaches.
hierarchical nonlinear transformations. The state-of-the-art in
numerous sequence processing applications, including music
IV. DEEP RNNS FOR STATE FORECASTING prediction and machine translation [9], has been significantly
Per time slot t, the PSSE scheme we developed in improved with deep RNN models. By stacking up multiple re-
Section III estimates the state vector vt ∈ R2N upon receiving current hidden layers (cf. (11)) one on top of another, deep RNNs
measurements zt . Nevertheless, its performance is challenged can be constructed as follows [21]
when there are missing entries in zt , which is indeed common  
slτ = f Rl−1 sl−1
τ + Rss,l slτ −1 + rl−1 , l ≥ 1 (12)
in a SCADA system due for example to meter and/or communi-
cation failures. To enhance our novel PSSE scheme and obtain where l is the layer index, slτ denotes the hidden state of the l-th
system awareness even ahead of time, we are prompted to pursue layer at slot τ having s0τ := vτ , and {Rl , Rss,l , rl } collect all
power system state forecasting, which for a single step amounts unknown weights. Fig. 5 (left) depicts the computational graph
to predicting the next state vt+1 at time slot t + 1 from the avail- representing (12) for l = 2, with the bias vectors rl = 0, ∀l for
able time-series {vτ }tτ =0 [6]. Analytically, the estimation and simplicity in depiction, and the black squares standing for single-
prediction steps are as follows step delay units. Unfolding the graph by breaking the loops and
connecting the arrows to corresponding units of the next time
vt+1 = φ(vt , vt−1 , vt−2 , . . . , vt−r+1 ) + ξ t (9) slot, leads to a deep RNN in Fig. 5 (right), whose rows represent
zt+1 = ht+1 (vt+1 ) + t+1 (10) layers, and columns denote time slots.
The RNN output can come in various forms, including one
where {ξ t , t+1 } account for modeling inaccuracies; the tun- output per time step, or, one output after several steps. The latter
able parameter r ≥ 1 represents the number of lagged (present
included) states used to predict vt+1 ; and the unknown (non- 1 Hidden state is an auxiliary set of vector variables not to be confused with
linear) function φ captures the state transition, while ht+1 (·) is the power system state v consisting of the nodal voltages as in (1).

Authorized licensed use limited to: University of Saskatchewan. Downloaded on September 14,2021 at 03:51:12 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: REAL-TIME POWER SYSTEM STATE ESTIMATION AND FORECASTING VIA DEEP UNROLLED NEURAL NETWORKS 4073

of the present paper, it is worth remarking that the residuals

zt+1 − ht+1 (v̂t+1 ) along with zt+1 − žt+1 can be used to un-
veil erroneous data, as well as changes in the grid topology and
the link parameters; see [6] for an overview.

V. NUMERICAL TESTS
Performance of our deep prox-linear net based PSSE, and
deep RNN based state forecasting methods was evaluated using
the IEEE 57- and 118-bus benchmark systems. Real load data
Fig. 5. An unfolded deep RNN with no outputs. from the 2012 Global Energy Forecasting Competition (GEFC)2
were used to generate the training and testing datasets, where
the load series were subsampled for size reduction by a factor
of 5 (2) for the IEEE 57-bus (118-bus) system. Subsequently,
the resultant load instances were normalized to match the scale
of power demands in the simulated system. The MATPOWER
toolbox [38] was used to solve the AC power flow equations
with the normalized load series as inputs, to obtain the ground-
Fig. 6. DNN-based real-time power system monitoring. truth voltages {vτ }, and produce measurements {zτ } that com-
prise all forwarding-end active (reactive) power flows, as well
as all voltage magnitudes. All NNs were trained using ‘Ten-
matches the rth-order nonlinear regression in (9) when approx-
sorFlow’ [1] on an NVIDIA Titan X GPU with 12 GB RAM,
imating φ with a deep RNN. Concretely, the output of our deep
with weights learned by the backpropagation based algorithm
RNN is given by
‘Adam’ (with starting learning rate 10−3 ) for 200 epochs. To
v̌t+1 = Rout slt + rout (13) alleviate randomness in the obtained weights introduced by the
training algorithms, all NNs were trained and tested indepen-
where v̌t+1 is the forecast of vt+1 at time t, and (Rout , rout )
dently for 20 times, with reported results averaged over 20 runs.
contain weights of the output layer. Given historical voltage
For reproducibility, the ‘Python’-based implementation of our
time series, the weights (Rout , rout ) and {Rl , Rss,l , rl } can be
prox-linear net for PSSE of the 118-bus system is publicly avail-
learned end-to-end using backpropagation [9]. Invoking RNNs
able at https://github.com/LiangZhangUMN/PSSE-via-DNNs.
for state-space models, the class of nonlinear predictors dis-
cussed in [6] is considerably broadened here to have memory. As
A. Prox-Linear Nets for PSSE
will be demonstrated through extensive numerical tests, the fore-
casting performance can be significantly improved through the To start, the prox-linear net based PSSE was tested, which es-
use of deep RNNs. Although the focus here is on one-step state timates {v̂τ } using {zτ }. For both training and testing phases, all
forecasting, it is worth stressing that our proposed approaches measurements {zτ } were corrupted by additive white Gaussian
with minor modifications, can be generalized to predict the sys- noise, where the standard deviation for power flows and for volt-
tem states multiple steps ahead. age magnitudes was 0.02 and 0.01. The estimation performance
So far, we have elaborated on how RNNs enable flexible non- of our prox-linear net was assessed in terms of the normalized
linear predictors for power system state forecasting. To predict root mean-square error (RMSE) v̂ − v2 /N , where v is the
v̌t+1 at time slot t, the RNN in (12) requires ground-truth volt- ground truth, and v̂ the estimate obtained by the prox-linear net.
ages {vτ }tτ =t−r+1 (cf. (9)), which however, may not be available In particular, the prox-linear net was simulated with T = 2
in practice. Instead we can use the estimated ones {v̂τ }tτ =t−r+1 and K = 3. The ‘workhorse’ Gauss-Newton method, a 6-layer
provided by our prox-linear net-based estimator in Section III. ‘plain-vanilla’ FNN that has the same depth as our prox-linear
In turn, the forecast v̌t+1 can be employed as a prior to aid PSSE net, and an 8-layer ‘plain-vanilla’ FNN that has roughly the same
at time slot t + 1, by providing the so-termed virtual measure- number of parameters as the prox-linear net, were simulated
ments žt+1 := ht+1 (v̌t+1 ) that can be readily accounted for in as baselines. The number of hidden units per layer in all NNs
(2). For example, when there are missing entries in zt+1 , the was kept equal to the dimension of the input, that is, 57 × 2 =
obtained žt+1 can be used to improve the PSSE performance by 114 for the 57-bus system and 118 × 2 = 236 for the 118-bus
imputing the missing values. system.
Figure 6 depicts the flow chart of the overall real-time power In the first experiment using the 57-bus system, a total of
system monitoring scheme, consisting of deep prox-linear net- 7, 676 measurement-voltage (zτ , vτ ) pairs were generated, out
based PSSE (cf. Fig. 1) and deep RNN-based state forecasting of which the first 6,176 pairs were used for training, and the
(cf. Fig. 4) modules, that are implemented at time t and t + 1. rest were kept for testing. The average performance over 20
Throughout, RNNs with l = 3 and r = 10 were used in our trials, was evaluated in terms of the average RMSEs over the
experiments. Our novel scheme is reminiscent of the predictor-
corrector-type estimators emerging with dynamic state estima- 2 https://www.kaggle.com/c/global-energy-forecasting-competition-2012-
tion problems using Kalman filters. Although beyond the scope load-forecasting/data

Authorized licensed use limited to: University of Saskatchewan. Downloaded on September 14,2021 at 03:51:12 UTC from IEEE Xplore. Restrictions apply.
4074 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 15, AUGUST 1, 2019

Fig. 7. Estimation errors in voltage magnitudes and angles of bus 10 of the Fig. 9. Estimation errors in voltage magnitudes and angles of all the 57 buses
57-bus system from test instances 100 to 120. of the 57-bus system at test instance 120.

NNs for all buses on test instance 120 are depicted in Fig. 9.
Evidently, our prox-linear net based PSSE performs the best in
all cases.
The second experiment tests our prox-linear net using the
IEEE 118-bus system, where 18, 528 voltage-measurement
pairs were simulated, with 14, 822 pairs employed for train-
ing and 3, 706 kept for testing. The average RMSEs over 3,
706 testing examples for the prox-linear net, Gauss-Newton, 6-
layer FNN and 8-layer FNN, are 2.97 × 10−4 , 4.71 × 10−2 ,
1.645 × 10−3 , and 2.366 × 10−3 , respectively. Clearly, our
prox-linear net yields markedly improved performance over
competing alternatives in this case (especially as the system
size grows large). The Gauss-Newton approach performs the
worst due to unbalanced grid parameters of this test system. In-
terestingly, it was frequently observed that the Gauss-Newton
iterations minimize the WLS objective function (resulting a
Fig. 8. Estimation errors in voltage magnitudes and angles of bus 27 of the loss smaller than 10−6 ), but converge to a stationary point
57-bus system from test instances 100 to 120.
that is far away from the simulated ground-truth voltage. This
is indeed due to the nonconvexity of the WLS function, for
1, 500 testing examples for the prox-linear net, Gauss-Newton, which multiple optimal solutions often exist. Depending crit-
6-layer FNN, and 8-layer FNN, are 3.49 × 10−4 , 3.2 × 10−4 , ically on initialization, traditional optimization based solvers
6.35 × 10−4 , and 9.02 × 10−4 , respectively. These numbers can unfortunately get stuck at any of those points. In sharp
showcase competitive performance of the prox-linear net. In- contrast, data-driven NN-based approaches nicely bypass this
terestingly, when the number of hidden layers of ‘plain-vanilla’ hurdle.
FNNs increases from 6 to 8, the performance degrades due partly In terms of runtime, the prox-linear net, Gauss-Newton,
to the difficulty in training the 8-layer FNN. 6-layer FNN, and 8-layer FNN, over 3, 706 testing examples are
As far as the computation time is concerned, the prox- 0.3323 s, 183.4 s, 0.2895 s, and 0.3315 s, resulting in an aver-
linear net, Gauss-Newton, 6-layer FNN, and 8-layer FNN over age per-instance estimation time of 6.5 × 10−5 s, 9.48 × 10−3 s,
1, 500 testing examples are 0.0973 s, 14.22 s, 0.0944 s, and 6.3 × 10−5 s, and 6.4 × 10−5 s, respectively. This corroborates
0.0954 s, resulting in an average per-instance estimation time of again the efficiency of NN-based approaches. The ground-truth
9.0 × 10−5 s, 4.9 × 10−2 s, 7.8 × 10−5 s, and 8.9 × 10−5 s, re- voltage along with estimates obtained by the prox-linear net,
spectively. These numbers corroborate the speedup advantage of 6-layer FNN, and 8-layer FNN, for bus 50 and bus 100 at test
NN-based PSSE over the traditional Gauss-Newton approach. instances 1, 000 to 1, 050, are depicted in Figs. 10 and 11, re-
The ground-truth voltages along with the estimates found by spectively. In addition, the actual voltages and their estimates for
the prox-linear net, 6-layer FNN, and 8-layer FNN for bus 10 the first fifty buses on test instance 1, 000 are depicted in Fig. 12.
and bus 27 from test instances 100 to 120, are shown in Figs. 7 In all cases, our prox-linear net yields markedly improved per-
and 8, respectively. The true voltages and the estimated ones by formance relative to competing alternatives.

Authorized licensed use limited to: University of Saskatchewan. Downloaded on September 14,2021 at 03:51:12 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: REAL-TIME POWER SYSTEM STATE ESTIMATION AND FORECASTING VIA DEEP UNROLLED NEURAL NETWORKS 4075

Fig. 10. Estimation errors in voltage magnitudes and angles of bus 50 of the Fig. 13. Forecasting errors in voltage magnitudes and angles of bus 30 of the
118-bus system from instances 1, 000 to 1, 050. 57-bus system from test instances 100 to 120.

B. Deep RNNs for State Forecating

This section examines our RNN based power system state
forecasting scheme. The forecasting performance was evaluated
in terms of the normalized RMSE v̌ − v2 /N of the forecast
v̌ relative to the ground truth v.
Specifically, deep RNNs with l = 3, r = 10, batch size 32,
and ReLU activation functions were trained and tested on the
ground-truth voltage time series, and on the estimated volt-
age time series from the prox-linear net. We will refer to the
latter as ‘RNNs with estimated voltages’ hereafter. The num-
ber of hidden units per layer in RNNs was kept the same
as the input dimension, namely 57 × 2 = 114 for the 57-bus
system, and 118 × 2 = 236 for the 118-bus system. For com-
parison, a single-hidden-layer FNN (2-layer FNN) [7], and a
VAR(1) model [12] based state forecasting approaches were
adopted as benchmarks. The average RMSEs over 20 Monte
Fig. 11. Estimation errors in voltage magnitudes and angles of bus 100 of the
118-bus system from instances 1, 000 to 1, 050.
Carlo runs for the RNN, RNN with estimated voltages, 2-layer
FNN, and VAR(1) are respectively 2.303 × 10−3 , 2.305 × 10−3 ,
3.153 × 10−3 , and 6.772 × 10−3 for the 57-bus system, as well
as 2.588 × 10−3 , 2.751 × 10−3 , 4.249 × 10−3 , 6.461 × 10−3
for the 118-bus system. These numbers demonstrate that our
deep RNN with estimated voltages offers comparable forecast-
ing performance relative to that with ground-truth voltages. Al-
though both FNN and VAR(1) were trained and tested using
ground-truth voltage time-series, they perform even worse than
our RNN trained with estimated voltages.
The true voltages and their forecasts provided by the deep
RNN, RNN with estimated voltages, 2-layer FNN, and VAR(1)
for bus 30 of the 57-bus system from test instances 100 to 120, as
well as all buses on test instance 100, are reported in Figs. 13 and
14, accordingly. The ground-truth and forecast voltages for the
first 50 buses of the 118-bus system on test instance 1, 000 are
depicted in Fig. 15. Curves illustrate that our deep RNN based
approaches perform the best in all cases.
The final experiment tests the efficacy of our real-time DSSE
Fig. 12. Estimation errors in voltage magnitudes and angles of the first 50 scheme on the 57-bus system assuming a fraction of measure-
buses of the 118-bus system at test instance 1, 000. ments are provided via deep RNN forecasts. For a fixed number

Authorized licensed use limited to: University of Saskatchewan. Downloaded on September 14,2021 at 03:51:12 UTC from IEEE Xplore. Restrictions apply.
4076 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 15, AUGUST 1, 2019

of pseudo measurements, 50 independent trials were executed. In

each trial, locations of the pseudo measurements were sampled
uniformly at random. Fig. 16 presents the sample mean and stan-
dard derivation of RMSE averaged over the 50 trials for different
numbers of pseudo measurements. Two pseudo measurements
generation schemes were compared: i) pseudo measurements
postulated with a deep RNN and ii) pseudo measurement values
set the same with their values at time t − 1. It is observed that
the DSSE with RNN postulated pseudo measurements offers
improved performance in all cases.

VI. CONCLUSIONS
This paper dealt with real-time power system monitoring (es-
timation and forecasting) by building on data-driven DNN ad-
vances. Prox-linear nets were developed for PSSE, that combine
NNs with traditional physics-based optimization approaches.
Fig. 14. Forecasting errors in voltage magnitudes and angles of all the 57
buses of the 57-bus system at test instance 100.
Deep RNNs were also introduced for power system state fore-
casting from historical (estimated) voltages. Our model-specific
prox-linear net based PSSE is easy-to-train, and computationally
inexpensive. The proposed RNN-based forecasting accounts for
the long-term nonlinear dependencies in the voltage time-series,
enhances PSSE, and offers situational awareness ahead of time.
Numerical tests on the IEEE 57- and 118-bus benchmark sys-
tems using real load data illustrate the merits of our developed
approaches relative to existing alternatives.
Our current and future research agenda includes specializing
the DNN-based estimation and forecasting schemes to distribu-
tion networks. Our agenda also includes ‘on-the-fly’ RNN-based
algorithms to account for dynamically changing environments,
and corresponding time dependencies.

REFERENCES
[1] M. Abadi et al., “TensorFlow: Large-scale machine learning on hetero-
geneous systems,” 2015. [Online]. Available: https://www.tensorflow.org/
[2] A. Abur and M. K. Celik, “A fast algorithm for the weighted least-absolute-
value state estimation (for power systems),” IEEE Trans. Power Syst.,
vol. 6, no. 1, pp. 1–8, Feb. 1991.
Fig. 15. Forecasting errors in voltage magnitudes and angles of the first 50 [3] P. N. P. Barbeiro, J. Krstulovic, H. Teixeira, J. Pereira, F. J. Soares, and J.
buses of the 118-bus system at instance 1, 000. P. Iria, “State estimation in distribution smart grids using autoencoders,”
in Proc. IEEE 8th Int. Power Eng. Optim. Conf., Shah Alam, Malaysia,
Mar. 2014, pp. 358–363.
[4] J. V. Burke and M. C. Ferris, “A Gauss-Newton method for convex compos-
ite optimization,” Math. Program., vol. 71, no. 2, pp. 179–194, Dec. 1995.
[5] A. S. Debs and R. E. Larson, “A dynamic estimator for tracking the state
of a power system,” IEEE Trans. Power App. Syst., vol. PAS-89, no. 7,
pp. 1670–1678, Sep. 1970.
[6] M. B. Do Coutto Filho and J. C. Stacchini de Souza, “Forecasting-aided
state estimation–Part I: Panorama,” IEEE Trans. Power Syst., vol. 24, no. 4,
pp. 1667–1677, Nov. 2009.
[7] M. B. Do Coutto Filho, J. C. Stacchini de Souza, and R. S. Freund,
“Forecasting-aided state estimation–Part II: Implementation,” IEEE Trans.
Power Syst., vol. 24, no. 4, pp. 1678–1685, Nov. 2009.
[8] G. B. Giannakis, V. Kekatos, N. Gatsis, S.-J. Kim, H. Zhu, and B. Wollen-
berg, “Monitoring and optimization for power grids: A signal processing
perspective,” IEEE Signal Process. Mag., vol. 30, no. 5, pp. 107–128,
Sep. 2013.
[9] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning.
Cambridge, MA, USA: MIT Press, 2016. [Online]. Available:
http://www.deeplearningbook.org
[10] K. Gregor and Y. LeCun, “Learning fast approximations of sparse coding,”
in Proc. 27th Int. Conf. Mach. Learn, Haifa, Israel, Jun. 2010, pp. 399–406.
[11] M. Hassanzadeh and C. Y. Evrenosoğlu, “Power system state forecasting
using regression analysis,” in Proc. IEEE Power Energy Soc. Gen. Meeting,
Fig. 16. Average RMSE w.r.t. the number of predicted values. San Diego, CA, USA, Jul. 2012, pp. 1–6.

Authorized licensed use limited to: University of Saskatchewan. Downloaded on September 14,2021 at 03:51:12 UTC from IEEE Xplore. Restrictions apply.
ZHANG et al.: REAL-TIME POWER SYSTEM STATE ESTIMATION AND FORECASTING VIA DEEP UNROLLED NEURAL NETWORKS 4077

[12] M. Hassanzadeh, C. Y. Evrenosoğlu, and L. Mili, “A short-term nodal [36] Y. Zhao, J. Chen, and H. V. Poor, “A learning-to-infer method for real-time
voltage phasor forecasting method using temporal and spatial correlation,” power grid multi-line outage identification,” 2018, arXiv:1710.07818.
IEEE Trans. Power Syst., vol. 31, no. 5, pp. 3881–3890, Sep. 2016. [37] R. D. Zimmerman, C. E. Murillo-Sanchez, and R. J. Thomas, “MAT-
[13] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image POWER: Steady-state operations, planning and analysis tools for power
recognition,” in Proc. Conf. Comput. Vision Pattern Recognit., Las Vegas, systems research and education,” IEEE Trans. Power Syst., vol. 26, no. 1,
NV, USA, 2016, pp. 770–778. pp. 12–19, Feb. 2011.
[14] K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual
networks,” in Proc. Eur. Conf. Comput. Vision. Amsterdam, The Nether-
lands: Springer, Oct. 2016, pp. 630–645.
[15] P. J. Huber, “Robust Statistics,” in Proc. Int. Encyclopedia Statistical Sci.,
Springer, 2011, pp. 1248–1251. Liang Zhang (S’13) received the B.Sc. and M.Sc.
[16] R. Jabr and B. Pal, “Iteratively reweighted least-squares implementation degrees in electrical engineering from Shanghai
of the WLAV state-estimation method,” IET Gener. Transmiss. Distrib., Jiao Tong University, Shanghai, China, in 2012 and
vol. 151, no. 1, pp. 103–108, Feb. 2004. 2014, respectively, and the Ph.D. degree in electri-
[17] A. M. Leite da Silva, M. B. Do Coutto Filho, and J. F. De Queiroz, “State cal and computer engineering from the University of
forecasting in electric power systems,” IEE Proc. C. Gener., Transmiss. Minnesota, Minneapolis, MN, USA, in 2019. Since
Distrib., vol. 130, no. 5, pp. 237–244, Sep. 1983. February 2019, he has been working on platforms
[18] H. Lin and S. Jegelka, “ResNet with one-neuron hidden layers is a universal applied machine intelligence with Google, Mountain
approximator,” in Proc. Adv. Neural Inf. Process. Syst., Montreal, CA, View, CA, USA. His research interests include the
Dec. 3–8, 2018, pp. 6169–6178. areas of large-scale optimization and learning.
[19] K. R. Mestav, J. Luengo-Rozas, and L. Tong, “State estimation for unob-
servable distribution systems via deep neural networks,” in IEEE Power
Energy Society General Meeting, Portland, OR, USA, Aug. 5–9, 2018,
Gang Wang (M’18) received the B.Eng. degree
pp. 1–5.
in electrical engineering and automation from the
[20] N. Parikh and S. Boyd, “Proximal algorithms,” Found. Trends Optim.,
Beijing Institute of Technology, Beijing, China, in
vol. 1, no. 3, pp. 127–239, 2014. 2011, and the Ph.D. degree in electrical and com-
[21] R. Pascanu, C. Gulcehre, K. Cho, and Y. Bengio, “How to construct deep
puter engineering from the University of Minnesota,
recurrent neural networks,” in Proc. Int. Conf. Learn. Representations,
Minneapolis, MN, USA, in 2018.
Banff, AB, Canada, Apr. 2014, pp. 1–10.
He is currently a Postdoctoral Associate with the
[22] W. S. Rosenthal, A. M. Tartakovsky, and Z. Huang, “Ensemble Kalman Department of Electrical and Computer Engineering,
filter for dynamic state estimation of power grids stochastically driven
University of Minnesota. His research interests focus
by time-correlated mechanical input power,” IEEE Trans. Power Syst.,
on the areas of statistical signal processing, optimiza-
vol. 33, no. 4, pp. 3701–3710, Jul. 2018.
tion, and deep learning with applications to data sci-
[23] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning repre- ence and smart grids. He was the recipient of a number of awards, including
sentations by back-propagating errors,” Nature, vol. 323, pp. 533–536,
paper awards at the 2017 European Signal Processing Conference and the 2019
Oct. 1986.
IEEE Power and Energy Society General Meeting.
[24] N. Samuel, T. Diskin, and A. Wiesel, “Deep MIMO detection,” in
Proc. IEEE 18th Int. Workshop Signal Process. Adv. Wireless Commun.,
Hokkaido, Japan, Jul. 2017, pp. 1–5.
[25] G. Wang, G. B. Giannakis, and J. Chen, “Learning ReLU networks on Georgios B. Giannakis (F’97) received the Diploma
linearly separable data: Algorithm, optimality, and generalization,” IEEE in electrical engineering from the National Technical
Trans. Signal Process., vol. 67, no. 9, pp. 2357–2370, May 2019. University of Athens, Athens, Greece, in 1981, and
[26] G. Wang, G. B. Giannakis, and J. Chen, “Robust and scalable power system the M.Sc. degree in electrical engineering, the M.Sc.
state estimation via composite optimization,” IEEE Trans. Smart Grid, to degree in mathematics, and the Ph.D. degree in elec-
be published, doi: 10.1109/TSG.2019.2897100. trical engineering from the University of Southern
[27] G. Wang, H. Zhu, G. B. Giannakis, and J. Sun, “Robust power system California, Los Angeles, CA, USA, in 1983, 1986,
state estimation from rank-one measurements,” IEEE Trans. Control Netw. and 1986, respectively.
Syst., to be published, doi:10.1109/TCNS.2019.2890954. From 1982 to 1986, he was with the University
[28] G. Wang, G. B. Giannakis, J. Chen, and J. Sun, “Distribution system state of Southern California. From 1987 to 1998, he was a
estimation: An overview of recent developments,” Frontier Inf. Technol. faculty member with the University of Virginia. Since
Electron. Eng., vol. 20, no. 1, pp. 4–17, Jan. 2019. 1999, he has been a Professor with the University of Minnesota, Minneapolis,
[29] Z. Wang, Q. Ling, and T. Huang, “Learning deep 0 encoders,” in Proc. MN, USA. He is the (co-) inventor of 32 patents issued. His research inter-
AAAI Conf. Artif. Intell., Phoenix, AZ, USA, Feb. 2016, pp. 2194–2200. ests include the areas of statistical learning, communications, and networking—
[30] A. J. Wood and B. F. Wollenberg, Power Generation, Operation, and Con- subjects on which he has authored or coauthored more than 440 journal papers,
trol, 2nd ed. New York, NY, USA: Wiley, 1996. 730 conference papers, 25 book chapters, 2 edited books, and 2 research mono-
[31] Y. Yang, J. Sun, H. Li, and Z. Xu, “Deep ADMM-net for compressive graphs (h-index 140). His current research focuses on data science, and network
sensing MRI,” in Proc. Adv. Neural Inf. Process. Syst., Barcelona, Spain, science with applications to the Internet of Things, social, brain, and power
Dec. 2016, pp. 10–18. networks with renewables. He is the (co-) recipient of nine best journal paper
[32] A. S. Zamzam, X. Fu, and N. D. Sidiropoulos, “Data-driven learning- awards from the IEEE Signal Processing (SP) and Communications Societies,
based optimization for distribution system state estimation,” 2019, including the G. Marconi Prize Paper Award in Wireless Communications, and
arXiv:1807.01671. the recipient of Technical Achievement Awards from the SP Society in 2000,
[33] A. S. Zamzam and N. D. Sidiropoulos, “Physics-aware neural networks from EURASIP in 2005, a Young Faculty Teaching Award, the G. W. Taylor
for distribution system state estimation,” 2019, arXiv:1903.09669. Award for Distinguished Research from the University of Minnesota, and the
[34] L. Zhang, V. Kekatos, and G. B. Giannakis, “Scalable electric vehicle IEEE Fourier Technical Field Award (inaugural recipient in 2015). He holds an
charging protocols,” IEEE Trans. Power Syst., vol. 32, no. 2, pp. 1451– ADC Endowed Chair, a University of Minnesota McKnight Presidential Chair
1462, Mar. 2017. in ECE, and is currently the Director of the Digital Technology Center with
[35] L. Zhang, G. Wang, and G. B. Giannakis, “Real-time power system state the University of Minnesota. He is a Fellow of EURASIP, and has served the
estimation via deep unrolled neural networks,” in Proc. Global Conf. Sig- IEEE in a number of posts, including that of a Distinguished Lecturer for the
nal Inf. Process., Anaheim, CA, USA, Nov. 26–28, 2018, pp. 907–911. IEEE-SPS.

Authorized licensed use limited to: University of Saskatchewan. Downloaded on September 14,2021 at 03:51:12 UTC from IEEE Xplore. Restrictions apply.