,(((,QWHUQDWLRQDO&RQIHUHQFHRQ&RPPXQLFDWLRQV&RQWURODQG&RPSXWLQJ7HFKQRORJLHVIRU6PDUW*ULGV 6PDUW*ULG&RPP

State Estimation in Smart Distribution Systems

with Deep Generative Adversary Networks
Kursat Rasim Mestav† and Lang Tong†

Abstract— The problem of distribution system state estima- hourly intervals, that is incompatible with the more rapid
tion using smart meters and limited SCADA (Supervisory Con- changes of DER. Unfortunately, existing techniques rarely
trol and Data Acquisition) measurement units is considered. To address the mismatch of measurement resolution among the
overcome the lack of measurements, a Bayesian state estimator
using deep learning is proposed. The proposed method consists slow timescale smart meter data, the fast timescale real-
of two steps. First, a deep generative adversary network is time measurements (e.g., current magnitudes at feeders and
trained to learn the distribution of net power injections at substations), and the need for fast timescale state estimation.
the loads. Then, a deep regression network is trained using State estimation for unobservable systems must incorpo-
the samples generated from the generative network to obtain rate additional properties beyond the measurement model
minimum mean-squared error (MMSE) estimate of the system
state. Our simulation results show the accuracy and the online defined by the power flow equations. To this end, we
computation cost of the proposed method are superior to the pursue a Bayesian inference approach where the system
conventional methods. states (voltage phasors) and measurements are modeled as
Index Terms— Distribution system state estimation, deep random variables endowed with (unknown) joint probability
learning, generative adversary networks, deep regression net- distributions. Given the highly stochastic nature of the re-
work, SCADA, smart meter, and Bayesian inference.
newable injections, such a Bayesian model is both natural
I. I NTRODUCTION and appropriate.
The problem of state estimation is considered for dis- The most important benefit of Bayesian inference is that
tribution systems. A major obstacle to state estimation in observability is no longer required. A Bayesian estimator
distribution systems is that such systems are nominally exploits probabilistic dependencies of the measurement vari-
unobservable [1], [2]. By unobservable it means that there ables on the system states; it improves the prior distribution
is a manifold of uncountably many states that correspond to of the states using available measurements, even if there are
the same measurement. System unobservability arises when only a few such measurements. Unlike the least squares tech-
the number of sensors is not sufficiently large—typical in niques that minimize modeling error, a Bayesian estimator
distribution systems—or sensors are not well placed in the minimizes directly the estimation error.
network. An observable system may become unobservable The advantage of Bayesian inference, however, comes
when sensors are at fault, sensor data missing, or data with significant implementation issues. First, the underlying
tampered by malicious agents [3]. joint distribution of the system states and measurements is
The popular weighted least-squares (WLS) estimator and unknown, and some type of learning is necessary. Second,
its variants can no longer be used when the system is even if the relevant probability distribution is known or can
unobservable because a small WLS error in model fitting be estimated, computing the actual state estimate is often
does not imply a small error in estimation; a large estimation intractable analytically and prohibitive computationally.
error may persist even in the absence of noise. A standard A. Summary of Results and Contributions
remedy of unobservability is to use the so-called pseudo
measurements based on interpolated observations or forecasts The main contribution of this work is an application
from historical data. Indeed, the use of pseudo measurements of deep learning technology for distribution system state
has been a dominant theme for distribution system state estimation when the system is unobservable by the deployed
estimation. These techniques, however, are ad hoc and do SCADAs. To this end, we develop a data-driven generative
not assure the quality of estimates. model coupled with a deep neural network that provides
The advent of smart meters and advanced metering in- SCADA timescale state estimates. As a major departure
frastructure provide new sources of measurements. Attempts of the predominantly pseudo-measurement approaches to
have been made to incorporate smart meter data for state state estimation when the power system is unobservable, the
estimation [4]–[6]. Not intended for state estimation, smart proposed solution to state estimation for the unobservable
meters measure accumulative consumptions. They often ar- systems takes a Bayesian inference perspective, assuming
rive at a much slower timescale, e.g., in 15-minute to the system states as random quantities. Consequently, the
proposed approach is not bound by the observability assump-
Kursat Rasim Mestav† , and Lang Tong† are with the School of Electrical tion as required by the traditional weighted least-squares
and Computer Engineering, Cornell University, Ithaca, NY, 14850 USA e- solutions.
mail: {krm264,lt35}@cornell.edu.
This work was supported in part by the National Science Foundation The Bayesian inference that minimizes the mean squared
under Awards 1809830 and 1932501. error (MSE) of the state estimate is given by the conditional

k,(((

Authorized licensed use limited to: University of Saskatchewan. Downloaded on October 04,2021 at 07:44:28 UTC from IEEE Xplore. Restrictions apply.
,(((,QWHUQDWLRQDO&RQIHUHQFHRQ&RPPXQLFDWLRQV&RQWURODQG&RPSXWLQJ7HFKQRORJLHVIRU6PDUW*ULGV 6PDUW*ULG&RPP

mean of the system state. Simple as it may appear, the although real-time measurements were used as optimization
conditional mean can be difficult to compute. Fundamen- constraints rather than as conditioning variables in Bayesian
tally, the underlying joint probability distribution of the inference. One approach to calculating conditional statistics
measurement and the system state is required. Even when is based on a graphical model of the distribution system from
such a probability distribution is given in closed-form, the which belief propagation techniques are used to generate
computation of the conditional mean is intractable in general. state estimates [21]. These techniques require a dependency
Furthermore, the lack of measurement-state samples makes graph of the system states and explicit forms of probability
it impossible to learn the joint distribution directly. distributions. Another approach is based on a linear approx-
The proposed deep learning approach consists of unsu- imation of the AC power flow [22].
pervised learning of the generative model of the network The approach presented in this paper belongs to the class
injection and supervised learning of the conditional mean of of Monte Carlo techniques in which samples are generated
the system states. Specifically, given direct or indirect mea- and empirical conditional means computed. In our approach,
surements of the net-injection, we consider several machine instead of using Monte Carlo sampling to calculating the con-
learning techniques for the underlying generative models of ditional mean directly as in [23], [24], Monte Carlo sampling
network injections. For distribution systems with smart meter is used to train a neural network that, in real-time, computes
measurements, such models can be learned from smart meter the MMSE estimate directly from the measurements.
measurements or estimates of network injections using para- The proposed technique builds on to our work on dis-
metric or nonparametric techniques, including deep learning tribution system state estimation [25] with several notable
techniques such as the generative adversary network (GAN) differences. First, the techniques of learning of generative
method. models are different. In the referred research it was assumed
The unsupervised learning of injection generative model is that the power injections follow a Gaussian mixture distri-
followed by supervised learning of the conditional mean of bution. Here we propose a more comprehensive technique
the network state. To this end, we exploit the physical model using generative adversarial networks (GANs). The GANs
of the power system by embedding the power flow equation are not only a more generic method to learn distributions
in generating training samples. A deep neural network with without having strong assumptions, it also allows us to learn
prewhitening first layer is proposed. We show that the pro- a distribution when the samples are not directly observable,
posed state estimator achieves several orders of magnitude but observable under an operation. We proposed to change
improvement in accuracy and online computation costs over the objective function of the GANs to train it to learn
the classical weighted least-squares (WLS) estimates. the distribution of power injections using the aggregated
smart meter measurements. The state estimation algorithm
B. Related Work using regression learning is improved with a prewhitening
State estimation based on deterministic state models has technique and more regularization techniques.
been extensively studied. See [1], [2] and references therein.
We henceforth highlight only a subset of the literature with II. S YSTEM M ODEL AND BAYESIAN S TATE E STIMATION
techniques suitable for distribution systems. The system state vector of the power grid at time t is
In some of the earliest contributions [7]–[10], it was well defined by xit = Vti ∠θti where Vit is the voltage magnitude
recognized that a critical challenge for distribution system and ∠θti is the phase angle for the state variable of bus
state estimation is the lack of observability. Different from i. The overall system state xt = [x1t , · · · , xN
t ]

is the
the Bayesian solution considered in this paper, most exist- column vector consisting of voltage phasors at all buses.
ing approaches are two-step solutions that produce pseudo A SCADA measures active/reactive power injections, power
measurements to make the system observable followed by flows measurements and voltage magnitude. The SCADA
applying WLS and other well-established techniques. measurement vector yt and system state xt are related by
From an estimation theoretic perspective, generating
pseudo measurements can be viewed as one of forecasting yt = h(xt ) + wt , t = 1, 2, . . . (1)
the real-time measurements based on historical data. Thus where t is the time index at the SCADA timescale
the pseudo-measurement techniques are part of the so-called (millisecond), h(·) is the measurement function, wt the
forecasting-aided state estimation [11], [12]. To this end, measurement noise.
machine learning techniques that have played significant
roles in load forecasting can be tailored to produce pseudo A. Weighted Least Squares Solutions and Observability:
measurements. See, e.g., [13]–[18]. The WLS estimator is optimal for observable systems in
Bayesian approaches to state estimation are far less ex- the absence of measurement:
plored even though the idea was already proposed in the
seminal work of Schweppe [19]. Bayesian state estima- x̂WLS (yt ) = arg min||yt − h(xt )||2 , (2)
xt
tion generally requires the computation of the conditional
statistics of the state variables. An early contribution that The goal of least squares is to minimize the modeling
modeled explicitly states as random was made in [20] where error. When there are not enough SCADAs installed, or some
load distributions were used to compute moments of states, SCADAs are faulty, the system becomes unobservable. WLS

Authorized licensed use limited to: University of Saskatchewan. Downloaded on October 04,2021 at 07:44:28 UTC from IEEE Xplore. Restrictions apply.
,(((,QWHUQDWLRQDO&RQIHUHQFHRQ&RPPXQLFDWLRQV&RQWURODQG&RPSXWLQJ7HFKQRORJLHVIRU6PDUW*ULGV 6PDUW*ULG&RPP

methods need pseudo measurements or extra constraints to (xt , yt ), it is difficult even to estimate the generative model.
estimate states when the system is unobservable. Second, even if we have f (xt , yt ), computing the conditional
mean are often intractable. Several early attempts [21],
B. Smart Meters in State Estimation: [26] employed belief propagations on a graphical model to
Smart meter measurements that, say, T times slower than compute the solution efficiently. These methods still require
the SCADA measurements. Then the smart meter measure- the underlying graphical model; they deserve a new look
ment vector z[n] is aggregated every T time: through the lenses of modern machine learning.
The advent of powerful deep learning tools and compu-

nT −1
tation resources such as GPU and cloud computing make
z[n] = g(xi ) + v[n] (3)
it possible to overcomes the above challenges of Bayesian
i=nT −T
state estimation. The key idea of our approach is to embed
where g(·) is the power injection function and v[n] the the underlying physical law in the neural network learning
measurement noise. process.
SCADAs are not widely deployed such that most distri-
bution systems are unobservable. The common practice is to
calculate pseudo-measurements using smart meters and solve
(2).

C. SCADA and Smart Meter Based Bayesian State Estima-

tion
We present an alternative way to exploit smart meter
measurements. As a significant departure from the WLS-
based techniques, we take a Bayesian viewpoint in which
system state xt is random and jointly distributed with the
measurement yt . Given that the power system is driven by Fig. 1: Flowchart of the algorithm.
increasingly stochastic renewable generations and demands,
Shown in Fig. 1 is a schematic that highlights the building
this assumption seems both natural and appropriate. By
blocks of the proposed deep learning approach to Bayesian
modeling xt as a random variable, we are able to make use
state estimation. The proposed scheme includes the online
of its prior distribution f (xt ) and the discriminative model
state estimation in the upper right and the offline learning
through the conditional distribution f (yt |xt ).
in the rest of the figure. The online state estimation is
Let yt be the SCADA measurement vector and xt the
through a deep neural network that approximates the MMSE
system state as defined in (1). Unlike the WLS techniques
estimator defined in with SCADA measurements yt and
that minimize the difference between the actual measurement
discrete network state Jt (such as breaker state) as its input,
yt and predicted measurement Hxt , a Bayesian state estima-
and the state estimate x̂t the output. The computation cost of
tion method places the optimization objective directly on the
producing state estimate in real-time is orders of magnitude
actual state xt and its estimate x̂t . Using the Euclidean norm
* as the measure of error, the Bayesian estimator is referred to
lower than the WLS-based solutions and can be further
reduced with special hardware implementations [27].
as the minimum mean-squared error (MMSE) state estimator
The offline learning includes distribution and regres-
given by a deceptively simple solution in the form of the
sion learning modules. Taking samples from smart meters,
conditional mean:
weather, and other external data, distribution learning pro-
duces an estimate of the probability distribution of the (fast
min E(||xt − x̂(yt )||22 ) → x̂M
t
M SE
= E(xt |yt ) (4) timescale) power injection. For each net injection sample
x̂t (·)
st drawn from F̂s , the solution of the power flow equation
A major advantage of the Bayesian formulation is that gives the system state xt that gives a measurement sample
system observability is no longer required, and the Bayesian yt . We now have a training sample (xt , yt ). A collection of
estimator gives the lowest possible mean squared error these training samples are used to set the neural network
among all estimators including all existing WLS based parameter w via an empirical risk minimization.
estimators regardless whether the system is observable. A. State Estimation via Deep Neural Networks:
III. BAYESIAN S TATE E STIMATION VIA D EEP L EARNING Using the samples of states and measurements, we trained
a multilayer feedforward neural network to approximate the
The Bayesian estimator comes with several nontrivial
optimal MMSE estimator. The neural network, Fig 2 consists
implementation challenges. First, the Bayesian formulation
of multiple layers of neurons. Neurons at each layer produce
requires knowledge on the generative model - the joint distri-
a vector output for the next layer using a (parameterized)
bution f (xt , yt ) - of xt and yt . Without access to samples of
nonlinear function of the output from the previous layer.
* Other norms such that l1 − norm or the probability of error measure are The universal approximation theorem (see, e.g., [28]) has
also of interest established that a neural network with a single hidden layer

Authorized licensed use limited to: University of Saskatchewan. Downloaded on October 04,2021 at 07:44:28 UTC from IEEE Xplore. Restrictions apply.
,(((,QWHUQDWLRQDO&RQIHUHQFHRQ&RPPXQLFDWLRQV&RQWURODQG&RPSXWLQJ7HFKQRORJLHVIRU6PDUW*ULGV 6PDUW*ULG&RPP

Fig. 3: Generative Adversarial Network.

to detect it. After training, the generative network implicitly

learns the probability distribution of the data. The objective
Fig. 2: Multilayer Feedforward Neural Network. function of GANs as described in [33] is a two-player
minimax game,
is sufficient to approximate an arbitrary continuous function. min max Ey∼fdata [log(D(y, WD ))]+
This means that with a sufficiently large neural network and Wg WD
(5)
appropriately chosen parameters, a neural network can well Eu∼U [log(1 − D(G(u, WG ), WD ))],
approximate the MMSE state estimator.
where D is the Discriminative network, G the Generative
Given the set of training samples
network, Wg and Wd are the parameters to learn, U the
S = {(xt , yt ) | k = 1, · · · , |S|} uniform distribution and fdata the distribution of data.
To be able to learn the fast-timescale power injection
generated, the weight matrix w is chosen to minimize the
distribution using the smart meter measurements, we changed
empirical risk defined by
the algorithm and objective function. Instead of generating
1 
L(w; S) = ||xt − K(yt ; w)||2 , a sample from the generator, we generated T samples and
|S| aggregated them to imitate the smart meter measurements in
k:(xt ,yt )∈S
each step. The objective function is replaced by,
w∗ = arg min L(w; S).
w
min max Ey∼fdata [log(D(y, WD ))]+
The empirical risk minimization problem above is well stud- Wg WD
ied for deep learning problems, and an extensive literature 
T (6)
exists. See, e.g., [29]. Eu∼U [log(1 − D( G(ui , WG ), WD ))].
For the state estimation problem at hand, the class of i=1
stochastic gradient descent algorithms is considered. The IV. S IMULATIONS R ESULTS AND D ISCUSSIONS
Adam algorithm [30] designed for non-stationary objectives
and noisy measurements is particularly suitable. A. Simulation Settings
A characteristic of deep learning is over-fitting, which
a) Systems simulated: IEEE 118-Bus Test System is
means that the number of neural network parameters tends
used in the simulations [34]. It is assumed the 99 of the
to be large relative to the training data set. To overcome
busses have both load and generation. To solve the power
over-fitting regularization techniques should be used that
flow equations MATPOWER toolbox [35] is used.
constraint in some way the search process in neural network
training. In training, we used dropout and batch normaliza- It is assumed there are two kind of meters in the system. (i)
tion. We used a constant prewhitening layer as the first layer Smart meters measures aggregated power injection at loads
to preprocess the inputs. every 15 minutes. They are placed at all nodes. (ii) SCADA
meters are assumed to measure the active and reactive power
B. Learning from Smart Meters via Generative Deep Learn- injection and power flow. It is assumed that the number
ing : of SCADA measurements are not enough to make the
A barrier to train a machine learning model for state esti- system observable. It is assumed there is an independent and
mation is the lack of training samples. Distribution learning identically distributed additive Gaussian measurement noise
can be parametric or non-parametric [31], [32]. Although with zero mean and variance set at 1% of the average net
parametrical methods are practical and easy to use, they consumption value.
have biasses and their learning capacities are often limited. b) Performance measure: The average squared er-
We propose to use aggregated historical power injection data ror(ASE) per-node defined by,
from Smart Meters to learn the distributions using a gener- 1 
ative machine learning model called Generative Adversarial ASE = ||x̂[k] − x[k]||2 , (7)
MN
Networks (GAN) described in [33]. k

As shown in Fig. 3 GAN consists of a Generative and a is used as a performance measure, where M is the number
Discriminative networks which are simultaneously trained. of Monte Carlo runs, k the index of the Monte Carlo run,
While the generative network learns to generate samples N the number of nodes, x̂[k] and x[k] the estimated and the
similar to the data, the discrimination network learns how state vectors, respectively. The objective function in the deep

Authorized licensed use limited to: University of Saskatchewan. Downloaded on October 04,2021 at 07:44:28 UTC from IEEE Xplore. Restrictions apply.
,(((,QWHUQDWLRQDO&RQIHUHQFHRQ&RPPXQLFDWLRQV&RQWURODQG&RPSXWLQJ7HFKQRORJLHVIRU6PDUW*ULGV 6PDUW*ULG&RPP

learning model is chosen as minimize mean-squared error to into training (12000 samples), validation(4000 samples) and
approximate the optimal estimator. test(4000 samples) sets. For each one, we solved the power
c) Distribution Learning via generative deep learning: flow equations by MATPOWER to obtain the power flow
We used aggregated power injection data from the Pecan values and states. For this network, 32 optimally placed
Street collection† for distribution learning. It is beneficial to SCADA meters make the system observable. We repeated
select the training samples of deep learning model using the the experiment with 8, 14, 20 and 26 SCADAs which are
historical data with similar features such as the season, hour guaranteed to be unobservable. For each case, measurements
of the day, weather, etc. We collected the net power injection are imitated by adding measurements noise and states. The
data for each bus fixing the hour of the day at 5 pm from noisy measurements of the SCADA meters are chosen as
the 1st of May to the 31st of August in 2018. the input of the network. Prewhitening is used on the inputs.
We assumed that the power injection distribution at each The states are chosen as the output. Then we trained a deep
bus is a linear transformation of one distribution, they differ neural network with 5 to 10 hidden layers to estimate states
only by mean and variance. We normalized all measurements on the test set.
to obtain the samples of that distribution. Then, we used The ReLU (Rectified Linear Units) activation function was
these samples to train the GAN. As the data samples are used for neurons in the hidden layers and linear activation
from smart meters, we modified the objective function of functions in the output layer. The Adam algorithm was used
training as described in (6). to train the neural network with mini-batches of 60 samples.
We trained the generative network with 2 hidden layers Early stopping was applied by monitoring validation errors.
and 100 neurons at each layer. Batch normalization and To select an initial point for the optimization, He’s normal
dropout with rate 0.2 are used in layers. Adam optimization method [36] was used. To have a better regularization, batch
[30] algorithm with mini-batches is used as the optimizer. normalization and dropout with 0.3 dropping rate are used
Leaky-ReLU at hidden layers and a linear activation function at hidden layers.
at the final layer are used as the activation functions. For e) Comparison of Performances: We implemented the
the discriminative network, we used two hidden layers with proposed deep learning approach to Bayesian state estimation
30 neurons. Leaky-ReLU at hidden layers and a sigmoid on the IEEE 118-bus. We compared the proposed Bayesian
activation function at the final layer are used as the activation state estimation with deep neural network (herein abbreviated
functions. as Bayesian NN) with two WLS-based methods in the
After training the GAN and we generated many power literature:
injection vectors using the generator network. To verify
the results, we aggregated the generated samples to imitate 1) WLS with pseudo measurements: referred to as Regu-
the smart meter measurements and plotted the cumulative larized WLS generates injection pseudo measurements
distribution function (CDF) of it and empirical CDF of by normalizing the smart meter measurement over the
the raw data in Fig. 4. The figure shows empirical CDF interval.
obtained by the samples and CDF obtained from the learned 2) Augmented WLS: uses only SCADA measurements to
distribution are similar. It verifies the training was successful. estimate states. An extra constraint is added on the (2).
We also observed the discriminative function’s output is Fig. 5 presents the performance of three algorithms on
converged to a constant value of 0.5 as expected, described five scenarios. It is demonstrated that for highly unob-
in [33]. servable systems, Bayesian methods are advantageous over
conventional WLS methods. Note that the MSE floor of the
augmented WLS method due to the SCADA - Smart Meter
mismatch.

Fig. 4: The comparison of the CDFs. The red curve is the

empirical CDF of the data, the blue curve is CDF
of the imitated measurements from the learned
distribution.

d) Neural network specification and training: We sep-

arated the power injection vectors generated by the generator Fig. 5: The comparison of algorithms on 5 test cases.
† http://www.pecanstreet.org/

Authorized licensed use limited to: University of Saskatchewan. Downloaded on October 04,2021 at 07:44:28 UTC from IEEE Xplore. Restrictions apply.
,(((,QWHUQDWLRQDO&RQIHUHQFHRQ&RPPXQLFDWLRQV&RQWURODQG&RPSXWLQJ7HFKQRORJLHVIRU6PDUW*ULGV 6PDUW*ULG&RPP

V. C ONCLUSION [15] E. Manitsas, R. Singh, B. C. Pal, and G. Strbac, “Distribution

system state estimation using an artificial neural network approach
This paper presents a comprehensive Bayesian state es- for pseudo measurement modeling,” IEEE Transactions on Power
Systems, vol. 27, pp. 1888–1896, Nov 2012.
timation algorithm for unobservable distribution systems [16] J. Wu, Y. He, and N. Jenkins, “A robust state estimator for medium
using SCADA and smart meters under a stochastic demand voltage distribution networks,” IEEE Transactions on Power Systems,
model with generation. The proposed approach employs 2013.
[17] A. Onwuachumba, Y. Wu, and M. Musavi, “Reduced model for power
two machine learning techniques: distribution learning of system state estimation using artificial neural networks,” IEEE Green
power injection using generative adversary networks and Technologies Conference, 2013.
regression learning of MMSE estimator using deep learning. [18] F. Adinolfi, F. D’Agostino, A. Morini, M. Saviozzi, and F. Silvestro,
“Pseudo-measurements modeling using neural network and fourier de-
The performance of traditional methods are not fulfilling composition for distribution state estimation,” in IEEE PES Innovative
when the system is unobservable, where Bayesian alterna- Smart Grid Technologies, Europe, pp. 1–6, Oct 2014.
tives are still viable options. The objective function of GANs [19] F. C. Schweppe, J. Wildes, and D. P. Rom, “Power system static state
estimation, Parts I, II, III,” IEEE Transactions on Power Apparatus
is modified to learn the distribution of power injections and Systems, vol. PAS-89, pp. 120–135, 1970.
from the historical smart meter measurements. The use of [20] A. K. Ghosh, D. L. Lubkeman, and R. H. Jones, “Load modeling
deep neural network also plays a crucial role in overcoming for distribution circuit state estimation,” IEEE Transactions on Power
Delivery, vol. 12, pp. 999–1005, April 1997.
computation complexity in Bayesian estimation, making the [21] P. Chavali and A. Nehorai, “Distributed power system state estimation
online computation significantly lower than the traditional using factor graphs,” IEEE Transactions on Signal Processing, vol. 63,
WLS solutions. Simulation results demonstrate the potential pp. 2864–2876, June 2015.
[22] L. Schenato, G. Barchi, D. Macii, R. Arghandeh, K. Poolla, and A. V.
of the Bayesian state estimation for cases that are otherwise Meier, “Bayesian linear state estimation using smart meters and pmus
intractable for conventional WLS-based techniques. measurements in distribution grids,” in 2014 IEEE International Con-
ference on Smart Grid Communications (SmartGridComm), pp. 572–
577, Nov 2014.
R EFERENCES [23] K. Emami, T. Fernando, H. H. Iu, H. Trinh, and K. P. Wong, “Particle
filter approach to dynamic state estimation of generators in power
[1] A. Abur and A. G. Expósito, Power System State Estimation: Theory systems,” IEEE Transactions on Power Systems, vol. 30, pp. 2665–
and Implementation. CRC Press, 2004. 2675, Sep 2015.
[2] A. Monticelli, State Estimation in Electric Power Systems: A Gen- [24] A. Angioni, T. Schlösser, F. Ponci, and A. Monti, “Impact of pseudo-
eralized Approach. Power Electronics and Power Systems, Springer, measurements from new power profiles on state estimation in low-
1999. voltage grids,” IEEE Transactions on Instrumentation and Measure-
[3] O. Kosut, L. Jia, R. J. Thomas, and L. Tong, “Malicious data attacks ment, vol. 65, pp. 70–77, Jan 2016.
on the smart grid,” IEEE Transactions on Smart Grid, vol. 2, pp. 645– [25] K. R. Mestav, J. Luengo-Rozas, and L. Tong, “Bayesian state estima-
658, Dec 2011. tion for unobservable distribution systems via deep learning,” IEEE
[4] A. Alimardani, F. Therrien, D. Atanackovic, J. Jatskevich, and E. Vaa- Transactions on Power Systems, pp. 1–1, 2019.
hedi, “Distribution system state estimation based on nonsynchronized [26] Y. Hu, A. Kuh, T. Yang, and A. Kavcic, “A belief propagation
smart meters,” IEEE Transactions on Smart Grid, vol. 6, pp. 2919– based power distribution system state estimator,” IEEE Computational
2928, Nov 2015. Intelligence Magazine, vol. 6, pp. 36–46, Aug 2011.
[5] A. Al-Wakeel, J. Wu, and N. Jenkins, “State estimation of medium [27] V. Sze, Y. Chen, T. Yang, and J. S. Emer, “Efficient processing of
voltage distribution networks using smart meter measurements,” Ap- deep neural networks: A tutorial and survey,” Proceedings of the IEEE,
plied Energy, vol. 184, pp. 207 – 218, 2016. vol. 105, pp. 2295–2329, Dec 2017.
[6] Y. Gao and N. Yu, “State estimation for unbalanced electric power [28] S. Sonoda and N. Murata, “Neural network with unbounded activation
distribution systems using ami data,” in 2017 IEEE Power Energy functions is universal approximator,” Applied and Computational
Society Innovative Smart Grid Technologies Conference (ISGT), pp. 1– Harmonic Analysis, vol. 43, no. 2, pp. 233 – 268, 2017.
5, April 2017. [29] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT
[7] I. Roytelman and S. M. Shahidehpour, “State estimation for electric Press, 2016. http://www.deeplearningbook.org.
power distribution systems in quasi real-time conditions,” IEEE Trans- [30] D. P. Kingma and J. L. Ba, “Adam: A method for stochastic optimiza-
actions on Power Delivery, vol. 8, pp. 2009–2015, Oct 1993. tion,” International Conference for Learning Representations, 2015.
[8] M. E. Baran and A. W. Kelley, “State estimation for real-time [31] L. Wasserman, All of Statistics. Springer-Verlag New York, 2004.
monitoring of distribution systems,” IEEE Transactions on Power [32] L. Wasserman, All of Nonparametric Statistics. Springer-Verlag New
Systems, vol. 9, pp. 1601–1609, Aug 1994. York, 2006.
[9] C. N. Lu, J. H. Teng, and W. . E. Liu, “Distribution system state [33] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-
estimation,” IEEE Transactions on Power Systems, vol. 10, pp. 229– Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative Adversarial
240, Feb 1995. Networks,” arXiv e-prints, p. arXiv:1406.2661, Jun 2014.
[34] R. Christie, “Power systems test case archive,” Univ. Washington,
[10] A. K. Ghosh, D. L. Lubkeman, M. J. Downey, and R. H. Jones,
Seattle, WA, USA., 1993. https://www.ee.washington.edu/
“Distribution circuit state estimation using a probabilistic approach,”
research/pstca/pf118/pg tca118bus.htm.
IEEE Transactions on Power Systems, vol. 12, pp. 45–51, Feb 1997.
[35] R. D. Zimmerman, C. E. Murillo-Sanchez, and R. J. Thomas, “Mat-
[11] M. B. D. C. Filho and J. C. S. de Souza, “Forecasting-aided state
power: Steady-state operations, planning, and analysis tools for power
estimation—part i: Panorama,” IEEE Transactions on Power Systems,
systems research and education,” IEEE Transactions on Power Sys-
vol. 24, pp. 1667–1677, Nov 2009.
tems, vol. 26, pp. 12–19, Feb 2011.
[12] M. B. D. C. Filho, J. C. S. de Souza, and R. S. Freund, “Forecasting- [36] K. He, X. Zhang, S. Ren, and J. Sun, “Delving deep into rectifiers: Sur-
aided state estimation—part ii: Implementation,” IEEE Transactions passing human-level performance on imagenet classification,” CoRR,
on Power Systems, vol. 24, pp. 1678–1685, Nov 2009. 2015.
[13] A. Bernieri, C. Liguori, and A. Losi, “Neural networks and pseudo-
measurements for real-time monitoring of distribution systems,” in
Proceedings of 1995 IEEE Instrumentation and Measurement Tech-
nology Conference - IMTC ’95, pp. 112–119, April 1995.
[14] R. Singh, B. C. Pal, and R. A. Jabr, “Distribution system state
estimation through gaussian mixture model of the load as pseudo-
measurement,” IET Generation, Transmission Distribution, vol. 4,
pp. 50–59, January 2010.

Authorized licensed use limited to: University of Saskatchewan. Downloaded on October 04,2021 at 07:44:28 UTC from IEEE Xplore. Restrictions apply.