2022 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm)

2022 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm) | 978-1-6654-3254-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/SMARTGRIDCOMM52983.2022.9960967

Distributed Nonlinear State Estimation in Electric

Power Systems using Graph Neural Networks
Ognjen Kundacina, Student Member, IEEE, Mirsad Cosovic, Member, IEEE, Dragisa Miskovic, Dejan
Vukobratovic, Senior Member, IEEE

Abstract—Nonlinear state estimation (SE), with the goal of integrate phasor measurements into the established model with
estimating complex bus voltages based on all types of measure- legacy measurements. A standard way to include voltage and
ments available in the power system, is usually solved using the current phasors coming from PMUs is to represent them in
iterative Gauss-Newton (GN) method. The nonlinear SE presents
some difficulties when considering inputs from both phasor the rectangular coordinate system [2]. The main disadvantage
measurement units and supervisory control and data acquisition of this approach is related to measurement errors, where mea-
system. These include numerical instabilities, convergence time surement errors of a single PMU are correlated and covariance
depending on the starting point of the iterative method, and matrix does not have diagonal form. Despite that, because of
the quadratic computational complexity of a single iteration the lower computational effort, the measurement error covari-
regarding the number of state variables. This paper introduces
an original graph neural network based SE implementation over ance matrix is usually considered as diagonal matrix, which
the augmented factor graph of the nonlinear power system SE, has the effect on the accuracy of the SE. The diagonal form
capable of incorporating measurements on both branches and of the covariance matrix could be preserved by representing
buses, as well as both phasor and legacy measurements. The voltage and current phasors coming from PMUs in the polar
proposed regression model has linear computational complexity coordinate system, which requires a large computational effort
during the inference time once trained, with a possibility of
distributed implementation. Since the method is noniterative and with a convergence time significantly depending on the state
non-matrix-based, it is resilient to the problems that the GN variables’ initialisation [3]. Additionally, using magnitudes of
solver is prone to. Aside from prediction accuracy on the test branch current measurements can cause numerical instabilities
set, the proposed model demonstrates robustness when simulating such as undefined Jacobian elements due to the “flat start”
cyber attacks and unobservable scenarios due to communication [4, Sec. 9.3]. Furthermore, different orders of magnitude of
irregularities. In those cases, prediction errors are sustained
locally, with no effect on the rest of the power system’s results. phasor and legacy measurement variances can make the SE
problem ill-conditioned by increasing the condition number of
Index Terms—Machine Learning, Graph Neural Networks, the estimator’s gain matrix [2]. A single iteration of the GN
Power Systems, State Estimation, Real-Time Systems
method involves solving a system of linear equations, which
results in near O(n2 ) computational complexity for sparse
I. I NTRODUCTION matrices, where n denotes the number of state variables.
Motivation: The power system state estimation (SE) is a Graph neural networks (GNNs) [5], [6] are a promising
problem of determining the state of the power system repre- method for iterative problems in large-scale systems because
sented as the set of complex bus voltages, given the available once trained, they have a O(n) computational complexity,
set of measurements [1]. The dominant part of the input data with the possibility of distributed implementation. Apart from
for the SE model consists of legacy measurements coming not being restricted to training and test examples with fixed
from the supervisory control and data acquisition (SCADA) topologies like the other deep learning approaches, GNNs are
system, which have relatively high variance, high latency, and permutation invariant, have a smaller number of trainable pa-
low sampling rates. Increasingly deployed phasor measure- rameters, require less memory for the deep learning model, and
ment units (PMUs), provided by the wide area measurement can easily incorporate topology information into the learning
system (WAMS), have low variance and high sampling rates process. Since GNN inference is a noniterative procedure, its
and are a potential enabler of real-time system monitoring. computational time depends on the power system size only and
Taking into account both legacy and phasor measurements is not sensitive to state variable initialisation. Furthermore, it
results in the SE model formulated by the system of nonlinear does not suffer from numerical instability since it is not based
equations and is traditionally solved using the iterative Gauss- on the matrix model of the power system.
Newton (GN) method. Different approaches can be used to Literature Review: Various studies suggest using well-
known deep learning models like feed-forward and recur-
O. Kundacina and D. Miskovic are with The Institute for Arti- rent neural networks to approximate the existing iterative
ficial Intelligence Research and Development of Serbia (e-mail: ogn- SE solvers [7] or to provide them with the state variable
[email protected], [email protected]); M. Cosovic is with Fac-
ulty of Electrical Engineering, University of Sarajevo, Bosnia and Herzegovina initialisation [8]. Recently, there have been several proposals
(e-mail: [email protected]); D. Vukobratovic is with Faculty of Technical for applying GNNs to similar power system analysis problems,
Sciences, University of Novi Sad, Serbia, (email: [email protected]). like power flow [9], [10] and probabilistic power flow [11].
978-1-6654-3254-2/22/$31.00 ©2022 IEEE The work described in [12] presents a hybrid model and a

8 May 24,2023 at 19:27:55 UTC from IEEE Xplore. Restrictions apply.

Authorized licensed use limited to: Dejan Vukobratovic. Downloaded on
2022 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm)

data-based method for simultaneous power system parameters h(x) comprises a collection of m nonlinear equations. Next,
and state estimation using GNNs. The model predicts active u ∈ Rm is a vector of uncorrelated measurement errors, where
and reactive power injections based on the bus voltage phasor ui ∼ N (0, vi ) represents a zero-mean Gaussian distribution
measurements, whereas it does not take into account other with variance vi . The nonlinear SE model (1) is a consequence
measurement types, such as legacy measurements and branch of legacy and phasor measurements provided by SCADA and
current phasor measurements. The regularisation term in the WAMS, respectively. The typical set of legacy measurements
SE loss function in [13] uses a GNN trained on the histor- provided by SCADA includes active and reactive power flow
ical dataset to forecast voltage labels. The GNN component and injection, branch current magnitude, and bus voltage
of the proposed algorithm considers only inputs from node magnitude measurements. The WAMS supports phasor mea-
voltage measurements, while the other components consider surements and includes magnitude and angle measurements of
different measurement types, resulting in an overall O(n2 ) bus voltages and branch currents [19, Sec. 5.6].
computational complexity during the inference phase. In [14], Each measurement is associated with the measurement
state variables are predicted using a GNN with gated recurrent value zi , the measurement variance vi , and the measurement
units trained on a time-series of node voltage measurements function hi (x). The form of the function hi (x) depends on
acquired from PMUs. the branch model of the power system. Usually, the two-port
Contributions: In this work, we propose a data-driven π-model is used, where the branch between buses i and j is
nonlinear state estimator based on graph attention networks settled using complex expressions:
[15]. The proposed approach is an extension of our previous 
1

work on linear SE with PMUs [16] and linear SE with PMUs ¯ ∗
   
Iij
=  τ 2 (yij + ysij ) −αij yij  V̄i , (2)
considering covariances of measurement phasors represented I¯ji ij V̄j
in rectangular coordinates [17]. The contributions are listed as −αij yij yij + ysij
follows: where the parameter yij = gij + jbij represents the branch
• We provide an original implementation of GNNs over a series admittance, and half of the total branch shunt admittance
factor graph [18] obtained by transforming the bus/branch is given as ysij = jbsi . The transformer complex ratio is
power system model. Furthermore, we augment the factor defined as αij = (1/τij )e−jϕij , where τij is the transformer
graph to increase the information propagation in unob- tap ratio magnitude, while ϕij is the transformer phase shift
servable scenarios. angle. The complex expressions I¯ij and I¯ji define the branch
• The proposed method takes into account all of the legacy currents from the bus i to the bus j and from the bus j to the
measurements, as well as bus voltage and branch current bus i, respectively. The complex bus voltages at buses i and
phasor measurements, and provides a trivial way to j are given as V̄i = Vi ejθi and V̄j = Vj ejθj , where Vi , θi and
remove or add additional measurements by altering the Vj , θj denote the voltage magnitude and angle at buses i and
corresponding factor nodes in the graph. j, respectively.
• We designed the GNN-based state estimator to be fast, ro- The GN method is typically used to solve the nonlinear SE
bust to ill-conditioned scenarios, and an accurate nonlin- model (1), where the measurement functions h(x) precisely
ear SE approximator, once trained on the relevant dataset, follow the physical laws derived on the basis of (2):
labelled by the nonlinear SE solved by GN. The proposed h i
model has O(n) computational complexity during the J(x(ν) )T WJ(x(ν) ) ∆x(ν) = J(x(ν) )T Wr(x(ν) ) (3a)
inference and can be distributed among geographically
x(ν+1) = x(ν) + ∆x(ν) , (3b)
separated processing units.
• We evaluated the performance of the proposed method by where ν = {0, 1, . . . , νmax } is the iteration index and νmax
testing on various data samples, including unobservable is the number of iterations, ∆x(ν) ∈ Rn is the vector
cases caused by communication errors or measurement of increments of the state variables, J(x(ν) ) ∈ Rmxn is
device failures, and scenarios corrupted by malicious data the Jacobian matrix of measurement functions h(x(ν) ) at
injections. x = x(ν) , W ∈ Rmxm is a diagonal matrix containing inverses
• In addition to the standalone application, the proposed of measurement variances, and r(x(ν) ) = z −h(x(ν) ) is the
approach can be used as a fast and accurate initialiser of vector of residuals. Note that the nonlinear SE represents
the GN method by providing it with a starting point near a nonconvex problem arising from nonlinear measurement
the exact solution. functions h(x) [20]. Due to the fact that the values of the
state variables x usually fluctuate in narrow boundaries, the
II. N ONLINEAR S TATE E STIMATION P RELIMINARIES GN method can be used.
In a general scenario, the SE model is described with the The SE model (1) that considers both legacy and phasor
system of nonlinear equations: measurements, where the vector of state variables x = [V, θ]T
and phasor measurements are represented in the polar coor-
z = h(x) + u, (1)
dinate system, is known as simultaneous. The simultaneous
where bus voltage magnitudes V and bus voltage angles θ SE model takes measurements provided by PMUs in the
are observed as state variables x = [V, θ]T ∈ Rn , the same manner as legacy measurements. More precisely, the
vector z ∈ Rm contains measurement values, and the vector PMU generates measurements in the polar coordinate system,

9 May 24,2023 at 19:27:55 UTC from IEEE Xplore. Restrictions apply.

Authorized licensed use limited to: Dejan Vukobratovic. Downloaded on
2022 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm)

which delivers more accurate state estimates than the other h1 k−1 Message
representations [2], but requires more computing time [3] and m1,j k−1
produces ill-conditioned problems [2]. To address these issues,
we propose a non-matrix-based and noniterative GNN base mj k−1
h2 k−1 Message Aggregate Update hj k
SE, which can be used as a standalone approach to solve (1), m2,j k−1
or as a fast and accurate initialiser of the GN method (3). ...

III. G RAPH N EURAL N ETWORK BASED N ONLINEAR S TATE Fig. 1. One message passing iteration, implemented as a GNN layer, consists
E STIMATION of several trainable functions, shown as yellow rectangles with indicated inputs
and outputs.
This section introduces the fundamentals of spatial graph
neural networks, the transformation of the power system’s
bus/branch model to an augmented factor graph, and the B. Augmented Power System Factor Graph and the Proposed
proposed GNN architecture, specialised for augmented factor GNN Architecture
graphs.
Inspired by the application of probabilistic graphical models
for the nonlinear power system SE [21], we first transform the
A. Fundamentals of Spatial Graph Neural Networks
bus/branch power system model into the power system’s factor
The spatial GNNs perform recursive neighbourhood aggre- graph, a bipartite graph comprised of the factor and variable
gation, also known as message passing [5], over the local sub- nodes. We use variable nodes to generate an s-dimensional
sets of graph-structured inputs to create a meaningful represen- node embedding and the predictions of the state variables,
tation of the connected pieces of data. More precisely, GNNs i.e., magnitudes and angles of the bus voltages, Vi and θi , i =
act as a trainable local graph filter which has a goal of trans- 1, . . . , n. Because variable nodes have no other input features,
forming the inputs from each node and its connections to a binary index encoding is added as an input feature to help
higher dimensional space, resulting in an s-dimensional vector the GNN model distinguish between similar subgraphs. Factor
embedding h ∈ Rs per node. The GNN layer, which represents nodes propagate legacy and phasor measurement values and
one message passing iteration, combines multiple trainable variances to neighbouring variable nodes via GNN message
functions commonly implemented as neural networks. The passing via corresponding node embedding. When creating the
message function Message(·) : R2s 7→ Ru calculates the factor graph from the bus/branch power system model, each
message mi,j ∈ Ru between two node embeddings, hi and phasor measurement generates two factor nodes, while each
hj . The aggregation function Aggregate(·) : Rdeg(j)·u 7→ Ru legacy measurement generates one factor node. As an example,
combines the arriving messages in a specific way, and outputs we consider a simple two-bus power system, in which we
the aggregated messages mj ∈ Ru for each node j. At the end placed one voltage phasor measurement on the first bus and
of one message passing iteration, aggregated messages are fed one legacy voltage magnitude measurement on the second
into the update function Update(·) : Ru 7→ Rs , to calculate bus. Additionally, we placed one current phasor measurement
the update of each node’s embedding. Node embedding values and one legacy active power flow measurement on the branch
are initialised with the node’s input feature vector transformed connecting the two nodes. The factor graph of this simple
to the s-dimensional space, after which the message passing power system consists of the generated factor and variable
repeats K times, with K being a hyperparameter known as nodes, connected by full-line edges, as shown in Fig. 2. In
the number of GNN layers. One message passing iteration contrast to approaches such as [12], in which GNN nodes
corresponding to the k th GNN layer is displayed in Fig. 1 correspond to state variables only, we find applying GNNs to
and also described analytically by (4): factor-graph-like topologies more flexible. Since factor nodes
mi,j k−1 = Message(hi k−1 , hj k−1 ) can be added or removed from any point in the graph, it is
simple to simulate the inclusion of different types and numbers
mj k−1 = Aggregate({mi,j k−1 |i ∈ Nj }) of measurements both on power system buses and branches,
(4)
hj k = Update(mj k−1 ) such as multiple legacy and phasor measurements on a single
k ∈ {1, . . . , K}, branch.
As shown in [16], augmenting the factor graph by adding
where Nj denotes the 1-hop neighbourhood of the node j,
direct connections between the variable nodes separated by
and the vector superscript corresponds to the message passing
only one factor node1 notably improves the model’s per-
iteration. The outputs of the message passing process are
formance when some portion of measurements are removed
final node embeddings hj K , which are then passed through
from GNN inputs, like when simulating communication failure
the additional neural network layers, creating the outputs that
scenarios. We simulate the measurement data loss by removing
represent the predictions of the supervised GNN model. GNN
the corresponding factor nodes from the factor graph, which
training is performed by optimising the model parameters
can eliminate the connections that physically exist in the power
using the variants of the gradient descent algorithm, with the
loss function being some measure of the distance between the 1 Although augmenting the factor graph in the described way no longer
labels and the predictions. For a comprehensive overview of makes it bipartite, we will continue to use terms such as augmented factor
GNN foundations and algorithms, please refer to [6]. graph, factor nodes, and variable nodes.

Authorized licensed use limited to: Dejan Vukobratovic. Downloaded 10

on May 24,2023 at 19:27:55 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm)

V2 θ2 hf K−1
fV2 ... Layerf

hv K
... Layerv Layerv
...
hv2 K−1
fI12 fθI12 fP12
label
...
output
Loss Pred

fV 1 fθ1
V1 θ1
Fig. 3. The high-level computational graph displays the final message passing
Fig. 2. Factor graph (full-line edges) and the augmented factor graph for iteration along with prediction and loss calculation for the variable node v.
a simple two-bus power system. Variable nodes are depicted as circles, and Yellow rectangles represent trainable functions implemented as feed-forward
factor nodes are as squares, colored differently to distinguish between phasor neural networks.
and legacy measurements.

batch of graphs:
system, reducing the message passing between the nodes. For 2nB
1 X pred
example, removing both phasor and legacy measurements from L(θGNN ) = (xi − xi label )2 , (6)
2nB i=1
the branch in the described two-bus power system divides the
factor graph into two isolated parts, disabling the message where 2n is the total number of variable nodes in a graph, B
passing between the pairs of variable nodes. Augmenting the is the mini-batch size, and θGNN represents all the trainable
factor graph preserves the physical connection between the parameters. Fig. 3 displays a high-level computational graph
variable nodes, improving the message passing in unobserv- of the final message passing iteration, state variable prediction,
able scenarios. In the two-bus power system example, the and the loss function calculation for the part of the augmented
augmented factor graph consists of nodes connected by both factor graph given in Fig. 2.
full and dashed-line edges in Fig. 2. The power system augmented factor graph has bounded
We propose a GNN architecture specialised for working node degree, prediction for a single node requires only infor-
on heterogeneous augmented factor graphs, consisting of two mation from the node’s K-hop neighbourhood, and the number
types of GNN layers, Layerf (·) : Rdeg(f )·s 7→ Rs used for of GNN layers is small. Analogously to [17], we can conclude
updating embeddings of factor nodes f , and Layerv (·) : that the computational complexity of the proposed model’s
Rdeg(v)·s 7→ Rs for variable nodes v, with different sets inference for a single state variable is constant, which implies
of trainable parameters. In addition to variable-to-factor and that the overall complexity of the proposed GNN-based SE is
factor-to-variable node message functions, with the augmen- O(n). The implementation of the GNN model in large-scale
tation, we introduce new variable-to-variable node messages networks can be further improved by distributing the inference
and we model them with separate trainable message func- computation among local processors in the power system,
tion Messagev→v (·) : R2s 7→ Ru . For all of the message avoiding the communication delays between the measurement
functions we used two-layer feed-forward neural networks, devices and the central processing unit in the centralised SE
single layer neural networks for the update functions and implementation. For arbitrary K, measurements required for
the aggregation function based on the attention mechanism the single complex bus voltage inference are located within the
from graph attention networks [15]. State variables predic- ⌈K/2⌉-hop neighbourhood of the corresponding power system
tions xpred are generated by feeding the final variable node bus.
embeddings hK into the additional two-layer neural network
Pred(·) : Rs 7→ R. Equations (5) describe the recursive IV. N UMERICAL R ESULTS AND D ISCUSSION
neighbourhood aggregation and the state variable prediction In this section, we describe the GNN model’s training
processes: process and test the trained model on various examples to
validate its accuracy, and its robustness under measurement
hv k = Layerv ({hi k−1 |i ∈ Nv }) data loss due to communication failure and cyber attacks in
hf k = Layerf ({hi k−1 |i ∈ Nf }) the form of malicious data injections. The described GNN
(5) model for augmented factor graphs is implemented using the
xv pred = Pred(hv K ) IGNNITION library [22] and trained for 100 epochs in all
k ∈ {1, . . . , K}, experiments using the following hyperparameters: 64 elements
in the node embedding vector, 32 graphs in a mini-batch, four
where Nv and Nf denote the 1-hop neighbourhoods of vari- GNN layers, ReLU activation functions, Adam optimiser, and
able and factor nodes v and f . We use the mean squared a learning rate of 4 × 10−4 . We conducted separate training
difference between state variable predictions and labels xlabel experiments for IEEE 30 and IEEE 118-bus test cases, for
as a training loss function, calculated over the entire mini- which we generated a training set containing 10000 samples

Authorized licensed use limited to: Dejan Vukobratovic. Downloaded 11

on May 24,2023 at 19:27:55 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm)

and validation and test sets containing 100 samples each. the described cases by randomly removing a percentage of
Each sample is created by randomly sampling the active and all input measurements, ranging from 0% to 95% with a
reactive power injections and solving the power flow problem. step of 5%. We create 20 test sets per IEEE test case, each
Measurement values are created by adding Gaussian noise to containing samples with the same percentage of excluded
the power flow solutions, and the nonlinear SE is solved by GN measurements, and show the average test set MSEs in Fig. 5.
to label the input measurement set in each sample. We used Proposed GNN models yields predictions in all examples,
a Gaussian noise variance of 10−5 for phasor measurements, with an expected growing trend in MSE as the number of
10−3 for bus voltage magnitude and active and reactive power excluded measurements increases. In comparison, the GN
flow legacy measurements, and 10−1 for active and reactive method could not provide a solution for many examples due
injection legacy measurements. to underdetermined and ill-conditioned systems of nonlinear
For the IEEE 30-bus test case, we placed 100 legacy SE equations. A possible explanation for significantly lower
measurements and three PMUs (i.e., three bus voltage phasors MSEs for the IEEE 118-test case in these scenarios is that it
and eight branch current phasors) in each sample, resulting in contains a greater variety of subgraphs for GNN training. To
2.03 measurement redundancy. The trained model performed investigate the GNN predictions further, we create a test set
well on the test set, with the average test set mean square by excluding five measurements connected to the two directly
error of 1.233 × 10−5 between predictions and ground truth connected power system buses from each test sample, resulting
labels; the average test set MSE for voltage magnitudes of in the average test set MSE of 1.488 × 10−4 . Fig. 6 shows
5.221 × 10−6 ; the average test set MSE for voltage angles of the results for one test sample, where vertical dashed lines
1.944 × 10−5 . Fig. 4 shows the average test MSE per each correspond to the buses in the 1-hop neighbourhood of the
bus, where the upper plot corresponds to voltage magnitudes excluded measurements. We can observe that the deviation
and the lower one to voltage angles. from the ground truth values manifests mainly in the vicinity
For the IEEE 118-bus test case, we placed 500 legacy of the excluded measurements, not affecting the prediction
measurements and seven PMUs (i.e., seven bus voltage phasors accuracy in the rest of the power system.
and 26 branch current phasors) in each sample, resulting in
2.39 measurement redundancy. The average test set mean 100
square error equals 2.038 × 10−5 , with the average test set
Average MSE

10−1
MSE for voltage magnitudes of 1.572 × 10−5 and the average
test set MSE for voltage angles of 2.505 × 10−5 . Based on 10−2
the insights from both experiments, we can conclude that the 10−3 IEEE 30
proposed GNN model is a good approximator of the nonlinear 10−4 IEEE 118
SE solved by GN. 10−5
0 20 40 60 80 100
10−3 Percentage of Excluded Measurements
MSE of Vi

10−6
Fig. 5. Average MSEs of test sets created by randomly excluding measure-
−9 ments.
10
Voltage Angle θi Voltage Magnitude Vi

10−12 1.09 Predictions

10−3 1.06 Labels
MSE of θi

10−6 1.03
1
10−9
0.97
10−12
0
1 5 10 15 20 25 30
Bus Index i −0.1
−0.2
Fig. 4. The test set MSE between the predictions and the labels per each bus
for voltage magnitudes and angles in the IEEE 30-bus test case. −0.3
−0.4
0 5 10 15 20 25 30
A. Robustness to Loss of Input Data Bus Index i
Next, we observe predictions of the GNN models previ-
Fig. 6. GNN predictions and labels for one test example with all measure-
ously trained on IEEE 30 and IEEE 118-bus test data when ments connected to two neighbouring buses removed. Dashed lines indicate
exposed to the loss of input data caused by communication the buses in the 1-hop neighbourhood of the excluded measurements.
failures or measurement device malfunctions. We simulate

Authorized licensed use limited to: Dejan Vukobratovic. Downloaded 12

on May 24,2023 at 19:27:55 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm)

B. Behavior Under Malicious Data Injections VI. ACKNOWLEDGMENT

We examine the robustness of the proposed GNN model This paper has received funding from the European Union’s
to malicious data injection type of cyber attacks by randomly Horizon 2020 research and innovation programmes under
altering the values of five neighbouring measurements in each Grant Agreement numbers 856967 and 871518.
test sample. We compare the proposed GNN model’s predic-
tions with the solutions of the GN method and the ground R EFERENCES
truth values obtained using the GN method applied on the [1] A. Monticelli, “Electric power system state estimation,” Proceedings of
uncorrupted measurement data. The GNN model demonstrated the IEEE, vol. 88, no. 2, pp. 262–282, 2000.
an order of magnitude better performance than the GN method, [2] A. Gomez-Exposito, A. Abur, P. Rousseaux, A. de la Villa Jaen,
and C. Gomez-Quiles, “On the use of PMUs in power system state
with the average test set MSEs 1.281×10−4 and 1.034×10−3 , estimation,” Proc. IEEE PSCC, 2011.
respectively. Fig. 7 depicts the comparison of the state variable [3] G. N. Korres and N. M. Manousakis, “State estimation and observability
predictions under corrupted input data for one example from analysis for phasor measurement unit measured systems,” IET Gener.
Transm. Dis., vol. 6, no. 9, pp. 902–913, September 2012.
the test set. [4] A. Abur and A. Expósito, Power System State Estimation: Theory and
Implementation, ser. Power Engineering. Taylor & Francis, 2004.
[5] J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl,
Voltage Angle θi Voltage Magnitude Vi

1.12 Predictions “Neural message passing for quantum chemistry,” in Proceedings of the
GN based SE 34th International Conference on Machine Learning. PMLR, 06–11
1.08 Aug 2017, pp. 1263–1272.
Ground truth [6] W. L. Hamilton, “Graph representation learning,” Synthesis Lectures on
1.04 Artificial Intelligence and Machine Learning, vol. 14, no. 3, pp. 1–159,
2020.
1 [7] L. Zhang, G. Wang, and G. B. Giannakis, “Real-time power system state
estimation and forecasting via deep unrolled neural networks,” IEEE
0.96 Trans. Signal Process., vol. 67, no. 15, pp. 4069–4077, 2019.
[8] A. S. Zamzam, X. Fu, and N. D. Sidiropoulos, “Data-driven learning-
0 based optimization for distribution system state estimation,” IEEE Trans.
Power Syst., vol. 34, no. 6, pp. 4796–4805, 2019.
−0.1 [9] B. Donon, B. Donnot, I. Guyon, and A. Marot, “Graph neural solver
−0.2 for power systems,” in 2019 International Joint Conference on Neural
Networks (IJCNN), 2019, pp. 1–8.
−0.3 [10] V. Bolz, J. Rueß, and A. Zell, “Power flow approximation based on graph
−0.4 convolutional networks,” in Proc. IEEE ICMLA, 2019, pp. 1679–1686.
[11] D. Wang, K. Zheng, Q. Chen, G. Luo, and X. Zhang, “Probabilistic
0 5 10 15 20 25 30 power flow solution with graph convolutional network,” in 2020 IEEE
Bus Index i PES Innovative Smart Grid Technologies Europe (ISGT-Europe), 2020,
pp. 650–654.
Fig. 7. GNN predictions and GN based SE solutions for one test example [12] L. Pagnier and M. Chertkov, “Physics-informed graphical neural network
with corrupted input data. for parameter & state estimations in power systems,” 2021.
[13] Q. Yang, A. Sadeghi, and G. Wang, “Data-driven priors for robust
PSSE via Gauss-Newton unrolled neural networks,” IEEE Journal on
Emerging and Selected Topics in Circuits and Systems, vol. PP, pp. 1–1,
01 2022.
V. C ONCLUSION [14] M. J. Hossain and M. Rahnamay–Naeini, “State estimation in smart grids
using temporal graph convolution networks,” in 2021 North American
In this paper, we introduced a method for nonlinear SE Power Symposium (NAPS), 2021, pp. 01–05.
considering legacy and phasor measurements based on the [15] P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Liò, and
GNN model specialised for operating on augmented power Y. Bengio, “Graph Attention Networks,” International Conference on
Learning Representations, 2018.
system factor graphs. The method avoids the problems that [16] O. Kundacina, M. Cosovic, and D. Vukobratovic, “State estimation in
traditional nonlinear SE solvers face, such as numerical insta- electric power systems leveraging graph neural networks,” in 2022 17th
bilities and convergence time depending on the state variable International Conference on Probabilistic Methods Applied to Power
Systems (PMAPS), 2022, pp. 1–6.
initialisation. Additional benefits of the proposed GNN model [17] O. Kundacina, M. Cosovic, and D. Vukobratovic, “Robust and fast data-
are linear computational complexity regarding the number of driven power system state estimator using graph neural networks,” arXiv
state variables during the inference phase and the possibil- preprint arXiv:2206.02731, 2022.
[18] F. Kschischang, B. Frey, and H.-A. Loeliger, “Factor graphs and the
ity of distributing the inference computation across multiple sum-product algorithm,” IEEE Trans. Inf. Theory, vol. 47, no. 2, pp.
processing units. By testing the GNN on power systems of 498–519, 2001.
various sizes, we observed the prediction accuracy in the [19] J. De La Ree, V. Centeno, J. S. Thorp, and A. G. Phadke, “Synchronized
phasor measurement applications in power systems,” IEEE Trans. Smart
normal operating states of the power system and the sensitivity Grid, vol. 1, no. 1, pp. 20–27, 2010.
when encountering false data injection cyber attacks and input [20] Y. Weng, R. Negi, and M. Ilic, “Graphical model for state estimation
data loss due to communication irregularities. in electric power systems,” in Proc. IEEE SmartGridComm, Oct. 2013,
pp. 103–108.
Since the proposed GNN model generates predictions even [21] M. Cosovic and D. Vukobratovic, “Distributed Gauss–Newton method
for underdetermined SE systems of equations, it could be for state estimation using belief propagation,” IEEE Trans. Power Syst.,
applied to highly unobservable distribution power systems. vol. 34, no. 1, pp. 648–658, 2019.
[22] D. Pujol-Perich, J. Suárez-Varela, M. Ferriol, S. Xiao, B. Wu,
Another application of the proposed model could be the fast A. Cabellos-Aparicio, and P. Barlet-Ros, “Ignnition: Bridging the gap
and accurate initialisation of the nonlinear SE solver, resulting between graph neural networks and networking systems,” Netwrk. Mag.
in a hybrid approach that is both model-based and data-driven. of Global Internetwkg., vol. 35, no. 6, p. 171–177, nov 2021.

Authorized licensed use limited to: Dejan Vukobratovic. Downloaded 13

on May 24,2023 at 19:27:55 UTC from IEEE Xplore. Restrictions apply.