Towards a Graph Signal Processing Framework for

Modeling Power System Dynamics

Xinyue Hu, Zhi-Li Zhang
Department of Computer Science and Engineering, University of Minnesota

I. I NTRODUCTION a framework offers several advantages over the conventional

DAE models. In this data-driven framework, the model of
The conventional approaches for modeling dynamic systems a power system is implicitly learned, its “parameters” are
are based on state-space methods using differential algebraic inferred, and thus it can be learned continuously to adapt to
equations (DAEs). Such models not only require that the changing environments. In addition, the development of Phasor
system dynamics can be precisely captured and expressed Measurement Units (PMUs) increases the rate of measurement
in mathematical equations, but also need detailed knowledge collection significantly, which further justifies such a data-
about the system parameters. Even when such DAEs are avail- driven approach in real-world power systems. Our stochastic
able, no closed-form solutions are available, and numerical modeling framework supports various machine learning (ML)
solutions can be computationally expensive. As an example, methods for decomposing and approximating the “state-space”
modern power systems are typically large complex networks of power systems. This, in particular, allows us to view a
comprising of hundreds or even thousands of buses. The power system as a network with dynamic signals, where node
dimension of the mathematical models can easily reach the signals represent time-varying loads on buses (nodes) and edge
order of several thousands of state variables for dynamic signals represent power flows on branches (edges). In other
simulation, trajectory sensitivity analysis, control, and so forth. words, we can model the dynamics of a power system from a
Therefore, analyzing these extremely high-order DAEs poses graph signal processing (GSP) perspective (see Figure 1b for
a huge computational burden[1]. an illustration).
To tackle this problem, model reduction – namely, pro- Given such a perspective, several questions arise: i) Node-
ducing a reduced-order model that preserves the (dominant) level spatial requirements: to capture and model the signals at
dynamics of the original system approximately via a smaller individual nodes/edges over time, e.g. power loads on individ-
number of state variables, is necessary. Generally, two main ual buses, or power flows on individual branches, is it neces-
groups of techniques have been developed for power system sary to model and train the entire system as a whole by using
applications: techniques based on singular value decompo- the historical power load data (or power flow data) at all nodes,
sition (SVD) [2], [3], [4] and techniques based on Krylov or does it suffice to use only the data collected/measured at
projectors, also known as moment matching methods [5], [6], individual nodes/edges? ii)Temporal requirements: to capture
[7]. The first group of techniques is computationally expensive; and predict the input-output relations, namely, given the power
the second group is efficient but does not possess global loads on the buses, to predict the (current) power flows on the
error bounds [8]. Besides, most of the aforementioned model incident branches, is it good enough to simply use the current
reduction techniques typically start with a high-dimensional power loads for the prediction (i.e., a memoryless model),
model to be reduced, under the assumption that the knowledge or will the past power loads on the buses help improve the
of the system state-space model is fully specified. However, prediction of the current power flows on the branches (i.e., a
the accurate high-dimensional model is not always available. model with memory)? and iii) Area-level spatial requirements:
Modern power systems are highly dynamic and uncertain given a large power system which is often divided into multiple
due to the presence of various generation, transmission and areas for adaptive load control and other reasons, if we are
distribution devices (e.g., the increasing adoption of renewable only interested in only one area of the system, do we need
generation sources such as wind, solar, and energy storage). to model the entire system, namely, using the data from the
Therefore, building an accurate model for those systems can whole system to train the model, or is it good enough to simply
be a difficult task with only DAE methods, especially when employ only the data from the area of interest to train a model?
the system is partially known or with uncertain states. In practice, it may be the case that only data from this area
In this paper, we propose an alternative – potentially promis- of interest is available to us.
ing new – data-driven framework to model power systems: In this paper, we carry out a systematic study to answer
as depicted in Figure 1a, instead of using DAEs, we model these questions. We consider data-driven models with machine
a power system as a stochastic system by inferring input- learning algorithms in two different families: “classical” algo-
output relations, behavior dynamics or the “control laws” rithms based on graphical models and popular deep learning
governing the system dynamics based on observed system neural network models. We conduct a sequence of experiments
inputs/outputs, (partially) measured network state, etc. Such to study and compare the efficacy of these data-driven machine

978-1-7281-2723-1/19/$31.00 ©2019 IEEE

 (a) From DAE dynamic system models to ML (b) Dynamic system viewed as
based stochastic models (input/output) graph signals
Fig. 1: Data-driven Framework: A Graph Signal Processing Perspective Fig. 2: IEEE 14 Bus System

learning models and investigate the trade-offs and impacts of P Fij = (Vi2 − Vi Vj cosθij )Gij − Vi Vj sinθij Bij , QFij =
various factors involved in the design of such models. The −(Vi2 − Vi Vj cosθij )Bij − Vi Vj sinθij Gij , where P Fij and
results show the performances of different machine learning QFij respectively represent the active and reactive branch flow
models learned with different levels of temporal and spatial from bus i to bus j, Vi is the voltage magnitude at bus i, θij
system information, which serve as a baseline of data-driven is the voltage phase angle difference between bus i and bus j,
methods for learning reduced-order models of power system Gij and Bij respectively represent the real (conductance) and
dynamics, help further employ and design more advanced imaginary (susceptance) parts of the (i,k)th element in the bus
machine learning algorithms (e.g. Spatiotemporal Sequence admittance matrix. With increasing instrumentation in (smart)
Forecasting methods [9]) in our framework, and provide a dif- power grids, it is possible to measure and calculate power
ferent way to study power system modeling, model reduction flows on (some of the) branches. With current and historical
and segmentation. measurement data of power flows, is it possible to accurately
predict future power flows? Such models will be useful in the
II. P OWER S YSTEM DYNAMICS M ODELING : F ROM DAE operations of power systems [10], e.g., for power flow analysis
M ODELS TO ML M ODELS and stability control.
Using power systems as a case study, we investigate how System Input-Output Relation Modeling. Given the power
effective we can develop a data-driven framework to model demands (loads), power system dynamics are governed by
the dynamics (e.g., power loads and power flows) of a dynamic physical laws expressed by a set of differential-algebraic
system using machine learning (ML) models, where the sys- equations (DAEs) [10], where besides the (algebraic) power
tem dynamics can be viewed as a network of node/edge signals flow balance equations above, there are different equations
that evolve over time. Power system dynamics are governed describing the power generation dynamics as well as economic
by physical laws (e.g., Kirchhoff’s law) and are traditionally dispatch signals. Due to space limitation, we do not include
modeled with differential algebraic equations (DAEs) [10]. them here1 . In practice, however, it is often difficult for power
We will primarily use the IEEE 14 and 118 bus system (see system operators to build accurate DAE models to model
Figure 2 and Figure 6) as examples to explore the questions and capture the power system dynamics, especially for power
posed in the introduction. distribution systems. The system topology and network param-
eters (e.g. admittance matrix) may be out-of-date or inaccurate.
A. Power System Dynamics Modeling and Prediction
It is even more challenging to cope with rapid, random,
We focus on modeling and predicting the following system and large fluctuations in the growing distributed (renewal)
dynamics, with some justifications for data-driven models. energy resources. On the other hand, with increasing power
Power Load Dynamics Modeling and Prediction. Forecast- system instrumentation such as PMUs, smart meters as well as
ing power load is critical in operations of power systems, e.g., other IoT (Internet-of-Thing) devices, more measurement data
for planning and controlling power generations to meet power become available. We are therefore interested in examining
demands. The factors affecting power load are complicated, whether it is possible to directly predict power flows on
including meteorology status (e.g. temperature, humidity, and branches given the power loads on buses – we refer to this
wind speed), human activities (e.g. business hours, holidays), as input-output relation mapping.
economics, electricity price, demographic, etc. Obviously,
there are no accurate DAEs or other mathematical models to B. ML Models for Power Load & Flow Dynamics Prediction
delineate the dynamics of power loads due to the complex and As a power system is a dynamic network consisting of an
uncertain factors. interacting set of buses (nodes) and branches (edges), we treat
Branch Flow Dynamics Modeling and Prediction. Given
the power demands (loads on buses) and power genera- 1 In [11], a detailed DAE model is given for a New England 39-bus

tion settings to match the power demands, power flows on test power system, where the authors developed a semi-definite quadratic
programming approach for approximately solving the DAE models. The paper
the branches are governed by the Kirchhoff’s law, which also highlights several factors of uncertainties that are not captured by the
can be expressed using a set of algebraic equations [10]: DAE model.
 (a) BNF performance (b) LSTM performance (a) BNF performance (b) LSTM performance
Fig. 3: Sample CDFs of load prediction errors: local data Fig. 4: Sample CDF of load prediction errors: global data

III. EXPERIMENTAL RESULTS

it as a network (graph) of signals, with (time series of) power We first describe the experiment setups and then report the
loads as node signals and (time series of) power flows as edge key results and takeaways from our experiments.
signals. We employ and compare two kinds of data-driven,
machine-learning approaches for modeling and predicting the A. Experiment Setup
node and edge signals: (probabilistic) graphical models from We test our data-driven models for the above power system
“classical” machine learning and deep learning models based case studies on IEEE 14 and 118 bus systems. We use real-
on neural networks. The graphical models explicitly make world load data from GEFCom 2012 [15] as power loads. The
assumptions (e.g., a mixture of Gaussian distributions) on how dataset records hourly real load power for a US utility from
data are generated and learn the parameters of the models. In Jan 1, 2004 to July 7, 2008. To simulate the branch flow data,
contrast, neural networks are “agnostic” and more general, we use the GEFCom power load data as the input and employ
relying on the fact that multi-layer neural networks with MATPOWER [16] to solve a set of power-balance equations
nonlinear activation functions are universal function approxi- at each time slot to obtain the associated branch flows.
mators [12].
B. Power Load and Flow Modeling & Prediction Results
Bayesian Network Flow Models. Building on TribeFlow
Two different spatial scales of historical power load data
[13], which is designed for the next-item recommendation, we
are used to build ML models for tracking the dynamics of
develop Bayesian Network Flow (BNF) models. BNF takes
power loads and flows. In the “local” scale setting, we use
a set of buses/branches and their corresponding loads/flows
only the historical data from a single bus (i.e. the time series
trajectories as input, and outputs a model for bus/branch-
of past power loads on the bus) to learn and predict the current
personalized load/flow prediction in the next time slot. Since
and future power loads on the said bus. In the “global” scale
power loads/flows change over time, BNF models their fluctua-
setting, we use the historical data on all buses (i.e. the time
tion patterns as latent environments. Given its load/flow trajec-
series of past power loads on all the buses) to learn and
tories, each bus/branch predicts the next load/flow according
predict the current and future power loads on all the buses
to an environment-unique probability transition matrix. The
simultaneously. Likewise, in the “local” data setting for power
jump from one fluctuation pattern (environment) to the next
flow dynamics prediction, we use only the historical data from
one follows the Dirichlet distribution assumption.
a single branch to train ML models, and predict the future
Deep Neural Network Models. Given the scale and complex- flow on the same branch. For the “global” data setting, we
ity of real-world power systems, deep learning neural networks use the historical data from all branches to train ML models,
work better when providing enough data and computation and predict the future flow on all branches. In all cases, we
resource. The first option is a Multilayer Perceptron (MLP) train BNF and LSTM models to forecast the future power
neural network, a feed-forward neural network. The second is loads/flows on the IEEE 14 bus system.
Long-Short-Term Memory (LSTM) Recurrent Neural Network For power load dynamic prediction experiments, the Mean
[14], designed for capturing both short-term and long-term Absolute Percentage Errors (MAPE) of BNF and LSTM in
dependencies. Unlike feed-forward neural networks, LSTM the local scale setting are 0.0965 and 0.0268 respectively.
has feedback connections that allow information to persist and MAPE of BNF and LSTM in global scale are 0.109 and
thus have memory. 0.0763 respectively. The absolute percentage error distribution
Using these ML models, we conduct a series of experiments of randomly sampled buses are illustrated in Figure 3 and
to investigate whether it is feasible to capture the local Figure 4. Table I shows the corresponding upper bounds of
dynamics, e.g., power load on a single or power flow on 95% load prediction errors. According to the above results,
a single branch, using only the local data, or to obtain a LSTM achieves much higher accuracy than BNF model when
partial view of the system using only the partial data. In given only local historic bus load information. As for in the
other words, we study how much the global data or whole global scale setting, LSTM achievers generally better results
system information helps model local dynamics and partial than BNF in global scale data setting, but, in some cases, its
system dynamics, and how much memory (recent or long past performance (e.g., on bus 1) gets far worse, perhaps due to
history) helps improve the prediction results. higher model complexity, which means that more parameters
 TABLE I: 95% Power Load Prediction Errors
Local Data Local Data Global Data Global Data
Bus ID BNF LSTM BNF LSTM
1 0.26 0.069 0.29 0.4
6 0.047 0.017 0.044 0.016
7 0.057 0.018 0.053 0.027
11 0.11 0.038 0.10 0.065
12 0.17 0.05 0.24 0.037

TABLE II: 95% Branch Flow Prediction Errors

Local Data Local Data Global Data Global Data (a) MLP performance (b) LSTM performance
Branch ID
1
BNF
0.0055
LSTM
0.0035
BNF
0.0054
LSTM
0.011
Fig. 5: Power flow prediction on branch 2
3 0.012 0.054 0.084 0.014
5 0.032 0.013 0.034 0.02
6 0.44 0.17 0.81 0.41
10 0.04 0.02 0.04 0.024

need to be trained. Besides, global information does not seem

to help power load prediction significantly, probably due to
the fact that loads on buses are rather independent. Area3

For the power flow prediction experiments, the MAPE of

BNF and LSTM in local data scale setting are 0.0794 and Area1 Area2

0.0317 respectively; in global data scale setting are 0.134 and

0.0718 respectively. Table II shows the corresponding upper
bounds of 95% flow prediction errors. Similar to the perfor-
mance of power load dynamics modeling, LSTM achieves
Fig. 6: IEEE 118 bus system divided into 3 areas
much higher accuracy than BNF model; global information
does not seem to help branch flow prediction significantly,
especially for LSTM model. to the temporal power load and generation pattern in power
Main Take-Aways: Global information, in general, leads to grids.
worse performance for power load and flow modeling, which Power Load to Branch Flow Mapping Inferring with
might be due to higher model complexity, various sample data Partial vs. Whole System View. In this set of experiments,
distributions and the relatively small amount of training data. different spatial scales of data have been used to learn the
mapping rules from power loads to branch flows. We use
C. System Input-Output Relation Modeling and Prediction the larger IEEE 118 bus system, where we divide it into
Results three areas, see Figure 6. We assume that we are only
Power Load to Branch Flow Mapping Inferring without interested in area 1, and investigate whether it is possible to
vs. with Memory. Two different time scales of data have been model and predict the branch flows in area 1 using only data
used to learn the mapping rules from power loads to branch collected from this area (which may be the only available data)
flows on IEEE 14 bus system. The first mapping inferring or it is necessary to use the data collected from the entire
method is MLPNN, which has “no memory” and is trained to system. Due to the space limitation, we will not be able to
predict the current branch flows on all the branches, given the provide detailed results. The key Take-Away is summarized
current power loads on all the buses in a system. On the other as: Prediction accuracy is in fact often better with data from
side, the other mapping inferring method is LSTM recurrent partial system observations than whole system, probably due
neural network, which has “memory” and is trained to predict to the trade-offs of model complexity and training data size
the current power flows on all the branches, given the current and the fact that large power systems can be segmented into
and past historical power loads on all the buses in the system. static zones [17].
The intuition to employ LSTM is that the temporal pattern of
power systems may improve learning the mapping rules. IV. C ONCLUSION
The Mean Absolute Percentage Errors (MAPE) of MLPNN
and LSTM are 0.0288 and 0.0105 respectively. The overall In this work, we presented a data-driven, machine-learning
prediction performance of LSTM is better than MLPNN. Fig- framework for modeling power systems as stochastic systems.
ure 5 shows the power flow predictions on a randomly selected Using power load and flow modeling as case studies, we
branch. The blue line is the true power flow; the red line is conducted a sequence of experiments to systematically inves-
the predictions on the training dataset; the yellow line is the tigate several important questions in the design and training of
predictions on the testing dataset. On branch 2, MLPNN model ML models. Our experimental results show that a data-driven
yields “acceptable” performance, while LSTM model with framework based on machine learning methods, especially
memory achieves far better prediction accuracy. Main Take- deep learning neural networks, for modeling power systems
Aways: Models with memory (e.g., LSTM) generally perform is feasible and viable. But there are still many unquestioned
better than models without memory (e.g., MLP), probably due research challenges in making them effective in practice.
 ACKNOWLEDGMENT
The research was supported in part by US DoD DTRA grant
HDTRA1-14-1-0040, and NSF grants CNS 1831140, 1814322
and 1901103.
R EFERENCES
[1] K. J. Åström and P. R. Kumar, “Control: A perspective.” Automatica,
vol. 50, no. 1, pp. 3–43, 2014.
[2] M. Safonov, R. Chiang, and D. Limebeer, “Optimal hankel model
reduction for nonminimal systems,” IEEE Transactions on Automatic
Control, vol. 35, no. 4, pp. 496–502, 1990.
[3] S. Lall and C. Beck, “Error-bounds for balanced model-reduction of
linear time-varying systems,” IEEE Transactions on Automatic Control,
vol. 48, no. 6, pp. 946–956, 2003.
[4] F. D. Freitas, J. Rommes, and N. Martins, “Gramian-based reduction
method applied to large sparse power system descriptor models,” IEEE
Transactions on Power Systems, vol. 23, no. 3, pp. 1258–1270, 2008.
[5] D. Chaniotis and M. Pai, “Model reduction in power systems us-
ing krylov subspace methods,” IEEE Transactions on Power Systems,
vol. 20, no. 2, pp. 888–894, 2005.
[6] C. Wang, H. Yu, P. Li, C. Ding, C. Sun, X. Guo, F. Zhang, Y. Zhou,
and Z. Yu, “Krylov subspace based model reduction method for transient
simulation of active distribution grid,” in 2013 IEEE Power & Energy
Society General Meeting. IEEE, 2013, pp. 1–5.
[7] Z. Zhu, G. Geng, and Q. Jiang, “Power system dynamic model reduction
based on extended krylov subspace method,” IEEE Transactions on
Power Systems, vol. 31, no. 6, pp. 4483–4494, 2016.
[8] A. C. Antoulas, Approximation of large-scale dynamical systems. Siam,
2005, vol. 6.
[9] X. Shi and D.-Y. Yeung, “Machine learning for spatiotemporal sequence
forecasting: A survey,” arXiv preprint arXiv:1808.06865, 2018.
[10] P. Kundur, N. J. Balu, and M. G. Lauby, Power system stability and
control. McGraw-hill New York, 1994, vol. 7.
[11] H. Choi, P. J. Seiler, and S. V. Dhople, “Propagating uncertainty
in power-system dae models with semidefinite programming,” IEEE
Transactions on Power Systems, vol. 32, no. 4, pp. 3146–3156, 2016.
[12] G. Cybenko, “Approximation by superpositions of a sigmoidal function,”
Mathematics of control, signals and systems, vol. 2, no. 4, pp. 303–314,
1989.
[13] F. Figueiredo, B. Ribeiro, J. M. Almeida, and C. Faloutsos, “Tribeflow:
Mining & predicting user trajectories,” in Proceedings of the 25th
international conference on world wide web. International World Wide
Web Conferences Steering Committee, 2016, pp. 695–706.
[14] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[15] T. Hong, P. Pinson, and S. Fan, “Global energy forecasting competition
2012,” 2014.
[16] R. D. Zimmerman, C. E. Murillo-Sánchez, and R. J. Thomas, “Mat-
power: Steady-state operations, planning, and analysis tools for power
systems research and education,” IEEE Transactions on power systems,
vol. 26, no. 1, pp. 12–19, 2010.
[17] A. Marot, S. Tazi, B. Donnot, and P. Panciatici, “Guided machine
learning for power grid segmentation,” in 2018 IEEE PES Innovative
Smart Grid Technologies Conference Europe (ISGT-Europe). IEEE,
2018, pp. 1–6.