You are on page 1of 3

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/329116278

Exploring Application of Machine Learning to Power System Analysis

Article · November 2018

CITATIONS READS
0 302

1 author:

Mike Zhou
China Electric Power Research Institute
39 PUBLICATIONS   737 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Ph.D. Program at Tsinghua View project

Low Frequency Oscillation Analysis View project

All content following this page was uploaded by Mike Zhou on 22 November 2018.

The user has requested enhancement of the downloaded file.


1

Exploring Application of Machine Learning


To Power System Analysis
M. Zhou, Member, IEEE

 input/output training pairs, derive a general NN model for the


Abstract--InterPSS has been integrated with Google's underlying input/output relationship, which will be used to
TensorFlow machine learning (ML) engine to create an open predict output values for any new unseen input. This approach
platform for exploring application of ML to power system is sometimes also called supervised learning.
analysis. The open platform is introduced with a discussion of its In power system steady-state analysis, the problem usually
system architecture. Using a loadflow study case as an example, can be formulated as to solve a set of non-linear equations, as
neural network models are trained to predict the bus voltage, the
shown below:
branch active power flow and max branch power flow by N-1
contingency analysis when the bus injection power is given. The (1)
preliminary results of exploring application of ML to power The non-linear equation is commonly solved by some
system analysis are presented. numerical method, such as Newton-Raphson for loadflow
Index Terms--Deep Machine Learning, Power System
Analysis, InterPSS.
analysis. We would like to explore the possibility to use the
ML approach to solve the problem, e.g. to train an NN model
I. INTRODUCTION and use it to predict the output directly.

A FTER the AlphaGo wining game, there is a renewed


interest in artificial intelligence (AI), especially in the
area of Deep Machine Learning (ML). Google open-
sourced its TensorFlow ML engine[1], which is the brain
power behind AlphaGo. InterPSS[2] has been integrated with
TensorFlow to create an open platform for exploring
application of ML to power system analysis. The primary goal
of the open platform is to make the powerful TensorFlow ML
engine accessible to the power engineering community for the
power system analysis purpose. InterPSS is an open power Fig. 1. A Single Neural Network Layer
system simulation software development project. It was Neural networks typically consist of multiple layers, and the
created in 2005 and has been continuous in development by a signal path traverses from the input to the output layer of
group of volunteers. neural units. Back propagation is the use of forward
The use of automatic learning for real-time security stimulation to reset weights on the "front" neural units. This is
assessment was first envisioned in the late 1960s by Dy- done in combination with training or optimization where the
Liacco. Despite repetitive attempts, there are still no large- correct result (output in the training data) is known. A typical
scale industrial applications of machine-learning to off-line NN model is shown in the Fig.1. The output of previous layer
and on-line power system analysis. The main reason is that the [x] is weighted-summarized to produce [y] for the next layer,
existing automatic-learning methods were not powerful as illustrated in the equation below:
enough, and the amount of possible study cases was limited by (2)
capability of the available computation tools.[4]
An overview of machine learning is first presented. Then where, W - weight matrix, b - bias vector.
the open platform is introduced and followed by a sample TensorFlow is an open source software library by Google
study case to illustrate possible ways to use ML for power for numerical computation using data flow graphs. Nodes in
system analysis. the graph represent mathematical operations, while the graph
edges represent the multidimensional data arrays (tensors)
II. MACHINE LEARNING communicated between them. The flexible architecture allows
The main ideal of neural network (NN) based machine to deploy computation to one or more CPUs or GPUs, which
learning (ML) can be summarized as: Given a set of relevant makes solving large-scale ML problems possible.

III. AN OPEN PLATFORM


This work was supported in part by the State Grid of China under the
"Thousand Talents Plan" special research grant (5206001600A3).
InterPSS has been integrated with TensorFlow to create an
Mike Zhou is with the State Grid Electric Power Research Institute of open platform for exploring application of ML to power
China, Beijing, China (e-mail: mike.zhou@interpss.org). system analysis.[3] The platform system architecture is shown
2

in the figure below: Comparison of the largest branch power flow prediction in N-
1 CA with the accurate results are shown in Table-III. The
preliminary results of exploring the application of ML to
power system analysis are very interesting, especially the
largest branch flow prediction case, since there is no
underlying mathematic model to describe the relationship.
TABLE I
Fig. 2. The Platform System Architecture Bus Voltage Prediction (LF) Accuracy Comparison
There are three modules in the system: 1) TensorFlow ML
engine; 2) InterPSS power system analysis model service; and
3) training case generator. The power system model service
module provides service to the TensorFlow ML engine, 1) to
generate training data to train NN model; 2) provide service
for checking model prediction accuracy. TABLE II
An NN model is in general trained for certain purpose using Branch P Prediction (LF) Accuracy Comparison
a set of training data relevant to the problem to be solved. The
system architecture allows different training case generator to
be plugged-in for different model training purposes. The
training data could be obtained from field measurement or
generated from power system analysis model. The platform
system architecture supports both types of training case
generation. Certain types of training data, for example, power TABLE III
system contingency analysis data, can be only generated using Largest Branch P Prediction (CA) Accuracy Comparison
the analysis model approach.

IV. SAMPLE STUDY CASE


Power system loadflow analysis problem could be
formulated as follows:
(i = 1, ... n) (3)
V. SUMMARY
There is a renewed interest in Machine Learning (ML).
Where, V, θ are bus voltage magnitude and angle; P, Q bus
InterPSS has been integrated with the TensorFlow ML engine,
active, reactive power injection; g+jb branch admittance.
and an open platform for exploring application of ML to
In loadflow study, P, Q are given for PQ bus type; bus P, V
for PV bus type; bus V, θ for swing bus type. Bus V, θ are power system analysis has been created and introduced. Using
solved for PQ bus type; bus Q, θ for PV bus type; bus P, Q for a loadflow study case as an example, neural network models
swing bus type. The IEEE-14Bus loadflow test system is used are trained to predict bus voltage, branch active power flow
for the study. In the test system, power is flowing from the and largest branch power flow in N-1 contingency analysis
two generation buses (Buses-1,2) to the load bus area (Buses- when the bus power injection is given. The preliminary results
12,13,14) in the system. We want to train NN models to are very interesting and promising. The platform has been
predict network loadflow solution (bus voltage or branch open-sourced so that others could explore more complex
active power flow) and largest branch power flow value of power system analysis scenarios using the ML approach by
each branch in the test system in N-1 contingency analysis extending the platform. Also more capability and sample study
(CA), when the total power flow from the generation buses to cases will be added to the platform.
the load buses is varying, for example, by a load scaling factor
of 1.2 (20% increase from the base case). VI. REFERENCES
In the exercise, the load scaling factor changes in the range [1] Google, "TensorFlow - An open-source software library for Machine
[0.5~1.5] to generate 50 training cases evenly scattered in the Intelligence," available on the Web: www.tensorflow.org
range by the training case generator to train the NN models for [2] InterPSS, "The InterPSS Project," available on the Web:
the prediction. The TensorFlow Gradient Descent Optimizer is www.interpss.org.
[3] InterPSS, "TensorFlow Based Deep Machine Learning (ML) For Power
used for the NN model training with a learning rate of 0.001 System Analysis," available on the Web: www.github.com/interpss/
and 10K training steps. DeepMachineLearning
Results by the NN model prediction and comparison of [4] L. Wehenkel, "Machine-Learning Approaches to Power-System Security
Assessment," IEEE Expert Sept/Oct 1997, pp60-72.
them with accurate results by loadflow and N-1 CA are shown Mike Zhou (M’90) obtained his B.S. from Hunan University M.S. and PhD.
in Tables-I~III. Comparison of the bus voltage prediction from Tsinghua University. He was an assistant professor at University of
against the accurate loadflow results for a set of selected buses Saskatchewan 1990-1992, served as the VP in charging of power distribution
and three load scaling factors are shown in Table-I. system software development at EDSA Micro Corporation 1992-1997. He
was a Senior Computer System Architect with TIBCO Software Inc. 2000-
Comparison of the branch active power flow prediction 2014. He joint State Grid Electric Power Research Institute of China as a
against the accurate loadflow results are shown in Table-II. Chief Scientist in 2014, sponsored by the "Thousand Talents Plan" program.

View publication stats

You might also like