IEEE Power & Energy Society TECHNICAL REPORT

September 2023 PES-TR112

Practical Applications of
Artificial Intelligence and
Machine Learning in Power
System Protection and
Control
PREPARED BY THE
Power System Relaying and Control Committee
Subcommittee C – System Protection
Working Group C43

© IEEE (YEAR) The Institute of Electrical and Electronics Engineers, Inc.

No part of this publication may be reproduced in any form, in an electronic retrieval system or otherwise, without the prior written permission of the publisher.
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

THIS PAGE LEFT BLANK INTENTIONALLY

2
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

WORKING GROUP C43

PRACTICAL APPLICATIONS OF ARTIFICIAL INTELLIGENCE /
MACHINE LEARNING IN POWER SYSTEM PROTECTION AND
CONTROL

Chair: Yi Hu
Vice Chair: Adi Mulawarman
Secretary: Zheyuan Cheng

Members and Contributors

Abder Elandaloussi
Alex Apostolov
Ali Bidram
Athula Rajapakse
Carolina Arbona
Dan Sabin
Jayaprakash Ponraj
Jean Raymond
Jörg Blumschein
Juan F. Piñeros S.
Matthew Reno
Nirmal Nair
Ratan Das
Robert Fowler
Sebastien Billaut
Sukumar Brahma
Sukumar Kamalasadan
Vahid Madani
Yu Liu
Yujie Yin

3
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

KEYWORDS

Artificial Intelligence (AI), Control, Machine Learning (ML), Power System, Practical
Application, Protection, Relaying

4
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

CONTENTS

1 INTRODUCTION ....................................................................................................... 1
2 SCOPE 2
3 AI/ML Technology ...................................................................................................... 2
3.1 Reasoning through bodies of knowledge ................................................................ 3
3.2 Learning by experience ........................................................................................ 6
3.3 Learning by evolution .......................................................................................... 8
3.4 Deep Learning .................................................................................................... 9
3.5 Hybrid Intelligent Systems ................................................................................. 10
4 AI/ML for Protection & Control .................................................................................. 12
4.1 Drivers ............................................................................................................. 12
4.2 Limitations of existing P&C technology ............................................................... 13
4.3 Opportunities for Artificial Intelligence in P&C .................................................... 14
4.4 Considerations before using AI in protection ........................................................ 15
4.5 Specific areas of AI/ML P&C application ............................................................ 16
4.5.1 Asset Protection ......................................................................................... 17
4.5.2 Network Protection ..................................................................................... 17
4.5.3 Post Event Analysis .................................................................................... 18
4.5.4 Summary of Potential P&C Areas for Application of AI/ML ........................... 18
5 APPLICATION ILLUSTRATIONS ............................................................................ 19
5.1 Existing Applications......................................................................................... 19
5.1.1 AI to Detect Secondary Arc ......................................................................... 19
5.1.1.1 The problem: secondary arc detection ................................................. 19
5.1.1.2 Implemented solution .......................................................................... 22
5.1.1.3 Training of the neural network ............................................................ 23
5.1.1.4 Results ................................................................................................. 23
5.1.2 Hybrid Intelligent Model for Protection Settings Optimization ......................... 25
5.1.2.1 The Problem: Finding Optimal Settings for Distance Protection ........ 25
5.1.2.2 Hybrid Intelligent System for Distance Protection Settings Optimization
25
5.1.2.3 Test Results.......................................................................................... 30
5.1.2.4 Real Life Application Example ........................................................... 30

5
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

5.2 Emerging Applications ...................................................................................... 33

5.2.1 High-Impedance Fault Detection with Artificial Neural Networks .................... 33
5.2.2 Use of synchrophasor and ML for Early Detection of Long-term Voltage Instability
35
5.2.2.1 Introduction ......................................................................................... 35
5.2.2.2 Problem Statement ............................................................................... 36
5.2.2.3 Real-time Voltage Stability Monitoring System ................................. 37
5.2.2.4 Feature Selection and Offline Training Process .................................. 37
5.2.2.5 Real-Time Implementation and Testing .............................................. 40
5.2.3 Detecting Relay Misoperations Using Field Data ............................................ 41
5.2.4 Post-Event Analysis .................................................................................... 47
5.2.5 Application of AI/ML in Travelling Wave Protection...................................... 49
5.2.5.1 TW Protection Background ................................................................. 50
5.2.5.2 TW Protection of AC Distribution Systems Using ML [98]: .............. 51
5.2.5.3 TW Protection of DC Systems Using ML [99]: .................................. 54
5.2.6 AI Based Wide Area Control with Windfarm ................................................. 61
5.2.6.1 Introduction ......................................................................................... 61
5.2.6.2 Problem Statement ............................................................................... 61
5.2.6.3 Machine Learning Process................................................................... 62
5.2.6.4 Implementation and Testing ................................................................ 63
5.2.7 Transient Instability Predictions ................................................................... 64
5.2.7.1 Introduction ......................................................................................... 64
5.2.7.2 Problem Statement ............................................................................... 64
5.2.7.3 Rotor angle stability prediction scheme .............................................. 64
5.2.7.4 Application to IEEE 39 bus test system .............................................. 65
5.2.7.5 Conclusions ......................................................................................... 67
6 PRACTICAL APPLICATION CONSIDERATIONS ..................................................... 68
6.1 Application Implementation ............................................................................... 68
6.1.1 Data Types, Sources, Formats, and Management ............................................ 68
6.1.1.1 Data Types ........................................................................................... 68
6.1.1.1.1 Waveform Samples ......................................................................... 68
6.1.1.1.2 Digital Status Values ....................................................................... 69
6.1.1.1.3 RMS Values .................................................................................... 70
6.1.1.1.4 Phasor Values .................................................................................. 70

6
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

6.1.1.1.5 Time Series (Data Logs) ................................................................. 71

6.1.1.2 Event Categorization ........................................................................... 72
6.1.1.3 Data Sources ........................................................................................ 72
6.1.1.4 Data Formats........................................................................................ 73
6.1.1.5 Data Management and Storage ............................................................ 73
6.1.2 Data Structures ........................................................................................... 73
6.1.2.1 Introduction ......................................................................................... 73
6.1.2.2 Defining Data Structures ..................................................................... 74
6.1.2.3 Data Management and Relationships .................................................. 75
6.1.3 Computational Platforms ............................................................................. 75
6.1.3.1 Hardware ............................................................................................. 76
6.1.3.1.1 Hardware Options ........................................................................... 76
6.1.3.1.1.1 Edge Hardware ......................................................................... 76
6.1.3.1.1.2 Centralized Processing ............................................................. 76
6.1.3.1.1.3 Cloud Computing ..................................................................... 76
6.1.3.1.2 Hardware Considerations ................................................................ 76
6.1.3.2 Software ............................................................................................... 77
6.1.3.2.1 Software Options ............................................................................. 77
6.1.3.2.2 Software Considerations ................................................................. 77
6.1.4 Testing and Validation ................................................................................ 77
6.1.4.1 Learning. Validation and Testing in Machine Learning (ML)
Architecture........................................................................................................... 77
6.1.4.2 Testing ................................................................................................. 78
6.1.4.3 Parameter and Hyperparameters .......................................................... 79
6.1.4.4 Estimating Performance ...................................................................... 79
6.1.4.5 Cross-validation Loop ......................................................................... 80
6.1.4.6 Field Implementation and Testing ....................................................... 82
6.1.5 User Training ............................................................................................. 83
7 Use of AI/ML for Protection: Risks, Challenges and Acceptance Criteria ......................... 83
7.1 Risks ............................................................................................................... 83
7.2 Challenges........................................................................................................ 85
7.3 Criteria for Accepting AI/ML Applications in Field............................................... 87
7.3.1 Planning .................................................................................................... 88
7.3.2 Effective Communication ............................................................................ 88

7
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

7.3.3 Progressive Implementation ......................................................................... 88

7.3.4 Process Changing ....................................................................................... 88
7.3.5 Main Risk Scenarios Testing ........................................................................ 89
7.3.6 Redundancy in Case of Failure ..................................................................... 89
8 CONCLUSIONS AND RECOMMENDATIONS .......................................................... 89
9 APPENDIX A – BIBLIOGRAPHY ............................................................................. 91

LIST OF FIGURES
Figure 3-1: Basic elements of an expert system ................................................................. 4
Figure 3-2: (a) Components of a fuzzy inference system (b) Examples of membership
functions .............................................................................................................................. 5
Figure 3-3: (a) Model of an artificial neuron (b) Example of a feedforward neural network
............................................................................................................................................. 6
Figure 3-4: Types of Machine learning algorithms ............................................................ 8
Figure 3-5: Scikit-learn Algorithm Cheat Sheet [112] ....................................................... 8
Figure 3-6: Conceptual representation of a deep neural network for image recognition . 10
Figure 3-7: General framework for hybrid architectures [78] .......................................... 11
Figure 3-8: Typical architectures for hybrid intelligent system [82] ................................ 11
Figure 4-1: Optimal setting area for reliability ................................................................. 14
Figure 4-2: Main stages of the P&C process that AI could impact .................................. 15
Figure 5-1: Simplified diagram showing the secondary arc after single phase tripping .. 20
Figure 5-2: Simplified diagram showing the secondary arc extinction after single phase
tripping .............................................................................................................................. 20
Figure 5-3: Unsuccessful reclosing if secondary arc is not extinguished after single phase
tripping .............................................................................................................................. 21
Figure 5-4: Successful reclosing if secondary arc is extinguished after single phase tripping
........................................................................................................................................... 22
Figure 5-5: Structure of the secondary arc detection function .......................................... 22
Figure 5-6: Learning process of the neural network used for secondary arc detection .... 23
Figure 5-7: Detection of secondary arc extinguishing after single phase tripping ........... 24
Figure 5-8: Secondary arc not extinguishing during the dead time of autoreclosure ....... 24
Figure 5-9: Detection of re-arcing after extinguishing of secondary arc .......................... 24
Figure 5-10: Problem of Finding Optimal Distance Relay Settings ................................. 25
Figure 5-11: HIS for Distance Protection Settings Optimization ..................................... 26
Figure 5-12: Solution structure for zones 1, 2, 3 forward and 4 reverse .......................... 27
Figure 5-13: Differential evolution modified algorithm of HIS ....................................... 27
Figure 5-14: Optimization Model used in the HIS ........................................................... 30
Figure 5-15: HIS test system results ................................................................................. 30
Figure 5-16: Real life application example grid ............................................................... 31
Figure 5-17: Zone 1 incursion problem for of real-life application example grid ............ 32
Figure 5-18: HIS Execution final solution ........................................................................ 33
Figure 5-19: Typical ANN structure. ................................................................................ 34
Figure 5-20: ANN output – relative HIF probability. ....................................................... 35

8
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 5-21: (a) P-V curve and Loadability Margin (b) Thévenin equivalent circuit of the
network at load bus-i ......................................................................................................... 36
Figure 5-22: Proposed real-time VSM system.................................................................. 37
Figure 5-23: Correlation heat map of candidate VSIs and LM for (a) IEEE 14 bus system
and (b) IEEE 118 bus system ............................................................................................ 38
Figure 5-24: (a)LM prediction accuracy comparison for IEEE 118 bus test system, (b)
Effect of weightages on the accuracy of ensemble output ................................................ 39
Figure 5-25: Experimental Setup ...................................................................................... 40
Figure 5-26: Examples of (a) voltage stability monitoring dashboard, and (b) variation of
real-time predicted LM with system changes/events for IEEE 14 bus test system .......... 40
Figure 5-27: Supervisory protection and dynamic event visualization. ........................... 41
Figure 5-28: Energy change pattern due to (a) loss of generator mechanical input, and (b)
loss of load ........................................................................................................................ 46
Figure 5-29: FSM Model of the Post-Event Analysis ...................................................... 48
Figure 5-30: Fault Tree Construction for Transformer Bank X ....................................... 49
Figure 5-31: MRA block diagram..................................................................................... 51
Figure 5-32: AI-based traveling wave fault location algorithm........................................ 52
Figure 5-33: Schematic of the system ............................................................................... 53
Figure 5-34: Fault location prediction errors relative to fault distance ............................ 54
Figure 5-35 Simple DC microgrid system ........................................................................ 54
Figure 5-36: Cable configuration of the DC system in Figure 5-35 ................................. 55
Figure 5-37: Parseval energy values for different fault locations on the cable of DC system
with 1 MHz sampling frequency in Figure 5-35: (a) MRA’s level 1; (b) MRA’s level 2; (c)
MRA’s level 3. .................................................................................................................. 56
Figure 5-38: Parseval energy values for different fault locations on the cable of the DC
system with 2 MHz sampling frequency. ......................................................................... 56
Figure 5-39: AI-based traveling wave fault classification and location algorithm for DC
systems .............................................................................................................................. 57
Figure 5-40: DC microgrid diagram ................................................................................. 59
Figure 5-41: Regression verification plots for C4: (a) bolted PP faults; (b) resistive PP
faults; (c) bolted PG faults; (b) resistive PG faults ........................................................... 60
Figure 5-42: Wide Area System Centric Controller (WASSCO) design [100] ................ 62
Figure 5-43: Reinforcement learning process [101] ......................................................... 62
Figure 5-44: Implementation platform.............................................................................. 63
Figure 5-45: Visualization for a two-area power grid with fault on area 2 [100] ............. 63
Figure 5-46: (a) Variations of the voltage magnitudes, rotor angles, and generator speeds
during a contingency (b) Synchronized sampling of signals using PMUs, and (c)
Arrangement of the transient stability prediction scheme ................................................ 65
Figure 5-47: (a) Training process (b) Rotor angle prediction accuracy with different
predictors and data window lengths.................................................................................. 66
Figure 6-1: Example Waveform Chart Showing Voltage Samples and Current Samples 69
Figure 6-2: Example Chart Showing Digital Status Values with Voltage and Current
Waveform Samples ........................................................................................................... 69
Figure 6-3: Example RMS Values Chart with Voltage and Current Samples Derived from
Waveform from Figure 6-2 ............................................................................................... 70

9
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 6-4: Example Phasor Chart Showing Voltage Phasor (Per Unit) and Current Phasors
(Amps) .............................................................................................................................. 71
Figure 6-5: Example time series chart for RMS Voltage on Phase AN, Phase BN, and Phase
CN ..................................................................................................................................... 72
Figure 6-6: Example of a data structure ............................................................................ 74
Figure 6-7: Coordinated learning framework based on offline/online machine learning . 78
Figure 6-8: Train/test splits ............................................................................................... 80
Figure 6-9: Cross-validation process ................................................................................ 81
Figure 6-10: Generic Field Implementation Computing Infrastructure ............................ 82
Figure 7-1: Pros and Cons of AI/ML based methods for fault-related applications ......... 86

LIST OF TABLES
Table 4-1: Potential areas for application of AI/ML ........................................................ 18
Table 5-1: Comparison Optimized vs Conventional Setting line bay PS to JM 115kV ... 32
Table 5-2: Candidate VSIs for MLMs (see [89] for VSI definitions/references) ............. 38
Table 5-3: Descriptions of different MLMs in the ensemble ........................................... 39
Table 5-4: Clusters formed by using agglomerative hierarchical clustering approach. ... 43
Table 5-5: Disturbance events identified for initially unknown clusters. ......................... 44
Table 5-6: Number of simulated events for the seven disturbance types under study. .... 44
Table 5-7: Training (3495 simulated events from 7 disturbance types), testing (81 actual
events with FLT and GL) .................................................................................................. 45
Table 5-8: Training (90% of all 3495 simulated events with 7 disturbance types), testing
(10% of all 3495 simulated events with 7 classes); ten-fold cross validation. ................. 46
Table 5-9: Fault type classification accuracy.................................................................... 53
Table 5-10: Mean and STD of prediction errors for fault location. .................................. 54
Table 5-11: Cable lengths and fault locations................................................................... 59
Table 5-12: Fault location estimation error for DC Microgrid 1 ...................................... 60
Table 5-13: Comparison of different ML techniques for Cable C2.................................. 61
Table 5-14: Overall prediction accuracy........................................................................... 66
Table 5-15: Prediction accuracy under topologies not contained in training data ............ 67
Table 5-16: Prediction accuracy with noisy data after re-training the classifiers with noisy
data .................................................................................................................................... 67

10
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

THIS PAGE LEFT BLANK INTENTIONALLY

11
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

1 INTRODUCTION
In practical terms, intelligence can be defined in many ways: advanced comprehension, a
capacity to further the existing reasoning, demonstrative knowledge, and learned decision-
making. Intelligence can be demonstrated using machines, similar to the natural
intelligence shown by humans. Adaptive development is required at various stages, such
as cognition, manipulation, rationalization, communication, and reaction to any common
transaction. Here, the continuous learning experience facilitates the challenge of
automated enhancement in overall system performance over time.
Artificial intelligence (AI) is the simulation of human intelligence in machines that think
and act like humans. An effective AI application requires many skill sets such as cognition,
manipulation, rationalization, communication, and reaction to be incorporated into the
scheme.
AI utilizes Machine Learning (ML) and other associated techniques such as Heuristics to
resolve real challenges. Various computational tools used for implementing these skill sets
include search and optimization, artificial neural network, fuzzy logic, probabilistic
methods for uncertain reasoning, reinforcement learning, and other supervised and
unsupervised learning methods. ML is a subfield of artificial intelligence that enables the
gathering and analyzing volumes of data to extract representative features based on
appropriate training (learning) and develop an equation or algorithm for deriving useful
information or action. As an example, text recognition within images is now a commonly
used tool (OCR, optical character recognition) that is powered by ML. OCR models can
search for the pertinent information using representative features to identify text
predictively rather than programmatically.
ML or its earlier vintages of ‘Artificial Neural Network’ has been in practice within the
power system industry, especially in operations, workforce management, and planning
over a considerable period of time. Some examples are forecasting, optimization,
scheduling, and unit commitment. Due to the recent digital transformation progression in
the areas such as Cloud Computing (CC), Internet of Things (IoT), and the use of ML to
expand the earlier techniques, an upsurge of AI or ML (sometimes used interchangeably,
also in this report) is conveying additional significance in many power system applications.
This report aims to introduce significant AI/ML technical concepts to a Protection and
Control (P&C) knowledge base and audience. Providing the foundation for a methodical
approach to incorporate this new technology within our industry while displaying
representative examples of possible applications.
This report initiates the reader on current developments and futuristic ideas in AI/ML, to
spark creativity in advancing P&C applications within Power System Industry. For
example:

• AI/ML can help detect and locate high impedance faults faster.

1
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

• AI/ML can assist in predicting and isolating potential component failures by

utilizing partial discharge sensor data.
• AI/ML can be used to evaluate challenges to the dynamic stabilities of the power
network and support decision-making by using Phasor Measurement and Control
Unit data and architecture, thus mitigating the stability control challenges due to
increasing low inertia resources.
The following section describes the report's detailed purpose, scope, and limitations as the
topic of artificial intelligence in the power system are very broad.

2 SCOPE
This report covers the practical applications of AI/ML in the protection and control of
power systems. The process of applying AI/ML for solving protection or control challenges
in the power system involves the design, development, validation, integration, field testing,
and deployment of AI/ML models. All these applications will focus on reliability,
availability, dependability, security, speed, and accuracy to support the operation of a
dynamic power system. The report will primarily provide insights from data collection to
detailed analytics and informed decision-making to support asset and network protection
and anticipate potential disturbances to lessen the impacts. The protection and control
applications in this report will be limited to the protection of the electric network or the
power assets.
The focus will be AI/ML applications in P&C with updates on:
● Basics of the AI/ML technology
● Practical applications in use and Emerging application areas
● Challenges of applying the technology in power system protection and control
● Considerations in developing, validating, field testing, and implementing practical
applications
● Risks, challenges and acceptance criteria for the use of AI/ML for protection
The report is not intended to be used as a technical guide but as a resource that provides
recent advances in AI/ML applications in the P&C area of power system.

3 AI/ML Technology
Since the beginning of computing technology, the idea of imitating human capabilities has
been a goal of computational development. In 1950, Alan Turing created a machine to do
a complex task faster than a human being, and established the principles of machine
intelligence [72]. There have been several definitions of artificial intelligence through its
evolution. Any human creation or effort to imitate a human skill in executing a complex
task in a wide range of environments to achieve a specified objective is considered artificial
intelligence [74].
To perform complex tasks, several types of algorithms have been developed. Based on the
type of human skills that these algorithms or techniques imitate; they can be grouped as:

2
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

• Reasoning through bodies of knowledge – imitates the human’s ability to apply

prior knowledge in similar situations via inference and reasoning.
• Learning by experience - imitates the human’s ability to learn by experience to
perform pattern recognition, classification, regression tasks, etc.
• Learning by evolution – imitates human’s ability to learn by trial and error by
considering actions and an objective that is tested for success.
The role of AI/ML technology in a complex system can vary depending on the nature of
the application, and some of the common elementary functions performed by the
technology include:

• Pattern recognition: application of one or more techniques or algorithms to find,

based on some criteria, a similar ‘pattern’ or ‘class type’ in a set of data.
• Classification: identification of the ‘class type’ of given data, according to a
predefined set of class types.
• Clustering: grouping of data based on the similarity of their characteristics
• Regression and function approximation: a statistical or learning process to create
a functional relationship between two or more correlated variables from data, often
to predict values of one variable when given values of the others.

3.1 Reasoning through bodies of knowledge

Reasoning based on knowledge is a key skill of human intelligence. Starting with
conventional logic based on if-then clauses, algorithms and techniques have been
developed to imitate this skill. These systems can capture the scarce expertise and/or
difficult-to-use sources of knowledge to perform specific tasks in a much faster and more
consistent way than a human user. The main approaches to using knowledge-based
reasoning are expert systems, case-based reasoning, and Fuzzy logic.
Expert Systems
An expert system consists of a knowledge base, an inference engine, and a user interface,
as shown in Figure 3-1. The process of extracting knowledge from the subject experts and
constructing a knowledge base is known as knowledge engineering. The knowledge base
stores both factual knowledge (widely accepted information) and heuristic knowledge
(information derived from an expert’s judgment, practice, and ability to guess and evaluate)
in the form of a set of if-then rules such as:
R1: IF (I > 2 pu) THEN (I is High) R2: IF (V < 0.9 pu) THEN (V is Low)
R3: IF (I is High) THEN (Fault = True) R4: IF (V is Low) THEN (Fault = True)

3
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Human Expert Knowledge Engineer Expert System User: Human or Machine

Inference engine
(facts/data)
Acquisi�on of inputs
Knowledge
Knowledge User
factual and heuris�c representa�on Applica�on
base interface
knowledge (if-then rules)
output
(recommenda�on)

Figure 3-1: Basic elements of an expert system

The inference engine acquires and manipulates the knowledge from the knowledge base to
arrive at a particular solution. This is achieved by applying rules repeatedly to the presented
facts (inputs) and resolving conflicts when multiple rules are applicable to a particular case.
The inference engine may add new knowledge to the knowledge base if required.
Depending on the nature of the problem, inference engines use either forward chaining or
backward chaining of rules to recommend a solution.
In forward chaining, the inference engine attempts to conclude based on given facts; rules
are applied to derive new facts as conclusions (forward chaining of rules) until the
requested final goal fact is found. It is data-driven and commonly used to solve more open-
ended problems such as prediction, control, design, or planning.
In backward chaining, the inference engine attempts to find the conditions (facts) that lead
to a given conclusion; it strives to match the assumed conclusion (the goal or subgoal) with
the conclusion (then part) of the rules, and if such a rule is found, its premise (if part)
becomes the new subgoal. The process is continued until the goal or subgoal is grounded
in initial facts. Backward chaining is best suited for applications in which the possible
conclusions are limited in number and well defined. Examples are classification or
diagnosis type systems.
Case-based reasoning
In case-based reasoning, the knowledge is represented using real cases, and when faced
with new cases or conditions, decisions are made based on the solutions to similar past
problems. Case-based reasoning is accomplished by gathering case histories and is
implemented by identifying significant features that describe a case. Case-based reasoning
systems offer the capability of incremental, sustained learning; each time a problem is
solved, a new experience is retained and can be applied for future problems. In general, the
case-based reasoning process entails:
1. Retrieve: Gathering from memory an experience closest to the current problem.
2. Reuse: Suggesting a solution based on the experience and adapting it to meet the
demands of the new situation.
3. Revise: Evaluating the use of the solution in the new context.
4. Retain: Storing this new problem-solving method in the memory system

4
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Fuzzy logic systems

Conventional expert systems are based on Boolean logic where only two possibilities are
considered; a rule can be either true or false under a given condition. In the real world,
problems are often not precisely defined, and the data can be uncertain or ambiguous. Thus,
it is hard to establish whether a certain statement (or rule) is True or False. Fuzzy logic,
originally invented by Lotfi Zadeh in 1965, combines numerical data and linguistic
knowledge to offer the flexibility of reasoning, especially when encountered with
uncertainty [76].
The principle is to work with degrees of possibilities instead of only two possibilities
considered in the conventional logic as a comparative approach to the real-life environment
[76]. For example, a voltage of 0.9 pu can be considered ‘Normal’ to a degree of 70%, and
at the same time, it can be considered ‘Low’ to a degree of 30%.
A fuzzy system consists of a rule base, an inference engine, a fuzzifier, and a defuzzifier,
as shown in Figure 3-2 (a). The process of assigning linguistic labels and membership
values for a crisp variable such as a voltage is called fuzzification. It is achieved using a
set of membership functions, each denoted by a linguistic label, as shown in Figure 3-2 (b).
Defuzzification is the opposite process. The rule base consists of if-then rules written in
terms of the linguistic labels of the inputs and outputs, for instance:
IF (V is Normal) THEN (Stability Margin is High)
IF (V is Low) THEN (Stability Margin is Very Low)
The inference engine uses a rule base to evaluate the membership values of the output
variable(s) for the given inputs. Then the process of defuzzification is performed if a crisp
output value is desired.

Rule base
Defuzzyfica�on
Fuzzyfica�on

Crisp Crisp
inputs output
Inference engine

(a)
Degree of Degree of
membership membership
Low Normal High Very Low Low Medium High Very High
1.0 1.0

Voltage Stability
(pu) margin
0.85

0.95

1.0
1.15
1.05

(b)

Figure 3-2: (a) Components of a fuzzy inference system (b) Examples of

membership functions
5
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

3.2 Learning by experience

The techniques and algorithms that accomplish learning by experience are broadly
considered as machine learning methods. In this type of learning, machine learning
techniques or algorithms process input data in order to perform classification, pattern
recognition, clustering, regression tasks, and more, based on the experience learned from
the data [76].
Artificial Neural Networks (ANN) are at the forefront of machine learning as enablers to
learning by experience. Warren McCulloch and Walter Pitts laid the foundations for the
artificial neural networks in 1943 by developing the first mathematical model inspired by
biological neurons and by Frank Rosenblatt in 1957 by creating the perceptron model as
an effort to imitate the human brain. A commonly used model of an artificial neuron is
shown in Figure 3-3 (a). An artificial neural network is a network of neurons consisting of
an input layer, which receives data from outside sources (data files, sensors, etc.), one or
more hidden layers that process the data, and an output layer that provides one or more
output data points based on the function of the network. An example of a feedforward
neural network is shown in Figure 3-3 (b). The weights of the neurons are adjusted
considering the relationship between the inputs and the outputs as defined by the
experience data [73].

synap�c bias input hidden output

inputs weights layer layer layer
𝑏
𝑀𝑀1 𝑤1 ac�va�on
𝑀𝑀2 𝑤2 func�on
⋮ output
𝑏+ 𝑤𝑖𝑖 𝑀𝑀𝑖𝑖 𝑦
𝑀𝑀𝑖𝑖 𝑤𝑖𝑖
⋮
𝑀𝑀𝑛 𝑤𝑛

(a) (b)

Figure 3-3: (a) Model of an artificial neuron (b) Example of a feedforward neural
network

There are several kinds of ANN architectures; among them, the main configurations are:

• Single Layer Feedforward Architecture

• Multiple Layer Feedforward Architecture
• Recurrent or Feedback Architecture
• Meshed Architecture
In addition to ANNs, other structures such as decision trees, support vector machine, etc.,
can facilitate learning by experience.
After defining the structure or architecture of an ANN (the number of hidden layers, the
number of neurons in each layer, etc.), it needs to be trained. Training is the process of
incorporating the experience (gathered from the training data) into trainable parameters of

6
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

the ANN. There are several approaches for training ANNs. They can be classified based
on the approach used. The main types of training approaches include:

• Supervised Learning – data used for training is tagged by a human or available

as pairs of inputs and the corresponding expected outputs. Typically used for
classification and regression applications.
• Unsupervised Learning – self-organization to capture patterns in input data as
neuronal predilections or probability densities. Three main domains of
applications are clustering, association rules learning and dimensionality
reduction.
• Reinforcement Learning – in model-free RL, the machine learns by interacting
with the environment as an agent and is given a numerical performance score
(reward) after performing the action as its guidance. Model-free RL is
popularly used in robotics. Model-based reinforcement learning algorithms can
model the environment to predict the outcomes of actions before performing
them and optimize iteratively from these predictions. Model-free RL can be
used in situations where the environment is better known, such as playing
chess.
The availability of the input data used to train the algorithm will be different for each
application. Sometimes, the entire dataset is available from the beginning; however, the
data that the algorithms need will grow over time. As a result, the learning by experience
approach taken considers the data availability and is broken down into offline and online
learning. When there is not enough data (experience) to perform an offline learning, the
focus normally is about recognition of types or classes (for example clustering) by learning
online with the input data.

• Offline Learning – batch learning techniques which generate the best ANN
parameters by learning on the entire training data set at once
• Online Learning – sequential learning in which data becomes available in a
sequential order and is used to update the ANN weights at each step. Online
learning is used where it is necessary for the algorithm to dynamically adapt to
new patterns in the data (distribution shift) or when it is infeasible to learn
from an entire dataset.
Normally, a neural network architecture will be chosen based on the application and data.
The types, size, and number of layers within the network, along with the optimization
strategy and loss equation being carefully chosen to allow the network to efficiently map
the input data to the application goal.
Figure 3-4 is an overview of types of machine learning algorithms, and Figure 3-5 is the
Scikit-learn algorithm cheat sheet [77].

7
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 3-4: Types of Machine learning algorithms

Figure 3-5: Scikit-learn Algorithm Cheat Sheet [112]

3.3 Learning by evolution

Trial and error approaches have been a key part to human learning. Several techniques have
been developed to imitate this human skill and focus on an objective while using the
experience to find a solution for a problem, by means of the evolution of a population.
The population in the context of evolutionary learning refers to a set of potential solutions
to a problem. For example, for the problem of finding setting of a non-directional
overcurrent relay (ANSI 51), a solution could be a pair of values representing the pickup
current and the time dial. The objective for application of device 51 to a transformer could
8
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

be to get a reference operating time for the internal and external faults. Given a range of
short circuit currents, the operating time function can be evaluated for potential solutions
for this individual in a population. A learning by evolution algorithm applies random
changes guided by natural evolutionary principles, swarm intelligence, colony behavior or
physical/chemical processes to find a group of solutions (in this case settings) to meet the
objective (the required operating times) or the best possible solution. Starting with an initial
population (often randomly generated), these kinds of algorithms can find better solutions
for complex problems by making variations in every population.
Evolutionary computing started in 1960s with the evolutionary strategies proposed by
Rechenberg, evolutionary programming proposed by Fogel, and the idea of the genetic
algorithms, which is a population-based algorithm, proposed by Holland [76]. In general,
the ideas of evolution in computing are based on the processes observed in biological
environments and focus mainly on cellular behaviors and multiple groups – colony
behaviors [74].
Metaheuristic algorithms are a set of techniques or algorithms applied to solve optimization
and search problems using population evolution. Some of the popular algorithms are
Simulated Annealing, Ants Colony, Bees Colony, Genetic Algorithms, Differential
Evolution, Particle Swarm Optimization, Harmony Search, Bat Algorithm, and Taboo
Search. Metaheuristics are widely used to solve complex optimization problems with a
single objective or multi-objective focus and the capability of problem generalization.

3.4 Deep Learning

Deep learning is a subfield of machine learning and its backbone is built on neural
networks. In fact, it is the number of node layers, or depth, of neural networks that
distinguishes a shallow neural network from a deep learning algorithm. A deep neural
network will have more than three layers. These algorithms use multiple layers to
progressively extract higher-level features from the raw input, with minimal human
intervention. Learning can be supervised, semi-supervised or unsupervised.
Traditional machine learning algorithms such as linear regression, logistic regression,
shallow neural networks, and decision trees are algorithms that find mappings between
structured representations of data (specific data with a predefined format) and an outcome
that is often seen as a prediction. These traditional machine learning algorithms often fail
to create good mappings between data and an outcome when the data is of an unstructured
or of complex nature (conglomeration of many varied types of data that are stored in their
native formats) such as images, waveforms, video, etc. To tackle this problem, data
scientists usually conduct data preprocessing and feature engineering processes to find data
representations that are suitable for traditional machine learning of algorithms [79]
In deep learning, suitable representations of data (features) are found in a more automated
way. Deep learning algorithms are capable of understanding concepts on their own, and
this is referred to as representation learning [80]. This capability makes the algorithm less
reliant on the prior domain knowledge of the person who performed the feature engineering
process and more reliant on the neural network architecture chosen (number of hidden
9
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

layers, and the number of neurons in each layer, activation functions, etc.). Deep learning
encompasses a set of algorithms that are multilayered and that can do automatic feature
engineering. More precisely, deep learning algorithms attempt to solve the representation
learning problem by creating hierarchical (layered) structures of concepts that can represent
complex concepts from simpler or lower-level ones [79] [81].
In deep learning, each level learns to transform its input data into a slightly more abstract
and composite representation. For example, in an image recognition application, the raw
input may be a matrix of pixels. The first representational layer may abstract the pixels and
encode edges. The second layer may compose and encode arrangements of edges. The third
layer may encode a nose and eyes. The fourth layer may recognize that the image contains
a face as illustrated in Figure 3-6. Importantly, a deep learning process can learn which
features to optimally place in which level on its own.

. . . .
. . . .
. . . .
. . . . .
.
edge contour object
inputs detec�on detec�on classifica�on
forma�on

Figure 3-6: Conceptual representation of a deep neural network for image

recognition

A deep learning model would require more data points to improve its accuracy, whereas a
traditional machine learning model relies on less data given the underlying data structure.
Deep learning models generally take longer time to train. Traditional machine learning
algorithms have limited hyperparameters for tuning, but deep learning algorithms can be
tuned in various ways.

3.5 Hybrid Intelligent Systems

A Hybrid Intelligent System (HIS) is a system that uses two or more of the techniques or
algorithms belonging to artificial intelligence. HIS are generally made to solve complex
problems by the combination of several techniques or algorithms [75].
HIS common combinations are Neuro-fuzzy Systems (FIS-NN), Evolutionary Fuzzy
Systems (EC-FIS), Evolutionary Neural Networks (NN-EC) and Hybrid Evolutionary
Algorithms. The General framework for hybrid architectures is shown in Figure 3-7.

10
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 3-7: General framework for hybrid architectures [78]

Typical architectures for hybrid intelligent system are shown in the Figure 3-8.

single
component fusion hierarchical hybrid

Figure 3-8: Typical architectures for hybrid intelligent system [82]

There are several ways to combine techniques or algorithms to create a hybrid intelligent
system. Techniques are chosen based on the problem to be solved. Pattern recognition
problems that include classifications usually use neural networks with fuzzy approaches to
create Hybrid Fuzzy or Neural Systems to improve the performance and to make a more
general capability according to variation of input data.

When finding an optimal solution of a problem, it is common to use a Hybrid Evolutionary

System by combining evolutionary techniques or algorithms with fuzzy logic or expert
systems. For example, expert rules can be defined with fuzzy logic in order to define the
starting population of an evolutionary technique or algorithm in order to improve the
convergence speed. In protection, it is common to use a combination of genetic algorithms

11
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

with fuzzy logic to optimize overcurrent protection and distance protection settings in
single objective and multi-objective problems of protection coordination.

When the complexity of the problems requires several modules of task, every one of them
with a complex level, the combination of techniques or algorithms becomes an intelligent
system by the integration of more than two techniques or algorithms to be able to be
successful. For example, to build a system capable to do a protection coordination study of
a grid element like a power transformer, several tasks need to be done and can include,
scenario analysis, topology recognition, expert rules application for settings, definition of
verification faults and coordination rules, evaluation of proper coordination margins and
an optimization method. All the problems therefore need the usage of several techniques
or algorithms capable of performing classification, rules application, optimization and can
be achieved by the combination for example of fuzzy logic, neural networks and
evolutionary techniques or algorithms using a hierarchical architecture.

4 AI/ML for Protection & Control

4.1 Drivers
The existing protection and control applications have been working for many years
assuming the conventional grid architecture, and they have performed exceptionally well
thus far. However, modern power systems dynamics have changed, leading to protection
and control facing new challenges to maintain/improve the network and asset performance
metrics. Many engineers and researchers are exploring new approaches such as biology-
inspired AI/ML-based solutions to the problems that cannot be properly resolved by
conventional physics-based approaches. Increasing availability of large amounts of high-
fidelity, rapidly sampled data, both in time domain and frequency domain is enabling the
development and deployment of AI/ML-based solutions to challenging protection and
control problems.
Some drivers leading the focus towards the Artificial Intelligence are:
● High penetration of Distributed Energy Resources
● Large scale inverter-based resources
● Growing need for advanced distribution protection and control applications
including fault detection and location
● Wide area instabilities
● Power system reconfiguration
● Increased use of HVDC converters, FACTS devices and compensators
● Natural disasters
● Need for improving the efficiency in protection systems design and engineering
● Mitigating human errors

12
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

4.2 Limitations of existing P&C technology

Faults can occur on any electrical equipment/component due to various causes such as
extreme weather, aging, impact due to repeated stress, adverse environmental conditions
resulting in degradation of the insulation, internal/external damages during construction,
etc. They result in outages of a specific asset or the entire network depending on the design
of power system, severity, coverage of protection zone (e.g., line, transformer, reactors,
buses, generators, loads).
Conventional power system protection schemes have been successfully applied to provide
adequate protection functions under most power system conditions including normal
operating conditions and various fault scenarios. However, there are conditions that
conventional protection schemes had shown limitations such that existing schemes either
could not provide adequate protections at all or the desired performance and/or
dependability-security balance of the protection schemes could not be achieved.
Conventional protection schemes work the best when the power system conditions that
require protection actions are very different from all other conditions for which when no
such actions are needed. Some protection schemes (e.g., out-of-step protection schemes,
system integrity protection schemes) work better when the power system where the
schemes are applied is not dynamic in terms of topology but may have difficulties to
balance the dependability and security requirements if the topology of the power system is
dynamic.
On the other hand, designing some protection schemes for dynamic or complex power
systems is a combination of science and art. This means that there may not be any
“standard” designs for such protection challenges. Solutions by different engineers could
be different from each other, which can result in differences among the power systems
where they are applied and/or the differences in personal experience and knowledge among
the designers. The situation described is common, for example, in new transmission grid
projects. New projects modify a part of the power system and they can create complex
topologies because of weak short-circuit conditions with a high source impedance ratio
(SIR), infeed, or special configurations like three-terminal lines. According to the
regulations of some countries, in many cases, the projects do not replace the existing
protection systems of existing elements, and the protection engineers would develop
settings for new and existing protection systems to maintain proper protection
coordination. Existing protection systems can present challenges if teleprotection is not
available for distance protections or if only overcurrent relays are available. In these cases,
regulatory fault clearing times are difficult to fulfill and even the performance of the backup
protection coordination may be insufficient if redundancy is not available.
In protection coordination studies the major challenge when protection settings are
calculated and verified is to achieve settings that are reliable for most operating scenarios
of the power system (optimal settings). Power system expansion, new dynamics of
equipment, and maintenance constantly add more complexity for finding the optimal
settings of protection systems. Figure 4-1 shows the conventional focus of the reliability in
protection systems.

13
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Reliability
Dependability Security
Equipment Focus Power System
Focus

Desired reliability
area

Optimal setting area for proper protection coordination

for maximum number of operating scenarios

Figure 4-1: Optimal setting area for reliability

Microprocessor-based relays are basically an industrial computer with settings as

determined by the protection algorithms. Pickup, dial, zone reach and operation time are
the first to come to mind, but there are many more that require attention as well. Examples
are polarization, phase selection, directionality, blocking logics, etc. Depending on the
protection element or function, it may require a revision to optimize settings for proper
protection coordination in all the possible operating scenarios.
In the presence of high inverter-based resources penetration, there are many challenges
when using classic analytical methods to solve complex contingencies to determine optimal
protection settings.

4.3 Opportunities for Artificial Intelligence in P&C

AI/ML can sometimes enhance existing power system protection and control functions that
microprocessor relays currently perform, such as the AI application in addressing
secondary arc example described in 5.1.1. It enables engineers to make more data informed
decisions by learning and creating models from historical and near real-time data. By
leveraging the data to make protection and control decisions, the engineers can efficiently
adapt their protection and control approach to account for a dynamic grid whose electric
properties and characteristics are changing.
Artificial intelligence could create efficiencies by adding value and insight into various
stages of the protection and control process. The main stages of the P&C process that AI
could impact are as shown in Figure 4-2:

14
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 4-2: Main stages of the P&C process that AI could impact

Each of these tasks above provide an opportunity for some aspect of AI/ML to assist either
the engineers or the P&C devices react faster, be more accurate, and become more efficient.
Applications and use cases that benefit from AI/ML are discussed in the subsequent
sections. These discussions emphasize how each application benefits from AI/ML and how
it impacts the various stages. In general, AI/ML may supplement and aid protection
engineers but will never replace them.

4.4 Considerations before using AI in protection

Although, as discussed in this section, there are opportunities for use of AI to protection,
before using any new method to improve existing technology and practices, a philosophical
as well as a technical evaluation of the pros and cons of the new method is in order.
Principles and applications of power system protection have been refined and perfected
over decades of practical experience and a number of revolutions in technology. Since
success of protection improves safety of equipment worth millions of dollars, and more
importantly, human lives, protection engineers have traditionally preferred transparent and
relatively simple physics-based models.
Interestingly, pattern creation and classification, two foundational components of ML
based methods that help build AI in a system, are already an integral part of protection.
Practically every relay uses a feature or a combination of features (pattern), and a rule-
based classifier. A feature is typically frequency-domain transformation of time-domain
measurement – current, voltage, frequency, power, impedance, harmonic content, high-
frequency content, etc. A pattern is created through the combinatory logic blocks provided
in numerical relays. A classifier typically has two-classes: Fault & No Fault. Separation
15
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

plane for the classifier is a threshold value (for the feature) or a combination of threshold
values (for the pattern). Values and combinations are determined through system analytics,
physics and physics-based models, and experience. Here, experience is simply a
representation of the human mind learning from past data. The difference is, the learning
in ML applications happens (or experience is gained) through algorithms developed and
refined by computer scientists. In applications using large sets of data, it has been shown
that these algorithms “digest” the data better than a human mind can. ML based methods
also tend to create and use quite complex patterns using mathematical transformations on
raw data. Such transformations, when used to create a data-driven model of a physical
process, often yield quantities that lose direct relationship to the physics underlying the
process it is trying to model.
These observations lead to two philosophical questions that need to be considered before
applying AI to improve or replace an existing physics-based method:
1. If AI uses the same features/patterns used by relays, why would it be able to create a
more dependable and secure classifier?
2. If AI uses more complex and abstract features/patterns that have no transparent relation
to the underlying physics, and thresholds (separation planes) are created simply by
learning through data, why would it work better?
Extending the thought behind these questions, if a legacy protection is not working in
certain system conditions, why would an AI based method work? If it uses the same
patterns the legacy relay is using, or can use, applying learning to these patterns may not
necessarily yield any better result, as the patterns have failed. A question about opaque
patterns to solve this problem is valid in this case as well.
This discussion is against replacing physics-based time-tested methods with AI-based
solutions without clear justifications, even if the existing methods fail occasionally.
Unfortunately, a large number of papers published on this topic takes this approach. These
papers also disregard the fact that power system protection is a system that encapsulates
interdependent localized protection-schemes. Replacing one/few such schemes by a ML-
based method may not reconcile with the holistic nature of power system protection [83].
A pragmatic approach could be to complement conventional physics-based approaches
with biology-inspired ML approaches to provide maximum protection coverage, and
optimize dependability and security based on asset/system requirement. This report will
present application examples of AI that have shown to or could potentially help improve
protections where underlying physical models are not precise or absent, leading to known
weaknesses.

4.5 Specific areas of AI/ML P&C application

For this report, power system protection is classified into two categories: asset protection
and network protection.
The asset protection function involves detection of faults or abnormal conditions,
discrimination of faulted zones or devices, identification of fault types, and making trip or
block decisions, sometimes adaptively.
16
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

The network protection functions such as System Integrity Protection Scheme (SIPS),
Remedial Action Schemes (RAS) and Wide Area Monitoring, Protection and Control
Systems (WAMPAC) requires recognizing rare critical conditions impacting the power
systems by monitoring the diverse set of variables. These systems, upon detection of
critical conditions, initiate emergency control actions.
Regardless of the focus, design of the protection function/algorithm, system analysis,
selection of settings, and coordination often require trade-offs between the security (no
false tripping), speed of operation, and the dependability (no missing operations). The more
secure the relay is (both the algorithm and its settings), the more it tends to be less sensitive
or operate slowly. And vice versa, the faster the relay is, the more it tends to operate falsely.
The design and coordination of asset and network protection functions to achieve desired
reliability and cost objectives may become a complex optimization problem in modern
power systems. As the grid becomes more dynamic, the optimization objective becomes a
moving target, and static settings and assumptions may not be sufficient.

4.5.1 Asset Protection

The limitations of classical asset protection techniques may arise from the extent of
information available through the approach of using the same measurements used by
existing relays (e.g., sampling frequency) and/or limitations of the relaying algorithms used
for detection and discrimination. These limitations can manifest in the form of less-than-
ideal speed and sensitivity to high impedance faults, or adaptability to various conditions
such as high source impedance, series compensated lines, inverter-based sources, etc. To
mitigate these limitations of the classical relaying principles, it is possible to improve and
extend the range of measurements available for decision-making and improve the
recognition process itself. For example, the high frequency transient signals contained in
the input signals can be utilized to detect and discriminate high impedance, arcing and
evolving faults, which are challenging to traditional protection function based on power
frequency phasors. However, circuit laws applied to simplified models cannot characterize
the relationship between the high frequency content and the fault location/severity. In such
situations, AI/ML techniques can be used to train classifiers to make decisions based on
various signatures contained in the high frequency components of the input measurements.
On a case-by-case basis, engineers can consider leveraging AI/ML models to augment the
selectivity, security, and speed of classical relaying principles by
- Optimizing settings and utilizing hybrid approaches that use multiple relaying
principles
- Automatically correcting CT and VT signals
- Employing non-conventional algorithms

4.5.2 Network Protection

With respect to network protection, significant advances have been made in the field of on-
line security assessment technologies that can determine critical contingencies, transfer
limits, and remedial measures. However, conventional numerical techniques based on
17
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

offline studies are usually time consuming and therefore are not always suitable for real-
time applications. Also, these methods suffer from the problem of misclassifications or/and
false alarms. Many studies show that AI/ML techniques can be applied to develop effective
RAS and WAMPAC systems to detect critical or abnormal network conditions in real-time
and propose or automatically activate remedial actions.

4.5.3 Post Event Analysis

Another context where AI/ML can be very helpful is post event analysis by using data
analytics. Modern data mining techniques can be essential when it is required to analyze
large data sets such as digital fault recorder data, synchrophasor data, utility meter readings,
weather data, etc. and their combinations. Analysis of large structured or non-structured
data sets is needed in many situations such as sequence of event analysis.

4.5.4 Summary of Potential P&C Areas for Application of AI/ML

Table 4-1 provides some examples of protection and related functions where application
of AI/ML techniques could be helpful. The current intelligent electronics devices (IEDs)
employ classical protection functions to achieve many of these functions with high degree
of success, but there is some room for improvement in terms of speed, sensitivity,
reliability, and adaptability as mentioned earlier. On the other hand, classical algorithms
are inadequate or too complex to apply in real-time for some functions such as system
stability prediction and control, islanding detection, controlled islanding, arc detection, etc.
In these applications AI/ML can play a crucial role.

Table 4-1: Potential areas for application of AI/ML

Asset Protection Network Protection

• High impedance fault detection • System level monitoring and
• Evolving fault detection protection
• Fault direction discrimination • Voltage stability monitoring and
• Faulted segment identification protection
• Detection of secondary arc extinction • Transient stability monitoring and
• Adaptive protection protection
• Condition monitoring and fault diagnosis of • Islanding detection and protection
transformers, reactors and capacitors • Frequency stability protection
• Detection of CT saturation, VT transients,
and inrush conditions, as well as self-
monitoring and diagnosis.
Common to both asset and network protection
• Alarm processing and diagnosis
• Settings/coordination optimization, adaptive settings, and setting-less protection
• Digital fault recorder data analysis
• Sequence of event data analysis
• Bad data detection and data conditioning

18
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

5 APPLICATION ILLUSTRATIONS
The following example applications illustrate how AI/ML technology could be applied to
resolve practical power system protection and control problems. The example applications
are divided into two broad categories: in-use examples and high-potential examples.
When an AI/ML application has been implemented in a product or has been used to solve
real-life power system protection and control problems, it is considered as an in-use
example.
The inclusion of high-potential application examples could be somewhat subjective to the
opinion of this working group: either there has been sufficient evidence, i.e., the current
R&D work has shown some very good results, or the assessment of this working group has
led us to believe that the application of AI/ML has a high-potential to provide a better
solution over existing technologies.
The report only included the example applications that are known to or evaluated by this
working group. There could be many other example applications that are currently in-use
or considered as high-potential ones under the active R&D efforts.

5.1 Existing Applications

Two existing example applications of AI/ML technology for power system protection and
control are described below. The first example application, where AI/ML technology is
applied to detect secondary arc, has been implemented in a real product. The second
example application, where AI/ML technology is applied to select the optimized relay
settings, has been in use by a utility.

5.1.1 AI to Detect Secondary Arc

5.1.1.1 The problem: secondary arc detection
The problem of secondary arc detection is related to single phase tripping and
autoreclosing. Single phase tripping and autoreclosing is used to clear temporary single
phase to ground faults. In case of a single phase to ground fault, the faulted phase is tripped
from both ends of the line and reclosed after a short time, e.g., typically less than 1 sec (30-
50 cycles). In most cases the fault has disappeared during this time and the reclose is
successful.
Single phase tripping refers to opening only the faulted phase at both ends of the line.
Therefore, no current can flow via the faulted phase directly into the fault anymore. The
primary arc extinguishes.
Unfortunately in some cases a secondary arc appears, which is fed by the two healthy
phases via capacitive coupling like shown in Figure 5-1. The voltage UM, measured at the
disconnected phase is characterized by the ohmic nonlinear behavior of the secondary
arc RARC.

19
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Phase A

CCA
Phase B

CBC
Phase C

UM RARC

Figure 5-1: Simplified diagram showing the secondary arc after single phase
tripping

If the secondary arc is extinguished, the equivalent circuit is changing to a different model
like shown in Figure 5-2. The voltage UM, measured at the disconnected phase after
extinguishing of the secondary arc is characterized by the linear capacitive behavior of
the phase to ground capacitance CCG of the open conductor.

Phase A

CCA
Phase B

CBC
Phase C

UM CCG

Figure 5-2: Simplified diagram showing the secondary arc extinction after single
phase tripping

A reclose at presence of the secondary arc is not successful as shown in Figure 5-3. After
reclosure, the fault persists which leads to a final trip of the protection. This unsuccessful
reclosure could be prevented using a secondary arc detection function.

20
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

I/kA
Fault
5,0 Unsuccessful
reclosure
2,5

0,0
0,00 0,25 0,50 0,75 1,00 1,25 1,50 1,75
t/s
-2,5

-5,0

U/kV

200
Secondary arc
100

0
0,00 0,25 0,50 0,75 1,00 1,25 1,50 1,75
-100 t/s
-200

-300

TRIP A
TRIP B
TRIP C
Close
0,00 0,25 0,50 0,75 1,00 1,25 1,50 1,75
t/s

Figure 5-3: Unsuccessful reclosing if secondary arc is not extinguished after single
phase tripping

A successful reclosure after extinguishing of the secondary arc is shown in Figure 5-4.
After tripping the line, the fault current disappears. At the same time the voltage starts the
typical nonlinear behavior of arcing. Later the secondary arc extinguishes, and the voltage
is changing to a linear capacitive behavior. After successful reclosing both voltage and
current go back to normal conditions. In this case the single-phase dead time could be
reduced from 1.2 s to 0.5 s, if the extinction of the secondary arc would be detected.

21
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

I/kA
5,0
Fault Successful reclosure
2,5
0,0
0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 1,1 1,2
-2,5 t/s
-5,0
-7,5

U/kV
200
Secondary arc Secondary arc extinguished
100
0
-100 0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 1,1 1,2 t/s
-200
-300

Figure 5-4: Successful reclosing if secondary arc is extinguished after single phase
tripping

5.1.1.2 Implemented solution

As described in [84], there are several algorithms to detect the secondary arc for instance
based on harmonics or phase angle of the open-phase voltage. All these algorithms try to
distinguish between the ohmic, non-linear behavior of the secondary arc and the capacitive
behavior of the voltage if the secondary arc is extinguished.
The behavior of arcing however can be quite different depending on several factors. That
is why another algorithm was added based on a neural network to detect the presence of
the secondary arc.
Finally, the secondary arc detection function uses several criteria to detect the secondary
arc as shown in Figure 5-5.

Algorithm based
Result Weighting factor 1
on harmonics

Algorithm based
Result Weighting factor 2
on voltage angles Logic End result
. .
. .
. .

Neural network Result Weighting factor n

Figure 5-5: Structure of the secondary arc detection function

22
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

As shown in Figure 5-5 the neural network was added as one criterion to detect a secondary
arc.
5.1.1.3 Training of the neural network
A neural network was added as a criterion to detect a secondary arc based on the waveform
of the open-phase voltage. A neural network results in an optimal algorithm for classifying
between “secondary arc” and “no secondary arc” detection if it was trained correctly with
a representative set of data.
The neural network shown in Figure 5-6 as example contains 9 neurons in three layers. In
a learning process these neurons were optimizing their weight factors. To achieve a good
result a lot of different training data is necessary. The training data consists of fault records
with an analog channel representing the open-phase voltage and a binary signal
representing the information about the presence of the secondary arc.
Most of the training data was computed by network simulation programs, simulating
different conditions of secondary arc. These training data was augmented by a large amount
of fault records from “real” secondary arcs, received from different customers around the
world.

Figure 5-6: Learning process of the neural network used for secondary arc detection

5.1.1.4 Results
The following three figures are illustrative of the successful function of secondary arc
detection of the implemented neural network. The figures are based on fault records
showing the voltage of the open phase in blue and the output signal of the neural network
in red color.

23
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 5-7 shows a secondary arc which is extinguished approximately 300 ms after the
single-phase tripping. The neural network detects the state change very fast and precise.

Figure 5-7: Detection of secondary arc extinguishing after single phase tripping

Figure 5-8 is showing a secondary arc which is changing its shape but does not extinguish
during the single-phase dead time. The output of the neural network is stable and indicates
the presence of the secondary arc during the whole time.

Figure 5-8: Secondary arc not extinguishing during the dead time of autoreclosure

Figure 5-9 shows a secondary arc which is re-arcing after extinguishing the first time. The
neural network detects the first extinguishing of the secondary arc and the re-arcing very
fast and precise. After re-arcing, the output of the neural network is correct but not as stable
as previous example.

Figure 5-9: Detection of re-arcing after extinguishing of secondary arc

24
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Due to the concept of using several criteria to detect the presence of the secondary arc the
risk of a malfunction in case of a secondary arc which is not detected with the neural
network is minimized.

5.1.2 Hybrid Intelligent Model for Protection Settings Optimization

5.1.2.1 The Problem: Finding Optimal Settings for Distance Protection
Distance protection settings are difficult to optimize because several electrical conditions
impact its performance, sometimes in competing manner. SIR, load flow, topology, type
of fault, fault resistance, zero sequence mutual coupling of paralleled lines, infeed of 3-
terminal lines can modify the impedance seen by protection relays and cause an overreach
or underreach problem. Figure 5-10 illustrates the challenges in finding optimal distance
relay settings.

Figure 5-10: Problem of Finding Optimal Distance Relay Settings

From the point of view of a power system protection engineer, the problem of finding
optimal settings in protection studies for distance protection means to find the reach values
of every zone, that is inductive reach and resistive reach, for reliable operation of the relay
in all the possible scenarios considering at least one level of contingency (N-1 criteria). It
is a highly time-consuming task to get a set of optimal settings.
5.1.2.2 Hybrid Intelligent System for Distance Protection Settings Optimization
A Hybrid Intelligent System (HIS) for distance protection settings optimization was
proposed, implemented and tested and the work and the results are reported in [85]. The
techniques used in the HIS model were Constructive Methods (sequential building of
solutions), Fuzzy + Conventional Logic, and Differential Evolution (DE). Figure 5-11
shows the structure of the HIS that contain a number of functional modules.

25
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Network DIgSILENT
Protected Line, Power Factory
reference bar

Relay Topology identifier (DPL) Rules

(Constructive Method) verification
1 faults
selection

Setting
Rules Verification Faults
Initial Solutions Generator
( Constructive Method) Programmer Algorithm
(Fuzzy Logic) 2
3

(DPL) Settings Loader

4.1
The Judge Verification Simulations
(Constructive Executor (DPL)
Method) ( Complete Method-
DIgSILENT Power Factory)
4.2

Performance Evaluation 4.3

stop criteria ok?

DIFFERENTIAL
EVOLUTION
4
New Generation
Generator

Report Generator

Figure 5-11: HIS for Distance Protection Settings Optimization

The HIS was designed to automate the whole process of calculation, verification and
optimization of distance protection relay settings by the execution of faults considering
normal operation scenarios and N-1 contingencies.
First part of the HIS considers automatic topology identification based on object-oriented
programming, which is labeled as Topology Identifier (1) in Figure 5-11. The topology
identification is used to program the faults to be applied to evaluate performance of the
settings. Programming of faults uses a hybrid logic that uses both fuzzy and conventional
logic. Then the module labeled Verification Faults Programmer Algorithm (2) in Figure
5-11 generates a set of faults according to the topology around the protected line, for
evaluating the performance of the settings.
26
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

After Verification Faults (VF) set is build, typical distance protection settings are
calculated by the Initial Solutions Generator (3) in Figure 5-11 as the initial population for
the optimizer algorithm that is based on differential evolution. Solution structure is shown
in Figure 5-12.

Figure 5-12: Solution structure for zones 1, 2, 3 forward and 4 reverse

The relay coordination problem is typically formulated as mixed integer non-linear

programming (MINLP). Most research efforts had been focused on overcurrent relays.
Considering ranges involved in variables in distance relay setting problem, differential
evolution was chosen over neural networks and vector support machines because of
generalization required to permit several network types (several sizes related to nodes and
topologies). The differential evolution was also chosen over genetic algorithms and swarm
techniques because this work was focused on only one relay and has several continuous
variables with known ranges. Coordination with other relays is done via constraints.
To find the global optimal, a modified version of the DE algorithm was implemented in
HIS, as shown in Figure 5-13.

Figure 5-13: Differential evolution modified algorithm of HIS

27
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

An innovative concept to quantify performance of a distance protection relay setting was

developed and implemented in the module labeled as The Judge (4.2) in Figure 5-11. The
basic formulation of Judge function is shown in Figure 5-11 and the performance
evaluation function (FDR) is explained in detail below.
The basic idea is to give points according to the performance of every fault (FV is the set
of faults to be evaluated) considering the ideal expected result and the simulated operation
based on the reliability concept. The FDR function has two components, DA and P. DA
component gives points to every fault according to expected result (REOF) and the
operation time of the relay for the fault being evaluated (TOF). Parameters k1 to k5 are
selected according to the performance priority (dependability of security). P component
gives to the fault a general weight factors according to how relevant and frequent is the
fault. The factors consider fault location FPUF, fault type FPTF and fault resistance FPRF.

The function for the evaluation of the performance of the relay is:

𝐹𝐹𝐹𝐹𝐹𝐹 = � 𝐷𝐷𝐷𝐷𝑖𝑖 (REOF𝑖𝑖 , TOF𝑖𝑖 ) × 𝑃𝑃𝑖𝑖 (FPUF𝑖𝑖 , FPTF𝑖𝑖 , FPRF𝑖𝑖 )

𝑖𝑖 ∈ 𝐹𝐹𝐹𝐹
where:
𝐹𝐹𝐹𝐹𝐹𝐹: performance function
𝐹𝐹𝐹𝐹: the set of faults to be evaluated
𝐷𝐷𝐷𝐷𝑖𝑖 : Component for evaluation of relay at the fault 𝑖𝑖
𝑃𝑃𝑖𝑖 : Component according to fault characteristics

Component 𝐷𝐷𝐷𝐷𝑖𝑖 is calculated as follow:

𝑖𝑖𝑖𝑖 𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖 ∈ 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 → 𝑘𝑘1 (according to expected)

𝐷𝐷𝐷𝐷𝑖𝑖 = � 𝑖𝑖𝑖𝑖 𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖 > 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 → 𝑘𝑘2 (delayed trip) � ×
𝑖𝑖𝑖𝑖 𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖 < 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 → 𝑘𝑘3 (too fast trip)
𝑖𝑖𝑖𝑖 𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖 < 𝑁𝑁𝑁𝑁 ∧ 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 = 𝑁𝑁𝑁𝑁 → 𝑘𝑘4 (𝑢𝑢ndesired trip)
� 𝑖𝑖𝑖𝑖 𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖 ∈ 𝑁𝑁𝑁𝑁 ∧ 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 < 𝑁𝑁𝑁𝑁 → 𝑘𝑘5 (trip omission) �
in other wise → 1

Where:
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑖𝑖 : Expected Range of Operation of the relay for the Fault 𝑖𝑖
𝑇𝑇𝑇𝑇𝑇𝑇𝑖𝑖 : Operation Time of the relay at Fault 𝑖𝑖
𝑘𝑘1 , 𝑘𝑘2 , 𝑘𝑘3 , 𝑘𝑘4 , 𝑘𝑘5 : factors selected according to the focus (dependability or security). Further details
in [85].

28
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Component 𝑃𝑃𝑖𝑖 is calculated as follow:

𝑃𝑃𝑖𝑖 = 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 × (𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 + 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 )

Internal fault → 1
𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 : Fault location weight factor for fault 𝑖𝑖: = � �
External fault → 0.7

single fault to ground → 0.9

double − phase → 0.25
𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 : Fault location weight factor for fault 𝑖𝑖: = � �
double − phase to ground → 0.4
three phase → 0.3

0 − 10 Ω → 1
⎡ > 10 − 20 Ω → 0.9 ⎤
⎢ ⎥
𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝑖𝑖 : Fault resistance weight factor for fault 𝑖𝑖: = ⎢ > 20 − 40 Ω → 0.8 ⎥
⎢ > 40 − 70 Ω → 0.5 ⎥
⎢> 70 − 100 Ω → 0.3⎥
⎣ > 100 Ω → 0.2 ⎦
The optimization model used is shown in Figure 5-14. The key concept is to find the best
setting according to the priority of dependability and security. For prioritizing the
dependability, performance is calculated using parameters focused on dependability to
promote tripping over undesired tripping. On the other hand, to prioritize security,
performance is calculated using parameters that prioritize points to avoid undesired
operations.
The optimization model is:

𝑀𝑀𝑀𝑀𝑀𝑀 𝐹𝐹𝐹𝐹𝐹𝐹1 × �1 − 𝑑𝑑𝑑𝑑𝑠𝑠𝑠𝑠𝑠𝑠 � + 𝐹𝐹𝐹𝐹𝐹𝐹2 × 𝑑𝑑𝑑𝑑𝑠𝑠𝑠𝑠𝑠𝑠

Where:
𝐹𝐹𝐹𝐹𝐹𝐹1 : performance of the relay for a set of faults 𝐹𝐹𝐹𝐹, for parameters 𝑘𝑘1 , 𝑘𝑘2 , 𝑘𝑘3 , 𝑘𝑘4 , 𝑘𝑘5 with
dependability focus.
𝐹𝐹𝐹𝐹𝐹𝐹2 : performance of the relay for a set of faults 𝐹𝐹𝐹𝐹, for parameters 𝑘𝑘1 , 𝑘𝑘2 , 𝑘𝑘3 , 𝑘𝑘4 , 𝑘𝑘5 with security
focus.
𝑑𝑑𝑑𝑑𝑠𝑠𝑠𝑠𝑠𝑠 : User input factor to define security vs dependability priority.

29
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 5-14: Optimization Model used in the HIS

5.1.2.3 Test Results

Figure 5-15 shows the test system used for initial testing of the HIS to prove the concept.
Conventional settings are optimized according to the electrical conditions leading to a
reduction in zone 1 reach setting to avoid undesired tripping.

Figure 5-15: HIS test system results

5.1.2.4 Real Life Application Example

During a blackout in a section of the Colombian Power System, caused by a double line-
to-ground bus bar fault at the substation JM in the level of 115 kV, the 115kV line relay at
PS looking to JM did not trip. Moreover, readings showed apparent impedances that
30
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

differed from the line length parameter of the protection study. The utility owner of the PS
– JM 115kV line corrected the line length to 6 km and thus a new protection coordination
study was needed. The new length presented a SIR of 25 at the 115 kV JM terminal at PS.
Short circuit level at substation PS is 6.4 kA and at substation JM is 7.3 kA. The high SIR
condition and very short length of the line presented a settings challenge. The protection
scheme available for the line bay PS to JM 115 kV was ANSI/IEEE 21, ANSI/IEEE
67/67N without teleprotection. Figure 5-16 shows the topology of the grid.

Figure 5-16: Real life application example grid

The HIS was used as a support for the decision about the best settings of the line bay PS to
JM 115 kV. The best solution after 22 generations is shown in Table 5-1 together with one
of the initial solutions based on conventional criteria. The percentages in settings of Table
5-1 are with respect to the impedance values of the line for inductive reach. For resistivity
reach, the percentage is with respect to the minimum load impedance, considering 450 A
of current capacity.

31
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Table 5-1: Comparison Optimized vs Conventional Setting line bay PS to JM 115kV

Optimization Conventional
Setting [%] [%]
Result [Ω] Criteria [Ω]
Zone 1 X 1.85 71.1 2.23 85.7
Zone 1 R 47.56 32.2 63.7 43.2
Zone 2 X 6.19 237.8 3.41 131.0
Zone 2 R 84.93 57.6 63.7 43.2
Zone 3 X 8.94 343.4 16.43 631.1
Zone 3 R 20 13.6 63.7 43.2
Performance [%] 50.2 27.8

After the analysis of the results of the optimal solution, it was identified an incursion in
zone 1 for 4 to 6 ohms single phase to ground fault at the remote busbar (Sub JM) of the
bay under study. Figure 5-17 shows the incursion of the single line to ground fault of 4
ohms.

Figure 5-17: Zone 1 incursion problem for of real-life application example grid

The incursion shows the balance of the HIS that tried to achieve dependability and security.
The optimal settings cover faults between 5 ohms or below to be detected along the
protected element. The conventional solution in this case has several zone 1 incursions for
faults in the adjacent elements and it does not cover all faults in the protected element with
a fault resistance between 0 to 5 ohms, leading to a low performance of the settings as the
Table 5-1 shows. The final decision considered the settings proposed by the HIS including
a timing of zone 1 of 100 ms. Zone 3 had an important reduction due to the incursion of
faults in the busbar of 230 kV substation JM. Figure 5-18 shows the best solution for every
generation.

32
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 5-18: HIS Execution final solution

5.2 Emerging Applications

In the following, example applications, where AI/ML technology may have a high potential
to these application areas, are described.

5.2.1 High-Impedance Fault Detection with Artificial Neural Networks

High-impedance fault (HIF) poses unique challenges to conventional protection schemes
operating on the premise of significant fault current availability, as it typically creates little
to no fault current. Depending on the HIF type, e.g., downed conductors, persistent arcing,
and intermittent arcing, the fault signature of each HIF type can be very different from each
other. Therefore, it is difficult to detect HIF using conventional protection methods.
Although HIF rarely causes secondary equipment damage, it is critical to detect and clear
the HIF promptly, to reduce safety hazards to the public and surrounding environment,
e.g., wildfire, electrical shock, etc.
AI approaches, such as ANN, have been applied to address this technical challenge as early
as the 1990s. Figure 5-19 presents a typical ANN structure. What is important when using
ANN is:
• Selecting what the inputs to the ANN will be
• Selecting the ANN architecture
• Using the right data for the training
• Being as close as possible to reality

33
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 5-19: Typical ANN structure.

As an example, this section describes an R&D project using ANN based HIF detection
method that was carried out at New York State Electric & Gas, Binghamton, NY in 1991.
It successfully demonstrated the usability of ANN for high impedance fault detection. As
a result of the project, two patents [86] [87] were issued to the team.
The method first performs pre-processing of the analog signals representing phase currents.
They are provided to the inputs of an analog-to-digital (A/D) converter operating with a
sampling rate of 80 samples/cycle. Digitized representations of the phase currents are
applied to a pre-processor via a data bus. The pre-processor performs several mathematical
and data formatting operations and then the pre-processed data is applied to the input of a
trained ANN.
Training the ANN is necessary to establish the weights to be used to classify events
occurring on the network as HIFs or normal switching events. Training was accomplished
by providing multiple sets of known data representing both HIF and normal network
conditions to the ANN along with the correct labels corresponding to each data set. Data
sets may be compiled using computer simulation techniques or may consist of actual field-
collected data corresponding to both fault and no-fault conditions. Actual field-collected
data and superposition combinations of field data were used to train the ANN. Over 300
sets of field-collected and superposition data were used for the training. A typical training
data set consisted of data points representing 10 seconds of sampling time on a real or
simulated network.
The backpropagation of errors method was the learning technique chosen to train the ANN.
A learning rate of 0.05 and a momentum (a method of changing weights based on a
previous weight) of 0.1 were chosen. The data processing also included the identification
of zero-crossing points in the data for each phase conductor. The maximum and minimum
currents for each cycle are determined.
It has been found through experimentation with both simulated and field recorded data that
calculating the first derivative is essential to the process of accurately detecting HIFs. For
purposes of illustrating the importance of the first derivative in detecting a high-impedance
fault, refer to Figure 5-20 (a) and Figure 5-20 (b). The bar graphs of ANN output (relative

34
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

value or relative HIF probability) vs. time are shown for a 10-second period for an actual
HIF condition occurring on a single phase of a 13 kV three-phase distribution line. For this
fault, the phase B leg of the distribution line was dropped onto ice. Series 1 data shows the
faulted condition while Series 2 data shows the line under no fault conditions. In Figure
5-20 (a), the first derivative has not been calculated and as may be seen, there is little
difference in the relative value of Series 1 or Series 2 data. In Figure 5-20, identical data is
plotted. However, the first derivative has been calculated, and consequently there is
pronounced difference in the output of the ANN in response to the HIF (relative value of
the Series 1 data).

(a) Without rate of change calculation (b) With rate of change calculation

Figure 5-20: ANN output – relative HIF probability.

In conclusion, ANNs have been around for more than half a century. With today’s huge
amount of data in digital substations and the integration of distributed inverter-based
energy resources they will play a very important role in the future protection and control
systems. This HIF detection ANN application project 30 years ago demonstrated how
useful such methods can be for detecting conditions that are difficult to protect using
conventional methods. The above is only one of the examples – Another application of
ANN for high impedance fault detection can be found in [113].

5.2.2 Use of synchrophasor and ML for Early Detection of Long-term Voltage

Instability
5.2.2.1 Introduction
Voltage stability refers to the ability of a power system to maintain steady voltages close
to nominal value at all buses in the system after being subjected to a disturbance, and the
phenomenon is divided into two categories: (i) Short-term voltage stability which involves
dynamics of fast acting load components such as induction motors, electronically
controlled loads, HVDC links and inverter-based generators, and (ii) Long-term voltage
instability which involves slower acting equipment such as tap-changing transformers,
thermostatically controlled loads, and generator current limiters [88]. This application
targets the long-term voltage instability, which typically manifests as a progressive voltage
drop in the system ultimately leading to the collapse of system. The long-term voltage
instability has been a contributing factor of many blackouts occurred around the world and
35
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

early detection enables taking manual or automatic remedial action to prevent voltage
collapse.
The phenomenon of long-term voltage instability is typically described using the P-V curve
(or Q-V curve) at a given load bus as shown in Figure 5-21 (a). The loadability margin
(LM) illustrated on the P-V curve shown in Figure 5-21 (a) is an easily understandable
indicator of proximity to voltage instability. Pre-determined LM thresholds can be used to
alert the system operators of impending voltage instabilities. Tracing of the P-V curve
starting from a given operating point needs a technique such as continuation power flow
(CPF) and requires significant computational effort to apply to a large power system.
Point of Operation
Voltage (pu)

Point of Voltage Collapse

Loadability
Margin
P P
0 max

Active Power (pu)

(a) (b)

Figure 5-21: (a) P-V curve and Loadability Margin (b) Thévenin equivalent circuit
of the network at load bus-i

5.2.2.2 Problem Statement

Voltage stability monitoring and prediction
The benefits of real-time voltage stability monitoring are well recognized. Although LM is
a good indicator of proximity to voltage collapse, its computation using a technique such
as CPF is time consuming for large systems due to iterative computations involved.
Furthermore, accurate knowledge of the network topology, which can be undergoing
changes during a disturbance, is required for CPF. Therefore, evaluation of LM using direct
methods is not practical for real-time voltage stability monitoring.
Use of voltage stability indices (VSI) is an alternative method of assessing the long-term
voltage stability. In general, VSIs are derived with respect to a load bus while considering
the rest of the system as an equivalent circuit as shown in Figure 5-21 (b). Most VSIs are
founded on the maximum power transfer theorem but differ in focus (load bus or transfer
corridor), extent of information used (local or wide-area measurements, requirement of
topology and system admittance data), simplifying assumptions, etc. There are some VSIs
based on other considerations such as reactive power reserves, power losses in a region of
interest, etc. A number of these VSIs can be computed using real-time measurements with
a computational effort feasible for real-time implementation. However, when the topology
and the operating conditions change (especially during an evolving abnormal situation),
the accuracy of different VSIs is affected depending on the underlying principle. Therefore,
relying on a single VSI to detect impending voltage instabilities has drawbacks.
36
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

There is a potential for improving the accuracy of voltage stability assessment by

combining several VSIs, particularly those based on different principles. Furthermore, an
easily understandable indicator such as LM would be more useful for system operators in
taking remedial actions. Therefore, the approach proposed here explores the possibility of
using multiple VSIs to predict the loadability margin. However, the relationship between
the LM and VSIs is not straightforward and therefore a machine learning type approach is
required to learn the implicit relationship between the LM and the VSIs.
5.2.2.3 Real-time Voltage Stability Monitoring System
This scheme consists of three subsystems: phasor measurement unit (PMU) infrastructure,
real-time monitoring system at the control center and the offline mashing learning model
(MLM) training system as shown in Figure 5-22. The PMU infrastructure provides real-
time synchrophasor measurements from selected locations of the power system. The
voltage stability monitoring (VSM) system collects PMU data and predicts the LM of the
power system using a set of MLMs trained offline. PMU measurements are validated and
conditioned to remove bad data, and then used to compute several VSIs. MLMs use
computed VSIs and phasor data for predicting the LM. Predictions of the ensemble of
MLMs are aggregated to obtain the final prediction, which will be displayed to the operator
through a voltage monitoring dashboard.

Figure 5-22: Proposed real-time VSM system

5.2.2.4 Feature Selection and Offline Training Process

The first step of the offline MLM training process is to generate a large sample of potential
operating conditions of the power system, including under different contingencies. These
operating conditions are then fed to a CPF program to calculate the bus voltages and LMs
as the system load increases. Then these data are used to calculate the set of candidates
VSIs listed in Table 5-2, which are selected based on the suitability for real-time
implementation. A large database is created by computing VSIs and the corresponding
LMs at different operating points obtained by sampling the CPF results.

37
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Table 5-2: Candidate VSIs for MLMs (see [89] for VSI definitions/references)
Name of VSI Abbreviation
Voltage Stability Load Index VSLI
Voltage Collapse Proximity Index VCPI
Voltage Stability Load Bus Index VSLBI
S Difference Criterion SDC
Simplified Voltage Stability Index SVSI
Impedance matching Stability Index ISI
Voltage Stability Margin VSM
Reactive Power Reserve Index ERPR
L-index L
Wide-Area Loss Index WALI

The second step of the development is the feature selection for MLMs. The correlation
between each VSI and LM is analyzed using the Spearman’s rank correlation coefficients
to select the most suitable VSIs to be used as input features for MLMs. This process can
also eliminate the VSIs that provide redundant information. Examples of this analysis
performed for IEEE 14 and 118 bus test system are presented in Figure 5-23.

(a) (b)

Figure 5-23: Correlation heat map of candidate VSIs and LM for (a) IEEE 14 bus
system and (b) IEEE 118 bus system

The third step is the offline training and validation of machine learning based LM
prediction system. The most important features (VSIs) selected in step two are used to
generate the input-output data pairs to be used in training and testing of MLMs. Each data
set includes a set of input VSIs computed at a particular operating point and the LM at that
operating point. These data sets are structured in the form of [VSI1, VSI2, …..VSIn: LM].
The data is divided into training and testing data (roughly in 3:1 proportion). Several
MLMs are designed considering other criteria such as those VSIs computed using only the
38
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

measurements at a local bus. Also, one MLM is designed to directly use the voltage
phasors as inputs and the data for it are structured in the form of [|V1|, δ1, …..|Vn|, δn : LM].
The locations of voltage measurements are selected based on a recursive feature
elimination method. Descriptions of MLMs are given in Table 5-3.

Table 5-3: Descriptions of different MLMs in the ensemble

MLM Inputs
MLM 1 Voltage magnitudes/phase angles of selected buses
MLM 2 All of the candidate VSIs
MLM 3 VSIs which are calculated using only the local measurements of the considered bus
MLM 4 VSIs with high correlation (correlation > 0.9) with LM

Random Forest (RF) regression machine learning algorithms is selected as the most
feasible algorithm among the other candidates through preliminary studies. After training
each MLM, the accuracy of the MLMs were tested and the less accurate MLMs were
discarded, for example MLM 2 in Table 5-3 is eliminated. Finally, a set of weights is
determined to aggregate the outputs from MLMs in the ensemble using weighted average
method. Typical prediction results for testing data, using the individual MLMs and the
MLM ensemble are compared in Figure 5-24 (a). When the prediction results are 100%
accurate, they would coincide with the red diagonal line. The predictions made with MLM
3, which uses only local measurements deviate significantly from the actual values, while
the predictions made with MLM 1, which directly uses voltage magnitudes and angles are
quite accurate. However, the most accurate predictions are obtained with the ensemble that
combines the outputs of all MLMs. The accuracy of predictions (RMS Error (RMSE) and
R2 value) obtained with different weightages in combining MLM outputs is shown in
Figure 5-24 (b). The 2:1:1 combination of MLMs 1, 3, and 4 is considered for the final
implementation, and it gives an R2 value of over 0.998 for the IEEE 14 bus system and an
R2 value over 0.975 for the IEEE 118 bus system.

(a) (b)

Figure 5-24: (a)LM prediction accuracy comparison for IEEE 118 bus test system,
(b) Effect of weightages on the accuracy of ensemble output
39
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

5.2.2.5 Real-Time Implementation and Testing

The proposed real-time VSM system was implemented and tested using experimental setup
shown in Figure 5-25, which consists of an RTDS® digital real-time digital simulator and
PhasorSmart software platform. The local area network implemented using a
RuggedCom™ RSG2288 utility grade Ethernet switch facilitates the communication of the
synchrophasor measurements from the PMUs simulated in RTDS to a PC that runs the
software platform. The synchrophasor application program consists of two main platforms:
ePDC virtual Phasor Data Concentrator (PDC) and Synchronous Data Server (SDS), which
are responsible for concentrating the PMU measurements from different PMUs. SDS
publishes data to the VSM application program developed using C++. This program
implements the VSI calculation and LM estimation using the trained MLMs. A monitoring
dashboard is implemented using web based Grafana visualization tool to provides a user-
friendly monitoring interface as shown in Figure 5-26 (a) [88]. A sample of LM predictions
using trained MLMs during various system events is shown in Figure 5-26 (b).

Figure 5-25: Experimental Setup

(a) (b)

Figure 5-26: Examples of (a) voltage stability monitoring dashboard, and (b)
variation of real-time predicted LM with system changes/events for IEEE 14 bus
test system

40
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

5.2.3 Detecting Relay Misoperations Using Field Data

Although relay non-operations are detected in real time and supported by local and remote
backup protection, their incorrect operations, or misoperations, are never detected in real
time. In a survey of four utilities conducted by PSRC, yearly misoperations were found to
be 15.4%, 9.5%, 52.6%, and 28.6% at 345 kV voltage levels [90]. A study of hidden
failures [91] shows these can be responsible for unintended operation of distance relays,
breaker failure relays, underfrequency relays, and undervoltage relays. Blackout reports
show that misoperations have contributed to either initiating or exacerbating the
progression of large-scale instabilities that take anywhere between a few minutes to a few
hours to result in blackouts [92]. Load encroachment, a common misoperation of distance
relays, has occurred during the majority of documented blackouts.
Since synchrophasor measurements capture signatures of dynamic events in power
systems, a disturbance identifier tool can be used to flag off any relay misoperation. The
concept, called supervisory protection and dynamic event visualization, is shown in Figure
5-27. As seen from the figure, the application would reside inside a PDC and classify
disturbances in real time. If a relay operates, but a fault is not detected by the classifier, a
flag is set off, indicating possible misoperation. If, on the other hand, classifier also detects
a fault, the relay operation is validated. This algorithm is not used to classify misoperations
due to slow operation or failure to operate when a fault is present.

Figure 5-27: Supervisory protection and dynamic event visualization.

There are many classifiers proposed to identify dynamic events using PMU data, but few
describe the exercise using field data. This section describes an effort made using field
data. It will underscore the obstacles faced in creating a classifier when field data are used.
It will also summarize how these obstacles can be tackled. A comprehensive description of
the material presented here can be found in [93].

41
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

The data available to the researchers in this case consisted of all the disturbance files
captured by four PMUs owned by a utility in the Southwestern USA from 2007 to 2010.
Data consisted of the positive sequence voltages, frequency, and rate of change of
frequency measured in the 345 kV transmission network at 30 fps reporting rate. After
discarding files which suffered significantly from data loss, and files that were simply
continuation of an event-recording, a total of 1013 disturbance event files were selected for
this study. These are not enough to train and test any classifiers, which constituted the first
hurdle. Additionally, the utility log captured only 84 of these 1013 events - 23 generation
loss events, 58 faults, and 3 line-trip events. Thus, majority of events recorded by PMUs
were unknown. Among them, many events might have occurred outside the utility’s
boundaries (i.e., in neighboring utilities). This situation creates two additional hurdles: 1)
the exact number of classes, which in this case would be types of disturbance events, are
unknown, and 2) ground truths required to test a classifier are inadequate.
To overcome these hurdles, the following strategy was adopted:

• Since the number of classes that would encompass all PMU disturbance data are
unknown, the first step is to know how many clusters these data can be segregated into.
One cluster would ideally correspond to one type of disturbance events and consists of
a group of disturbance files recorded for all occurrences of this type of events.
• After obtaining the clusters, each cluster will be mapped to a disturbance type. Ideally
ground truths would be used for this step. However, since the ground truths are
available for only three events, domain expertise was used.
• A pattern-creation technique and a classifier need to be chosen for precise
classification.
• Since the field data are inadequate to train and test a classifier, additional data will be
generated through simulation of the utility’s network. Preferably, field data will be used
for testing the chosen classifier.
An unsupervised clustering method – agglomerative hierarchical clustering technique
called AGNES (AGglomerative NESting) - was chosen to find clusters contained within
the 1013 disturbance files [94]. A previously published method for pattern creation –
Minimum Volume Enclosing Ellipsoid, or MVEE – was used to create and feed patterns
to the clustering technique. This method uses domain transform on all data-streams coming
to the PDC to create a multi-dimensional ellipsoid. The geometrical properties of the
ellipsoid are then taken as patterns. This is a computationally intensive approach which
was improved in the later part of our study. The results from the clustering technique are
shown in Table 5-4. Nineteen clusters emerged. Some of the clusters were identified
straight away as they contained the known ground truths. However, 10 clusters remained
unidentified. The good news was that no two ground truths fell into the same cluster, which
provided credibility to this approach.
In order to identify events falling in the unknown clusters, domain knowledge was used.
The unknown clusters were identified as shown in Table 5-5. Here, LS stands for load
switching, and SCS stands for shunt capacitor switching. Pumping units were synchronous
motors in the utility system. To further test the veracity of the disturbance types determined
by domain knowledge, all the disturbances were simulated on the PSLF® software, using
42
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

the WECC model of the utility system available in the software. Twenty instances of each
disturbance were simulated and then all the event files from simulated and field data were
grouped together and re-classified using the same method. The same 19 clusters emerged
and there was no cross-classification. This provided credibility to the process of
discovering groups hidden in the PMU dataset.

Table 5-4: Clusters formed by using agglomerative hierarchical clustering

approach.

43
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Table 5-5: Disturbance events identified for initially unknown clusters.

Once the disturbances were discovered, the next step was to build a classifier. However,
due to too few field data, more simulated data had to be generated for all types of
disturbances. Table 5-6 shows the simulated data generated for each of the seven major
events (line-trip was excluded).
Table 5-6: Number of simulated events for the seven disturbance types under study.

Since the disturbance traces were rich with time-domain signatures, a time-domain pattern
creating method called shapelets was adopted to extract the patterns. This method was
implemented in our work as domain-specific shapelets (Dshapelets) [95] to save the
computational efforts involved with domain transforms used by the MVEE method. Slope
Sequence of Shaplets (S3) was a pattern derived from the shapelets. A slope formed from
raw data (S2) was also chosen as a pattern. Other more popular patterns involving domain
transformation – Discrete Fourier Transform (DFT), Discrete Wavelet Transform (DWT),
fast variant of discrete S transform (FDST), and Principal Component Analysis (PCA) –
were also chosen for a comparative study. Finally, some statistical metrics formed from
time-series data were also chosen as a pattern. The measurement noise in the field
44
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

waveform [96] was kept as it was. Additionally, since data streams from all PMUs would
be affected in the same way by a given disturbance, the data stream with the strongest
signature was identified using an energy-based index, and used for the classification, thus
reducing the computation burden on the classifier [95].

Table 5-7: Training (3495 simulated events from 7 disturbance types), testing (81
actual events with FLT and GL)

Support vector machines (SVM) and K – Nearest Neighbor (1NN when K=1) were chosen
as classifiers, as they had shown the best performance in earlier studies [92]. Table 5-7
shows the results from training the classifiers for all seven disturbances with simulated data
and testing it with all the field data. Slope Sequence of Shaplets (S3) is seen to outperform
other patterns. It is observed that voltage and frequency both need to be used as inputs,
since each has unique signatures for disturbances [95][6]. Also observe that the raw data
based pattern (S2) performed better than the methods requiring change of domains. This
underscores the fact that there is no one “magic” pattern that works for all applications.

45
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Table 5-8: Training (90% of all 3495 simulated events with 7 disturbance types),
testing (10% of all 3495 simulated events with 7 classes); ten-fold cross validation.

Now, the classifier was tested for all seven events using only simulated data due to
unavailability of field data for all seven disturbances. The results are shown in Table 5-8.
Although both S3 and S2 patterns give the best performance, the accuracy drops. Further
investigation revealed that the only confusion was between Generation Loss and Load
Switched On, since they impact voltage and frequency in the same way from a physical
viewpoint, so the signatures are not distinguishable. This is a drawback of the data-driven
method. To overcome this drawback, a physics-based method based on the system energy
function was designed [93]. In this method, components of the energy function were
monitored to track each event. Figure 5-28 (b) shows the component that captures the
change in energy due to loss of load and Figure 5-28 (a) shows the component that captures
change in energy in generator mechanical input [93]. The two events are distinctly
different. This exercise underscores the importance of an approach of physics-informed
machine learning, whenever feasible.

(a) (b)

Figure 5-28: Energy change pattern due to (a) loss of generator mechanical input,
and (b) loss of load
46
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Finally, it was observed that faults had such distinct signatures that they were identified
with 100% success. Time taken to determine whether an event is a fault or not was less
than 0.4 ms. Even after adding sensing, PMU measurement and communication delay,
which are reasonably estimated between 50-130 ms according to the IEEE C37.118.2-2011
standard, the supervisory protection can be performed in real-time, because the time taken
for misoperations to cause large scale instabilities has been observed to be in minutes to
hours. Most of the misoperations do not even result in instabilities.

5.2.4 Post-Event Analysis

In digital substations, IEDs with event recording capabilities are used for the protection
and control functions of the power equipment. Some of the substations also have Digital
Fault Recorders (DFRs) that capture event data for post-event analysis. When a fault or an
event happens in the power grid, the IEDs and DFRs can perform the pre-fault and after-
fault data recording of circuit statuses and electric parameters by event triggers. A
cascading fault event often involves a large area and many IEDs. To manually analyze the
root cause of the fault requires a lot of engineering effort. The post-event analysis can
utilize AI/ML to help with creation and analysis of fault trees using the protection and
control logic in the IEDs [97]. Each primary equipment (e.g., transformer, line, generator.)
has a fault tree linked to the IEDs that are used to perform the protection and control. It
starts by collecting the event COMTRADE files from the IEDs and finding the matched
fault trees in the library such that the picked up elements will be searched through the tree
and highlighted in the fault tree together with a time stamp. For a cascading fault event that
involves multiple IEDs, the tool will highlight all the fault trees involved and create a
combined COMTRADE file with all the elements that have picked up.
A Finite State Machine (FSM) model of the post-event analysis and visualization is shown
in Figure 5-29. When started, the system is continuously monitoring the event trigger. The
monitoring time window is user configurable and any other triggers received within this
window of the aforementioned trigger will be considered as related events and similarly
processed. The monitoring time window is also a wait-time for “Event/COMTRADE
Collect” state. All the IEDs that have new Event or COMTRADE files will be
automatically connected for event file downloading and the retrieved files will be stored in
a specific folder for analysis.

47
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 5-29: FSM Model of the Post-Event Analysis

The application takes the new event files and processes them by matching the fault trees in
the database. Both the IED name and event tag will be used for the searching criteria. Once
the event tag is found in the fault tree, it will then be highlighted in the fault tree and
displayed with the event pickup time. Figure 5-30 shows an example of traversed fault tree
after a transformer bank tripped. The user can document the instantiated fault tree and raw
data for later reporting.
In the case of an event tag which was picked up but is not included in the fault tree, the
user can choose to add it to the fault tree or to the ignoring list. Machine learning can be
implemented in the tool with some built-in operation matrix for each type of relay. The
ignoring list contains the list of event tags that may be duplicated or not related to the root
cause of fault, thus no need to include in the fault trees.

48
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 5-30: Fault Tree Construction for Transformer Bank X

5.2.5 Application of AI/ML in Travelling Wave Protection

This section provides application examples of ML for traveling wave (TW) protection of
AC and DC systems. In the following, first, an ML-based algorithm for the TW protection
of AC distribution systems is discussed. Then, ML application for the TW protection of
DC systems is elaborated. While TW protection is well developed for AC transmission
49
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

lines and HVDC lines, it is significantly more difficult to apply to distribution systems with
many taps, transformers, and customers. For this application, AI/ML is used to learn the
expected high-frequency fault signatures from multiple reflections in networks to
determine the fault type and location in less than 1 ms.
5.2.5.1 TW Protection Background
TWs are electromagnetic waves propagating along the power system equipment such as
lines or cables when a disturbance (e.g., fault, lightning, switching, etc.) occurs. When the
high-frequency TWs reach a new environment with different circuit parameters (e.g., at a
line terminal), a portion of the incident TW is reflected while the other portion is refracted
to the neighboring equipment. Depending on the circuit parameters at both sides of the
terminal, the amplitude of reflected and refracted TWs changes accordingly. The TWs are
also reflected and refracted again after they reach the fault location or line terminals. One
of the common approaches to extract high-frequency TWs is to utilize Discrete Wavelet
Transform (DWT). A DWT is any wavelet transform (WT) for which the wavelets are
discretely sampled. The WT has been widely utilized as an effective tool for the
simultaneous analysis of waveforms in time and frequency domains.
In order to effectively construct wavelets over a wide frequency range, multiresolution
analysis (MRA) is a practical approach for fully implementing the DWT. MRA details the
procedure to obtain an orthonormal wavelet basis with compact support. MRA can be
implemented by a series of high-pass and low-pass filters and decimators as shown in
Figure 5-31. As seen, the outputs of low and high-pass filters at each level are ai [n] and
di [n] . The output of the low-pass filter at each level is passed through the next level for
constructing wavelets for the next decomposition level. The low-pass filter outputs are
referred to as scaling coefficients while the high-pass filter outputs are called wavelet
coefficients. Assuming that the initial sampling frequency of WT is set as f s , then the
frequency range of each level is shown in Figure 5-31. The wavelet coefficients are of
interest since they better represent the high-frequency behavior of TWs.
Once the MRA-based wavelet coefficients are identified, Parseval’s theorem is used to
calculate the energy corresponding to the identified coefficients. If the scaling function and
mother wavelet form an orthonormal basis, then Parseval’s theorem can be used to build a
relationship between the calculated wavelet coefficients and the energy spectrum of the
fault signal (i.e., cable’s voltage or current measurement). Under this condition, Parseval’s
theorem states that the energy of the fault signal can be described mathematically in terms
of the expansion coefficients (i.e., the integral or sum of the square of the original function
is equal to the sum of the square of the coefficients). The DWT can split the energy of fault
signals in time and frequency domains. With Parseval’s theorem, one can effectively
interpret the high-frequency signatures of TWs by relating the current or voltage of the TW
energy to the energy of the wavelet coefficients. The fault signal can be described by the
wavelet coefficients of different ranges.

50
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Sampled Signal

GLPF,1[n] HHPF,1[n]

2 2

a1 d1
[fs/4 , fs/2]
GLPF,2[n] HHPF,2[n]

2 2

a2 d2
[fs/8 , fs/4]
GLPF,3[n] HHPF,3[n]

2 2

a3 d3
[fs/16 , fs/8]

Figure 5-31: MRA block diagram.

5.2.5.2 TW Protection of AC Distribution Systems Using ML [98]:

The goal of this approach is to use TW measurements as inputs to a machine learning
algorithm to detect and locate different types of faults in a distribution feeder.
ML Approach
The method can be separated into two main stages: the signal processing block that aims
to process the measured signals (+/- 50 μs since TW time of arrival) and extract useful
information, and the Machine Learning stage where several algorithms are trained to
perform the actual fault location and classification. The workflow is shown in Figure 5-32.
Signal Processing: As it can be seen, the first part of the signal processing block consists
of applying some transforms to both the 3-phase voltage and current signals. Here, the
Clarke, Karrenbauer, and DQ0 transforms are employed. Numerical results showed that
the larger number of features, the better performance. Each transform performs a different
operation on the measured data, which is later used to create additional features to train the
ML algorithms.

51
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 5-32: AI-based traveling wave fault location algorithm

Herein, a comparison between several algorithms has been made, with special emphasis on
ensemble learning, which consists of combining multiple ML models in order to create a
more accurate and powerful model. Currently, it is one of the most used techniques to
improve the performance. There are several techniques to implement ensemble models,
such as bagging, boosting, or stacking. For comparison, a Random Forest (RF) algorithm
(example of bagging), implemented on scikit-learn library, has been selected, along with
Google’s TensorFlow Boosted Trees (BT) estimator (Gradient Tree Boosting). Some
notions of these techniques will be given later. These particular algorithms have been
selected because of their excellent accuracy, computational efficiency, and scalability for
a larger number of samples/ larger system (which is where they outperform other common
ML algorithms, such as Support Vector Machines (SVMs)). Finally, some of the state-of-
the-art Machine Learning algorithms integrate different types of models into one ensemble
model that takes advantage of the individual skills of each of the components (known as
stacking). In this line, one model that groups the two aforementioned individual methods
(RF and BT) has been implemented.
Simulation Test System: The system consists of one 12.47 kV variable-length distribution
line that is connected to a load of 300 kVA per phase. Faults occur on the load bus.
Simulations were performed in PSCAD for 160 fault locations (increasing the length in
steps of 25 meters), for 3 fault types (single-line-to-ground, line-to-line, and 3-phase), and
for 7 resistance values ranging from 0.01 to 10 Ω. This sums up to 3360 simulations. Three-

52
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

phase voltages and currents are measured on the secondary side of the transformer at a
sampling frequency of 10 MHz. The system is shown in Figure 5-33.

Figure 5-33: Schematic of the system

Fault Type Classification: For the fault type classification task, the algorithms are able to
discern whether the measurements are of a Single-Phase fault, Double-Phase fault, or
Three-Phase fault. Accuracy is defined as the number of correctly predicted fault types
over the total number of faults. Table 5-9 below summarizes the individual results of both
RF and BT algorithms. As it can be observed, accuracy is close to 100% in both of them.
The reason for this excellent performance is the large energy magnitude variations among
each type of fault, which makes this task easy.

Table 5-9: Fault type classification accuracy.

Classifier Accuracy
Boosted Trees 99.84%
Random Forest 100%

Fault Location: The results for the Boosted Trees and Random Forest regressors, along
with the Stacking method can be observed in Figure 5-34. The average and the standard
deviation of the errors are shown in Table 5-10. Most of the faults are located with an error
well below 200 m.

53
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 5-34: Fault location prediction errors relative to fault distance

Table 5-10: Mean and STD of prediction errors for fault location.

Standard
Regressor Mean (m)
Deviation (m)
Boosted Trees 72.64 99.14
Random Forest 53.64 113.01
Stacking 62.95 94.77

As it can be seen in Figure 5-34, the Random Forest model tends to have lower errors in
most of the cases, but there are a few outliers. However, the Boosted Trees algorithm is
much more consistent. The stacking of both algorithms leads to lower values and deviation
in prediction errors.
5.2.5.3 TW Protection of DC Systems Using ML [99]:
In order to study the high-frequency fault signatures of DC systems, a simple DC system
(shown in Figure 5-35) is modeled in PSCAD/EMTDC. This circuit includes a controllable
DC voltage source, one cable with the length of 3000 m, and a DC load with a resistance
of 10 Ω. The nominal voltage of this system is ±375 V. The cable is modeled using the
frequency-dependent distributed parameter model available in PSCAD/EMTDC. The
cable specifications are provided in Figure 5-36. It is assumed that each pole is buried 1 m
deep. The core conductor resistivity is 2×10-8 Ωm while sheath resistivity is 30×10-8 Ωm.

AC

DC Sensor C1
AC Grid Main Load
Converter

Figure 5-35 Simple DC microgrid system

54
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

0.5m

Sheath

Core
Conductor

0.015m
0.017667m
0.020667m

Figure 5-36: Cable configuration of the DC system in Figure 5-35

MRA is an effective tool to demonstrate the high-frequency fault signatures in different

frequency ranges. In order to show how the Parseval energy of wavelet coefficients of each
MRA’s level is impacted for different fault locations, phase-to-phase faults are applied at
every 75 m of cable C1 in Figure 5-35. First, it is assumed that DWT’s sampling frequency
is 1 MHz. The Parseval energy values for three levels of MRA applied to the current
measured at the sensor in Figure 5-35 are shown in Figure 5-37. Herein, levels 1, 2, and 3
are associated with [250 kHz, 500 kHz], [125 kHz, 250 kHz], and [67.5 kHz, 125 kHz]
frequency ranges. As seen in Figure 5-37 (a), the Parseval energy value is generally
decreasing as the fault location gets closer to the end of the cable, however, some local
peaks are observed that happen at every 375 m. A similar pattern is observed in Figure
5-37 (b); however, the local peaks occur at every 750-800 m. Finally, Figure 5-37 (c) shows
that the local peaks occur at every 1500 m. As a rule of thumb, the number of local peaks
approximately doubles from level 3 to level 2 and as well as from level 2 to level 1. In
Figure 5-38, it is assumed that the DWT’s sampling frequency is 2 MHz. Doing so, level
1 and 2 are associated with [500 kHz, 1 MHz] and [250 kHz, 500 kHz] frequency ranges,
respectively. As seen in Figure 5-38, a similar pattern to Figure 5-37 can be observed. In
general, with a higher DWT’s sampling frequency, (i) more oscillations on the Parseval
energy profile of fault currents are observed, and (ii) the first incident of TW can be
detected faster. The latter is based on an inherent feature of TWs in which the higher
frequency TWs travel faster with a lower magnitude.

(a)

55
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

(b)

(c)

Figure 5-37: Parseval energy values for different fault locations on the cable of DC
system with 1 MHz sampling frequency in Figure 5-35: (a) MRA’s level 1; (b)
MRA’s level 2; (c) MRA’s level 3.

Figure 5-38: Parseval energy values for different fault locations on the cable of the
DC system with 2 MHz sampling frequency.

The proposed fault classification and location algorithm is shown in Figure 5-39.

56
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

PPG
Voltage Level 1
Level 2 Parseval EPRS,VPG
MRA ∑
Energy

Level N
Calculate
EPRS,VPG /EPRS,VPP

PP Level 1
Voltage Parseval
Level 2
MRA ∑
Energy EPRS,VPP

Level N

> λFT
Yes No

PG Fault PP Fault

Pole Current
Pole Current Parseval Energies
Parseval Energies for N Levels
for N Levels SVM SVM
PPG Voltage Classifier Classifier PP Voltage
Parseval Energies Parseval Energies
for N Levels for N Levels

Bolted Resistive Bolted Resistive

Fault Fault Fault Fault

Select the GP
engine based
of fault type.

Pole Level 1 Gaussian

Current Parseval Process Fault
MRA Level 2
Location
Energy Regression
Engine
Level N

Figure 5-39: AI-based traveling wave fault classification and location algorithm for
DC systems

Fault Classification: To distinguish between Pole-to-Pole (PP) and Pole-to-Ground (PG)

faults, the Parseval energy of PP voltage at the sensor location is compared against the
Parseval energies of the positive or negative pole to ground voltage. This procedure is
shown in Figure 5-39. Simulation results show that for PG faults, the Parseval energies of
positive or negative pole to ground voltage (PPG or NPG) are significantly higher than the
Parseval energy of PP voltage. To this end, the algorithm first calculates the summation of
the first N levels of Parseval energy values related to PP voltage and PPG voltage (i.e.,
EPRS,VPP and EPRS,VPG respectively). Then, EPRS,VPP and EPRS,VPG are compared against each
other to determine the fault type. Herein, the ratio of EPRS,VPG /EPRS,VPP is calculated and
compared against λFT threshold. This threshold can be found by trial and error on the
microgrid system. Since EPRS,VPG is significantly higher than EPRS,VPP for PG faults, λFT is

57
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

always greater than 1. As a rule of thumb, the threshold is selected at around 10% of the
EPRS,VPG /EPRS,VPP ratio for the remote end PP and PG faults.
Once the fault type (i.e., PP versus PG) is identified, a Support Vector Machine (SVM)
classifier is used to determine if the fault is resistive or bolted. SVM has been proposed as
a strong classification tool. For this purpose, multiple bolted and resistive faults are
simulated at different locations of a cable (e.g., every 25 m) in a simulation software
package (e.g., PSCAD/EMTDC). It should be noted that the fault resistance values adopted
in the simulations depend on the DC microgrid conditions like voltage level or
geographical location. After the simulation results are gathered, the SVM classifier is
trained using the labeled Parseval energy values. For PP faults, the inputs to the SVM
classifier are the N level Parseval energy values of pole current and PP voltage at the sensor
location. For PG faults, the inputs to the SVM classifier are the N level Parseval energy
values of pole current and PPG voltage at the sensor location. The output of the classifier
is the fault resistance value.
Fault Location: Once the fault type is classified, first, the proposed approach determines
if the fault is located on the primary cable. To this end, the Parseval energy value of the
current flowing through the protection relay sensor, EPRS,I, is calculated. Depending on the
fault type, EPRS,I is compared against the precalculated Parseval energy value related to the
remote end bolted PP, resistive PP, bolted PG, or resistive PG fault. The proposed fault
location algorithm relies on the Parseval energy values gathered from MRA. The algorithm
utilizes the first N levels of MRA, calculates the Parseval energy of the first TW incidents,
and then utilizes Gaussian Process (GP) regression engines to find the fault location.
Simulation Results: The DC microgrid test system is illustrated in Figure 5-40. This DC
microgrid system is based on a real DC microgrid system in the city of Albuquerque, NM.
We have utilized high frequency (in the order of 10 MHz) field measurements from the
actual microgrid to calibrate the created model in PSCAD/EMTDC. The DC microgrid is
supplying four residential houses. Each residential house is described as a nanogrid (NG).
Each NG included the residential house load, a PV system, a Battery Energy Storage
System (BESS), and DC-DC converters to integrate NG into the rest of the microgrid. The
PV system is associated with a maximum power tracking scheme. The size of the PV
system in each NG is 10 kW while the size of BESS is 6 kW/12 kWh. The load of NG1,
NG3, and NG4 is 56.25 kW and the load of NG2 is equal to 50 kW. It is assumed that the
microgrid includes a community BESS and PV system with the size of 20 kW/40 kWh and
18 kW, respectively. The microgrid’s main converter is modeled as a multi-level voltage
source converter. The microgrid’s grounding happens at the middle point of the DC link of
the microgrid’s AC/DC converter. The size of this converter is 500 kW. Load L1’s size is
10 Ω. RC1 is 25 mΩ and LC1 is 20 µH. In Figure 5-40, relays R1 to R6 identify the
location of the proposed fault detection and location algorithms. The cables configuration
is provided in Figure 5-36. It is assumed that each cable has its own local protection.
To verify the effectiveness of the proposed scheme, bolted and resistive PP and PG faults
are applied at different locations of all six cables in the DC microgrid test system shown in
Figure 5-40. The fault resistance for both PP and PG faults is equal to 5 Ω. The length of
each cable and its simulated fault locations are summarized in Table 5-11. In these
58
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

simulations, DWT’s sampling frequency is 8 MHz. Six levels of MRA are used for the
fault classification and location algorithm. For cables C1 and C6, 40 different fault
locations are simulated; for cables C2 and C5, 30 different fault locations are simulated;
for cable C3, 38 different fault locations are simulated; for cable C4, 22 different fault
locations are simulated.
AC RC1 LC1
NG1
DC R1 C1 R6 C6
AC Grid Community
Main
PV/Storage
Converter

NG4
L1 R2 C2 R4 C4 R5 C5
R3
NG3
C3

NG2

Figure 5-40: DC microgrid diagram

Table 5-11: Cable lengths and fault locations

Cable C1 C2 C3 C4 C5 C6
Length 2000 m 800 m 1000 m 600 m 800 m 2000 m
Fault At every At every At every At every At every At every
locations 50 m 25 m 25 m 25 m 25 m 50 m

The algorithm was able to effectively distinguish PP faults from PG ones by selecting 10
as the λFT threshold in Figure 5-39. To verify the performance of the SVM classifier for
distinguishing bolted faults from resistive ones, two separate datasets were used for training
and testing. 60% of the available data were used for training and 40% of them were used
for testing randomly. In the SVM, a linear Kernel is used while the penalty factor for
misclassified data is set to 1. The verification results rendered 100% precision in
classifying bolted faults versus resistive faults using the six levels of current and voltage
Parseval energies at cable C1. In Table 5-12, the fault location estimation errors using GP
engines for all cables and different types of faults are summarized. The estimation error
percentage is equal to the mean absolute error of the testing dataset over the length of the
cable. For each cable, around 65% of the gathered datasets are used for training and the
rest are used for testing. The training and testing datasets are selected randomly. The fault
location estimation errors in Table 5-12 verify the effectiveness of the proposed fault
location algorithm. In Table 5-12, the estimation error depends on the number of datasets
available for training and the length of the cable. In general, simulating faults at more
locations can increase the number of training datasets which in turn improves the
performance of the GP regression engine in estimating fault location. The regression results
for Cable C4 are illustrated in Figure 5-41. As seen, for both resistive and bolted PP and
PG faults, the GP regression engine can effectively locate faults with small estimation
errors. In all cases, the proposed algorithm is able to find the fault location in 200 µs.

59
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

600 600
Actual
Actual
Estimated
Estimated
500 500

400 400

Fault Location (m)

Fault Location (m)

300 300

200 200

100 100
1 2 3 4 5 6 7 1 2 3 4 5 6 7

Testing Dataset Testing Dataset

(a) (b)
600 600
Actual Actual
Estimated Estimated
500 500

400 400

Fault Location (m)

Fault Location (m)

300 300

200 200

100
100
1 2 3 4 5 6 7
1 2 3 4 5 6 7
Testing Dataset
Testing Dataset

(c) (d)

Figure 5-41: Regression verification plots for C4: (a) bolted PP faults; (b) resistive
PP faults; (c) bolted PG faults; (b) resistive PG faults

Table 5-12: Fault location estimation error for DC Microgrid 1

Cable Bolted PP Resistive PP Bolted PG Resistive PG

C1 4.9% 3.6% 2.5% 4.1%
C2 5.1% 2.9% 3.9% 1.1%
C3 4.7% 6.1% 4% 2.7%
C4 1.1% 0.4% 2.1% 0.2%
C5 6.3% 0.6% 0.8% 0.6%
C6 5.3% 4.8% 6.2% 3.4%

In the Table 5-13 below, the performance of GP for fault location is compared against some
other regression techniques, including Artificial Neural Network (ANN), Decision Tree, ε-
SVM, and Nu-SVM. The comparisons are performed for Cable C2. As seen, GP renders
higher accuracy compared to the other regression techniques.

60
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Table 5-13: Comparison of different ML techniques for Cable C2

Fault Type GP ANN Decision Tree ε-SVM Nu-SVM

Bolted PP 5.1% 8.6% 9% 8.2% 15%
Resistive PP 2.9% 4.8% 6% 5.9% 14%
Bolted PG 3.9% 8.9% 5.9% 7.7% 9%
Resistive PG 1.1% 2.9% 2.2% 2% 6%

5.2.6 AI Based Wide Area Control with Windfarm

In this section, an AI engine-based real-time wide-area damping controller design
considering multiple synchronous generators and wind farms is discussed. The architecture
is based on, a reinforcement learning (RL) Adaptive Critic Design (ACD) framework
designed and tested to perform the optimal control of multiple generators based on energy
function theory. The approach connects Lyapunov energy function theory with machine
learning to design an architecture that damps inter-area oscillations.
5.2.6.1 Introduction
Real-time transient stability assessment and wide-area protection and control have been
studied widely in the literature. Methods such as numerical integration and direct energy
functions have been discussed and their advantages and disadvantages are mentioned in
the earlier studies. Numerical methods are in general vulnerable to converge in real-time
and direct energy functions, even though promising, are difficult to model without
excessive simplifications. Measurement-based learning framework in this sense has a
better advantage as such methods can be developed with AI and can be trained using offline
and online techniques. This increases the chances of convergence as learning accurate but
complex models through training provides accurate results with successive training
approaches. RL-based offline and successive online training is a great candidate for such
application without a prior knowledge of the learning objectives. The method can thus be
used as a wide area system-centric controller and observer (WASCCO) and can be
deployed based on a central learning framework (such as cloud computing) for offline
learning and a grid edge device for subsequent online training/learning.
5.2.6.2 Problem Statement
Rotor angle stability theories focus on the machine angle oscillations and transient energy
that creates such oscillations at the advent of an event such as a fault. Transient energy
interactions and instabilities are caused due to the power mismatch between the generator's
electrical and mechanical power. This can be captured in the form of a swing equation.
During internal oscillatory events such as a fault, the machine rotor angles oscillate.
Modeling of the generator dynamics can be presented classically using a third-order model
of the conventional generator including the generator angle, speed, and excitation voltage
equations with a Center of Inertia (COI) as reference. Then such energy functions can be
related as a Lyapunov function such that a cost function that is proportional to the square

61
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

of the generator speed and voltage deviation can be defined. Minimizing the cost function
then means minimizing the deviations in the voltage and speed in turn minimizing the
transient energy that is generated during an abnormal event such as a fault. From the third-
order dynamics of the generator including the excitation system, an energy function as a
cost function can be developed where J represents the total cost function. The goal of the
RL-based AI is to learn this cost function from the deviation in voltage (ΔV) and speed
(Δω) measurements as x(t). The output u(t) is the voltage or power reference. The overall
process is illustrated in Figure 5-42.
5.2.6.3 Machine Learning Process
The process of intelligent system construction starts with developing a database for training
three neural network engines viz., Wide Area Neural network identifier (WANNID), critic
neural network (Critic NN), and an action neural network (action NN). These three neural
network engines conform to the family of ACD’s.

Figure 5-42: Wide Area System Centric Controller (WASSCO) design [100]

Figure 5-43: Reinforcement learning process [101]

The goal of ACD is to learn the Hamilton-Jacobi-Bellman equation based on optimal

control theory through the Critic NN and find the right control or policy signal through the
62
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

action NN. Learning is done through optimization over time. ACDs are developed as
parametric structures that can optimize over time including noise. During offline training,
the goal is to minimize the estimated cost function J. For this purpose, three Feed Forward
Neural Networks (FFNN) are designed. The output of each network is computed by
calculating the inner product between the Weights (W) and the state-dependent vector (ɸ).
WANNID develops the model of the system based on measurement, CriticNN optimizes
the cost function and critic the policy, and Action NN develops the control signal. This
process is repeated until convergence is achieved. Action NN is connected to the
generator’s exciter. In the online training phase, the training for WANNID, CriticNN, and
ActionNN is performed sequentially in that order.
5.2.6.4 Implementation and Testing
The implementation platform is illustrated in Figure 5-46. The offline training is performed
using a standard intel computer based on the model output from the real-time simulator.
Once the offline training is completed, the controller is deployed on an edge device (TI
controller). The measurements are compiled using the PMU and data concentrator. The
online training process is based on the data streamed from the concentrator.

Hardware implementation

Controller training

Figure 5-44: Implementation platform

a) Power grid model with wind farm b) Oscillations in angle and speed

Figure 5-45: Visualization for a two-area power grid with fault on area 2 [100]

63
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

For visualization purposes, a self-clearing fault in area 2 is illustrated in Figure 5-45(a).

The fault is kept for 30 ms. The dynamics of the rotor angle and speed of one generator are
illustrated in Figure 5-45(b). The illustration shows oscillations for four scenarios, a) with
no damping controller (labeled as No control) b) with a Power System Stabilizer (labeled
as PSS), c) an optimal control with the mathematical model of the system (labeled as Conv.
WAC) and d) with the ML-based wide-area control (labeled as Prop. WAC). It can be seen
that with suitable ML architecture and training the damping controller can outperform the
conventional WAC and PSS.

5.2.7 Transient Instability Predictions

5.2.7.1 Introduction
Although power systems are designed to recover after being subjected to various
disturbances such as faults, loss of generation or large loads, they are not immune to every
disturbance which they may encounter. Large disturbance rotor angle instability events
appear as aperiodic angular separations due to insufficient synchronization torque. Fast
detection of impending rotor angle instabilities can allow taking automatic emergency
control actions such as load or generator shedding to avoid potential system wide
instabilities or blackouts.
5.2.7.2 Problem Statement
The typical local measurement-based out-of-step protection provided by distance relays
has a limited capability of identifying an area wide disturbance. This shortcoming is
typically addressed by using wide-area special protection system (SPSs). These systems
are usually event-based systems designed to directly detect pre-identified outages which
lead to rotor angle instability using status signals and initiate predetermined remedial
actions. Implementation of an event based SPS can be complex, cumbersome, and costly
due to the necessity of large number of teleprotection systems to convey all the system
contingency information.
There is a potential for developing simpler response based wide area protection systems
using synchrophasor measurements and Machine Learning techniques. Typically, the
generator rotor variations are used to assess the rotor angle stability. The post-fault
generator voltage magnitude and speed measurements also contain valuable information
about the stability status after the disturbance. However, it is difficult to develop a simple
physics-based relationship between the stability and the post-disturbance generator voltage
magnitudes, rotor angles, and the speeds without resorting to full time domain simulations.
Thus, it is useful to explore the potential of using system wide synchronized measurements
and AI/ML techniques for predicting large disturbance rotor angle instabilities.
5.2.7.3 Rotor angle stability prediction scheme
Post-fault variations of generator voltage magnitudes, rotor angles, and frequencies of a
power system subjected to a three-phase short circuit are in Figure 5-46(a). These
waveforms are obtained by simulating a fault in the 39-bus New England test system. The

64
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

fault is applied at 0.1 s and cleared after five cycles, and the system exhibits rotor angle
instability after the fault.
It is hypothesized that variation of voltages, phase angles and frequencies immediately
after clearing the fault can indicate whether the system is going to be transient stable after
the disturbance. Thus, it is possible to find a relationship between the observed variations
and the transient stability status. For a multi-machine system, this relationship is not
straightforward, but can be learned from examples using machine learning techniques. In
this study, a binary classifier is used to learn this complex relationship. The arrangement
of a transient stability status prediction scheme based on this principle is shown in Figure
5-46(b) and (c). The inputs to the classifier are the sampled values of the predictor variable,
which could be the generator rotor angles, speeds, or the voltage magnitudes, while the
output is the stability status. The algorithm is triggered by the voltage dip due to a fault,
monitors the bus voltage recovery profiles following the clearance of the fault by the action
of local protection relays, and collect the synchronously measured input features to the
classifiers [102].

(b)

(a)
(c)

Figure 5-46: (a) Variations of the voltage magnitudes, rotor angles, and generator
speeds during a contingency (b) Synchronized sampling of signals using PMUs, and
(c) Arrangement of the transient stability prediction scheme

5.2.7.4 Application to IEEE 39 bus test system

A Support Vector Machine (SVM) classifier is trained to predict the transient stability
status after a fault. The training and testing data for the IEEE 39-bus system is generated
using dynamic simulations in PSS/E software, considering a large number of pre-fault
operating points and contingencies. The process is shown in Figure 5-47(a). The results of
65
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

on-line prediction accuracy with testing data are shown in Figure 5-47(b). The graph shows
the variation of accuracy for testing data with the number of sample points considered in
the input data window.

(a) (b)

Figure 5-47: (a) Training process (b) Rotor angle prediction accuracy with different
predictors and data window lengths.

The samples of all variables are measured at a rate of one sample per cycle (60 fps). Each
point on each line corresponds to a different classifier configuration. For example, the SVM
corresponding to the first point on the rotor angles curve takes 40 inputs (4 samples of 10
generator rotor angles) while the SVM corresponding to the last point on the rotor angles
curve takes 120 inputs (12 samples from 10 generators). The best accuracy was obtained
with the classifier which used the voltage magnitudes as inputs. This classifier achieved
over 98% accuracy with just four consecutive samples of each generator bus voltage. The
generator speeds are also good predictors of transient stability, but to achieve an accuracy
comparable to that achieved with generator voltage magnitudes, 12 consecutive samples
were required. The sooner the prediction is completed, the longer the time available to take
control actions to avoid a possible system collapse, thus voltage magnitudes are the best
predictors for this application.
The classifier that uses voltage magnitudes is further tested and some results are presented
in Table 5-14, Table 5-15 and Table 5-16. For these tests, the system was simulated using
PSCAD and contingencies included unbalanced faults. The classifiers were applied phase-
wise, and a data window of just four samples were considered.

Table 5-14: Overall prediction accuracy

66
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Table 5-15: Prediction accuracy under topologies not contained in training data

Table 5-16: Prediction accuracy with noisy data after re-training the classifiers with
noisy data

5.2.7.5 Conclusions
Transient stability can be predicted accurately and fast using post-fault bus voltage
magnitudes. According to Table 5-14, the scheme has 100% security, and shows 97%
accuracy in correctly predicting unstable cases. The noisy data reduces the accuracy, but
when the classifiers are re-trained with noisy data, accuracy improves. The classifiers are
able to predict the instabilities with over 95% accuracy, even under topologies that were
not included in the training data patterns. The study also showed that prediction accuracy
increases when the loads include induction motors. Simulation studies with Venezuelan
power system showed 100% accuracy [102]. A more sophisticated version of the
classification scheme is presented in [103].

67
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

6 PRACTICAL APPLICATION CONSIDERATIONS

Once the acceptance criteria discussed in Section 7.3 is satisfied, there are several logical
steps to be taken to implement protection and control using artificial intelligence. This is
to allow that the grid incrementally is introduced to protection that is data driven and
security, reliability, and resiliency are not compromised.
The applications can be in individual devices and functions, substation functions or system-
wide functions. This part of the report will cover driver, benefit, industry use cases,
algorithms, AI approach and potential challenges for various applications.

6.1 Application Implementation

Application implementation includes understanding the data types, sources, formats and
management, defining the data structure, computational platforms, and learning
framework, testing, and validation. Each of these are discussed in the following
subsections.

6.1.1 Data Types, Sources, Formats, and Management

This section focuses on the sources of data for AI and ML applications in protective
relaying as well as data formats used by the data sources and common methods used to
manage the data. In the text below, the word ‘sensor’ is used as a generic term that
encompasses the IEDs, merging units, PMUs, etc.
6.1.1.1 Data Types
Sensors that are deployed in an electric power system can record many different data types,
such as waveforms, rms samples, phasors, and more.
6.1.1.1.1 Waveform Samples
Waveform measurements record short periods of three-phase instantaneous voltage and/or
current and can have varying sample rates (Figure 6-1). They can be triggered by
programmed algorithms detecting electrical disturbances or scheduled periodically. Most
measurements have defined start and finish times based on the triggering algorithm. Some
sensors offer continuous waveform recording and may use lossy compression and a FIFO
buffer. Waveform sampling can be synchronized to either an external clock or the current
system frequency. Additional channels, such as neutral current, and derived values like
residual current and system frequency may be included. Alongside waveform samples,
floating point values of intermediate or final values of relay targets or flags, derived from
the samples, may also be provided at the same sampling rate.

68
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Va Vb Vc Ia Ib Ic
100000

50000
Voltage (V)

-50000

-100000
2000

1000
Current (A)

-1000

-2000

0.00 0.05 0.10 0.15 0.20 0.25 0.30

Time (s)

Figure 6-1: Example Waveform Chart Showing Voltage Samples and Current
Samples

6.1.1.1.2 Digital Status Values

Many sensors that record waveforms or rms samples also record digital status values. The
values can be used to represent a time series that has two states, such as on/off, high/low,
true/false, or 1/0.
Figure 6-2 provides an example with voltage and current waveforms and three digital status
values (Feeder 05, Feeder 06, and Feeder 10), which provide the state of the breaker for
each of these three feeders. The digital status for Feeder 05 shows a thin line when the
associated circuit breaker is closed and a thick line when the associated circuit breaker if
open.

Figure 6-2: Example Chart Showing Digital Status Values with Voltage and Current
Waveform Samples
69
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

6.1.1.1.3 RMS Values

RMS values can be recorded due to a fault or disturbance at the same time as waveforms
or in place of waveforms. They can be comprised of the root mean square (RMS) value of
a single cycle or can be the averaged over a window such as 2 cycles, 3 cycles or 200 ms
(as per IEC 61000-4-30). Figure 6-3 is an example chart showing the voltage and current
RMS values derived from the waveform in Figure 6-1.

Va Vb Vc Ia Ib Ic
65000

60000
Voltage (V)

55000

50000

1200

1000
Current (A)

800

600

400

0.0 0.1 0.2 0.3 0.4 0.5 0.6

Time (s)

Figure 6-3: Example RMS Values Chart with Voltage and Current Samples Derived
from Waveform from Figure 6-2
6.1.1.1.4 Phasor Values
Phasor values can be recorded to summarize the magnitude and/or phase angle of voltage
and current values. Phasor values are usually derived for the fundamental frequency
component of the power system (e.g., 50 Hz or 60 Hz) but they can be derived for
harmonics of the power system (e.g., 3rd harmonic for a 50 Hz system or 150 Hz). Figure
6-4 is an example showing three-phase voltage and current phasor values.

70
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Figure 6-4: Example Phasor Chart Showing Voltage Phasor (Per Unit) and Current
Phasors (Amps)

6.1.1.1.5 Time Series (Data Logs)

Time series records a range of basic and advanced electrical quantities, such as rms voltage,
rms current, phasor magnitude, phase angle, line frequency, power factor, real power,
reactive power, apparent power, total harmonic distortion, and harmonic magnitudes.
Sensors can be programmed to provide these values at regular intervals, like every 15
minutes, 10 minutes, or even every second. They may also record values at a faster
sampling rate during specific events, like when rms voltage drops below a threshold.
To save storage space, some sensors or their associated data collection gateways or servers
use lossless or lossy compression algorithms. Lossy compression may include settings that
determine the allowable loss of precision in historical measurements.
Sensors can store either instantaneous values in a time series, such as the rms voltage for
one cycle at a specific time, or statistical summaries at regular intervals, like the minimum,
average, and maximum voltage every ten minutes. Values are typically recorded and stored
by phase (e.g., Phase AN, Phase BN, and Phase CN), but some meters might average line-
neutral voltage/currents or line-line voltages to save on memory or file space. Time series
can display values for different voltage phases, as exemplified by Figure 6-4 which shows

71
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

RMS voltage for Phase AN, Phase BN, and Phase CN.

Figure 6-5: Example time series chart for RMS Voltage on Phase AN, Phase BN,
and Phase CN

6.1.1.2 Event Categorization

Waveforms, RMS values, and phasor charts can be categorized using many IEEE standards
or custom codes.
For example, IEEE Std 1782 [109] provides categories and subcategories for the cause of
an interruption (e.g., vegetation, lightning, wildlife, weather, equipment, etc.) IEEE Std
1159 [110] provides broad categories related to power quality (e.g., momentary
interruption or instantaneous voltage sags). IEEE Std C37.114 [111] provides descriptions
of fault types as single-phase-to-ground, phase-to-phase, phase-to-phase-to-ground, and
three-phase.
A subject matter expert may review the waveforms, RMS values, and phasor charts and
categorize them due to their cause (e.g., cable failure or load tap changer failure) or their
source (transmission system distribution system, customer system, neighboring electric
system, etc.)
Cause information may also be extracted automatically from an electric utilities outage
management system (OMS) or distribution management system (DMS).
6.1.1.3 Data Sources
Waveforms, rms values, or phasor values can be recorded by a number of different sensors:
72
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

• Microprocessor Relays
• Digital Fault Recorders
• Remote Terminal Units (RTUs)
• Power Meters
• Energy Meters
• Power Quality Monitors
• Continuous Waveform Recorders
• Sequence of Event Recorders
• Lightning Detection Systems
• Manually Supplied Information to a Utility Automation System (e.g., the cause of
an interruption as reported by a customer)
• Weather sensors
Not all aforementioned sensors record all data types.
6.1.1.4 Data Formats
Data from these sensors can be stored in file formats

• IEEE Std C37.111 COMTRADE Files: Waveform Samples, RMS Samples,

Phasor Values, Digital Status Values, Time Series
• IEEE C37.118.1 Synchrophasor: Phasor Magnitude and Phase Angle, RMS
Values
• IEEE Std 1159.3 PQDIF Files: Waveform Samples, RMS Samples, Phasor
Values, Digital Status Values, Time Series
• Text Files: Custom CSV Files, Custom Sequence of Event Logs
6.1.1.5 Data Management and Storage
Waveform samples, digital status values, rms value, and other values can be stored in a
folders and subfolders. For example, a simple system may employ folders and subfolders
of IEEE COMTRADE Files or IEEE PQDIF Files.
Some data users employ commercially available or proprietary databases or data
warehouses to store the measurements in a relational database or time series historian.
When storing data in a cloud-based platform, additional storage options include data lakes.

6.1.2 Data Structures

6.1.2.1 Introduction
Data structuring and organizing is critical in any AI/ML based application. The
fundamental concept of data structures is about how data is managed to perform tasks like
organizing, storing, accessing, and editing. In modeling problems, the management of data

73
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

includes the relations or connections between data most often developed using complex
structured database.
Basic data structures are single variables, vectors and arrays. The focus of this kind of
structures is to access the data using position indices, normally the numbers of rows or row
and columns in an ordered sequence.

Figure 6-6: Example of a data structure

With the development of Object-Oriented Programming and modern programming

languages, data types were expanded to consider a combination of several types of data.
For example, it enables the capability of storing a string and numeric value in one variable
structure like an array that uses a key to access data or even to include properties of the
data (dictionaries). Other concepts like sets or containers were introduced by the evolution
of programming and with this expanding the concept of programming using close to real
world structures. Currently depending on the programming language used, there are many
different data structures that can be considered when developing an artificial intelligence
application. In this section key factors that needs to be considered while using data
structures for application of AI/ML are discussed.
6.1.2.2 Defining Data Structures
According to the algorithm or techniques to be implemented, data structure can be defined
based on the attributes and specific data to be used. For example, an algorithm that uses
population principles needs data about the size of population and solution fitness
(objective). An approach that can be used to define a data structure is to perform a manual
derivation of the algorithm such as checking the required variables, and how the data can
be packed in structures for running the algorithm in a high-level way. Relating this to AI
application, assume that a genetic algorithm for an optimization application of an AI
problem needs to be implemented. It is necessary to define the structure for storing the
solution considering the use of an identifier (id) for every solution and include among the
data attributes the value of the fitness function (objective).

74
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Special attention should be paid to the performance requirements when data structures are
used. By principle, the more complex are the data structures, the worse the performance
becomes. Therefore, the size of data structures inversely proportional to the performance
of the algorithm. The performance can be optimized considering the design principle of
minimum computation cost by processing only data structure parts that are fundamental
for algorithms execution. It is highly recommended to test the performance of data
structures by the execution of a draft algorithm implementation before totally defining the
data structure to be used in an algorithm.
6.1.2.3 Data Management and Relationships
The first aspect to be considered is the level of generalization of the problem to be solved
and input capabilities. When the level of generalization (capacity to process several kinds
of inputs – problems) is higher and data structures are bigger and complex, it is very
important to analyze limits of data to avoid algorithm collapsing due to large quantity of
data. Data segmentation and control the size of data structure are needed based on the
application in hand to avoid crashing during the execution. Second, determining correct
volume of data is critical. For this studying the capability of programming language to
define the data structure is needed, followed by an evaluation of the size of the data in the
structure and its impact on the algorithm’s performance. Special attention is required when
dynamic data structures are used. Control of size and segmentation may be needed to have
acceptable algorithm performance. Third, the data relationships will be studied well based
on the application when complex algorithms are designed. Understanding the data
relationship provides access to the right information, helps avoid passing all data to a
process with significant loss of execution time, and it allows for more efficient
programming to scale the models in the future. A key point that needs to be considered is
the usage of identifiers in a simple and consistent way to avoid data duplication and
confusion. A best practice relating to data structure relation definition is to include data
attributes in data structures to reduce the computational burden. It is also important to think
about how the program or algorithms will be maintained, how easy will be to scale and
how feasible is the algorithm to obtain the desired performance and user friendliness.

6.1.3 Computational Platforms

Computational platforms in power system protection and control are formed by hardware,
software, and communication systems for data collection, processing, and dissemination.
The selection of a specific computation platform is highly dependent on the type of AI/ML
applications to be implemented, and how/where an AI/ML application is going to be
implemented. These may be realized in various ways. One practical example of such a
computational platform would be Centralized Protection and Control (CPC) systems within
a substation [104]. CPC systems are modular and collect data from all the devices within a
substation. Due to the modularity of the hardware, CPC platforms can be expanded as
needed to provide computing power for AI/ML-based applications. CPC systems are great
computation platforms for AI/ML-based applications as they have access to substation-
wide real-time data for learning and inference. If employing expanded communications
schemes, CPC systems may also obtain additional data from neighboring substations, edge
devices in feeders, or from the cloud, giving rise to a hybrid edge/centralized system.
75
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Advanced high-impedance fault detection and location is an example of a challenging issue

that could be addressed with an AI/ML-based algorithm running on a CPC system. High-
impedance fault detection algorithms are difficult to implement in traditional asset-specific
IEDs due to the computational power requirements and the extensive requirement of on-
line training. If on-line training is moved to the cloud, with data sharing from and to field
devices (such as reclosers, fault indicators, or feeder V/I sensor monitors), then less
computationally intense algorithms can run on the edge. Still, though fault detection is
possible on IEDs, fault location remains difficult with the limited environmental
information available at the edge. Security and dependability can both greatly be improved
in a centralized system. Using a CPC, information and fault detection from all the
algorithms on the edge can be combined. In addition to improved data, more
computationally intense algorithms can be employed in the centralized system to verify the
results received from the edge. The location and relative separation of the edge devices
help give the CPC a clearer picture with which to resolve the problem of fault detection
and location.
6.1.3.1 Hardware
6.1.3.1.1 Hardware Options
To effectively run the machine learning models once they are built in a production
environment, where the models will reside is important to consider. This decision will be
different for different utilities, depending on the existing architecture of the backend data
repositories/software, cybersecurity practices, communication infrastructure, and general
preferences. There are three different options for storing the operational machine learning
models: hardware on the edge, centralized servers, and the cloud.
6.1.3.1.1.1 Edge Hardware
There are many solutions available for edge processing. However, edge processing
sometimes could be limited in computational power depending on the device chosen. Some
vendors offer all in one solution where they provide sensing in addition to computational
power. Such hardware may reside in the field or in a substation.
6.1.3.1.1.2 Centralized Processing
Centralized processing is more powerful computer located inside a facility. It enables to
collect data from various sensors across the system and build/run machine models on a
larger scale.
6.1.3.1.1.3 Cloud Computing
Cloud computing is an efficient approach to collect data from all sensors, process it, and
use it to build ML models without the need for hardware. There are many vendors that
offer cloud computing/data storage services safely.
6.1.3.1.2 Hardware Considerations
There are many considerations that will help identify the hardware needed and the
architecture (Edge vs Central vs Hybrid) to choose. Some may include:
- Location of the sensors/devices collecting the data relative to each other. If they
are all located in the same area, one might choose edge computing

76
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

- Amount of data collected from sensor. The larger the data, the more
consideration one would make for some level of edge processing
- Number of data sources that model relies on. The more data sources contribute
to the model, the greater the need for either a centralized computing resource or
cloud computing.
- Cybersecurity

6.1.3.2 Software
6.1.3.2.1 Software Options
The software space for machine learning can be very confusing. There are new companies
being created every year that claim to provide the best software to develop and run machine
learning models. It is therefore imperative that utilities be very clear on their objectives for
what a software platform should be provided. Their understanding of the end goal will
enable them to identify the level of UX/UI interface that is needed depending on who the
end user is.
6.1.3.2.2 Software Considerations
There are many considerations that will help identify the software needed to deploy the
models and view the results
- Software needs to comply with cybersecurity/IT standards and requirements
- Software needs to provide adequate level of visibility and flexibility that is
appropriate for the end user
- Software is scalable to accommodate the exponential amount of data without
sacrificing speed, reliability, etc.
- Software adapts to the machine learning models language, libraries
- Software is able to provide strong visualization that is specific to utility
applications and not too generic

6.1.4 Testing and Validation

6.1.4.1 Learning. Validation and Testing in Machine Learning (ML) Architecture
There are two different ways the ML architecture uses the data, a) offline or batch machine
learning b) online machine learning.

In the batch machine learning process, the data is first gathered and divided into training
and validation sets. Then the feature set has been extracted and further the features from
the set have been used for training including the data labeling. This is used as the input to
ML/ANN. The outcome is the learning model that has been developed based on the
training. This model is then used for validation and testing. For validation purposes, part
of the data from training is used which is then evaluated using a different set of data called
testing.

Once the training and testing of the batch learning are complete then the model can be used
in production (for the problem that needs to be solved such as prediction).

77
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

In the online training, ML algorithm/ANN is connected to the streamed data from a system
through a feature set, feature extraction and data labeling. With the online training, the
learning model continue to learn and update with the help of the ML algorithm so that the
new scenarios that have not been trained can be learned in this process.

Similarly, the prediction can be performed with the previously acquired data namely offline
prediction or using the streamed data namely online prediction.

Figure 6-7: Coordinated learning framework based on offline/online machine

learning

Figure 6-7 illustrates the process of combined offline and online training. The offline
training develops a learning model that can predict or infer the possible outcomes. The
model can be updated using an update algorithm evolved from the offline trained machine
learning algorithm. In the process of independent online learning without offline
counterpart, the learning starts from an initial condition and learning model updates are
performed from the initial state.
6.1.4.2 Testing
Testing in the context of an AI/ML system pertains to the process of ascertaining the
performance of such system. There are two possible aspects to this:
1. Ensuring that the AI/ML system exhibits satisfactory with respect to the error rate
requirements. The required error rates can be imposed on each type of errors
individually (e.g., false positives, false negatives), or combined.
2. Ensuring that the computational environment in which the AI/ML system is
deployed has proper resources to accomplish sufficient data acquisition in the
field, on-line inference, or on-line training.

78
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

In this section the process of offline learning is discussed in detail. Preparing the system to
perform classification or estimation in the field is referred to as training, while the
classification or estimation in the field is referred to as inference.
6.1.4.3 Parameter and Hyperparameters
Without any loss of generality, the following discussion assumes an AI/ML system that
needs to perform a task of classification. In this context it is further assumed that the
training of the system is performed in a supervised manner: data that describes observations
in the field is available for training along with the correct labels for each observation. In
general, every AI/ML algorithm has a set of direct parameters that are set through training,
using acquired data. Once the architecture for the algorithm is selected, these direct
parameters are produced through optimization, forcing the algorithm to try and conform to
the labels for each of the data points.
Take an ANN as an example. Each such network can have several layers in the network,
and several neurons per layer. Neurons between layers can be fully connected, or scarcely
connected. Each neuron can have various activation functions. There are other aspects to
take into considerations as well. The totality of all such aspects necessary to create an
AI/ML algorithm is covered within the term architecture. It is easy to show that multitude
of aspects can be defined for other ML algorithms as well. All the aspects within the
architecture are referred to as hyperparameters.
It should be clear at this point that the training mentioned above, through optimization,
results in optimal parameters for the model (given the architecture), while selection of
hyperparameters is not obtained in the same manner. This means that once the architecture
is selected, the available data can be used along with the optimization method to compute
the optimal parameters, which in turn sets the performance of the system with respect to
the error rates etc. as mentioned earlier. Thus, the only way to improve performance further,
if the performance is not satisfactory, is to change the architecture, or hyperparameters.
6.1.4.4 Estimating Performance
Before looking into how to change the architecture, it is imperative to look into how the
performance in the field can be estimated. Every AI/ML system is designed and trained
using finite amount of available data. However, once the system is deployed, it is most
likely to encounter data samples that had not been used in training. Therefore, it is
imperative to assess the inference performance on the samples not previously used.
A common way to approach this problem is to split the available data into training and
testing. In this manner the system is trained on a subset of data (training set), and its
performance is tested on the complement subset (testing set), which accomplishes testing
on the samples not previously seen.
Assume there are N samples available in total. Of these N samples, M samples are used for
training, and N-M samples are used for testing. This selection is referred to as a test/train
split. At this point we can use the M samples to train the model (optimize the parameter of
the system) and assess the performance on the N-M testing samples. We can then randomly
79
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

choose a different subset of M testing samples, retrain the network on the N-M samples,
and reassess the inference performance. This selection of splits can be performed multiple
(K) times, and the performance can be averaged over all (K) splits. Once the assessment is
performed, the system can be trained with all N samples before deployment, and the
estimate of the performance should hold in the field. Figure 6-8 illustrates this process of
selecting train/test splits.

Test (N-M) Train (M) Train (M)

Test (N-M)
Train (M)
Test (N-M)

Test (N-M)
Train (M)
Test (N-M) Train (M)

Figure 6-8: Train/test splits

Now consider a situation where the system was deployed and was forced to infer K samples
from the field. As mentioned before, the samples encountered in the field are most likely
samples that had not been used in training. However, the system was trained on N samples
and asked to infer on K samples, as opposed to the earlier situation where the training was
done on N-K samples and inference was done on K samples. It should be clear that the
performance in the field would be better than the estimate in the paragraph above, because
the system was trained on more data before deployment.
Note that in the training process, the testing samples were never used to change the
architecture of the model, but only to assess the inference performance on “previously
unseen” samples by the ML algorithm.
6.1.4.5 Cross-validation Loop
The development described thus far assumed that the architecture of the algorithm is not
changed throughout the training process. However, changing the hyperparameters could
improve the performance. Staying with the ANN example, maybe performance of the
algorithm could be realized if the number of layers is increased, number of neurons per
layer is changed, or activation functions per layer are changed. Searching for the optimal
hyperparameters is not trivial and cannot be accomplished by using the same approach as
optimizing the parameters of the algorithm. The process for optimizing hyperparameters is

80
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

most frequently performed by a grid search over possible combinations of hyperparameters

and is illustrated in the Figure 6-9.

Hyperparameter
Testing combination/selection
subset

K Splits

Training
subset

Figure 6-9: Cross-validation process

As Figure 6-9 shows, for each combination of hyperparameters, K splits are generated,
much like in the process outlined in Figure 6-8, except the splits are now generated from
the training subset (not the entire dataset), and the testing subset is left untouched during
this entire process. The K splits are now referred to as test/validation or cross-validation
splits, and the process is called cross-validation. For each combination of
hyperparameters, performance is assessed by averaging the performance for each of the K
splits. Optimal set of hyperparameters is the one that gives the optimal performance over
the K cross-validation splits.
Once the hyperparameters are optimized, the final performance is assessed by training the
model using the optimal architecture (hyperparameters) and the entire training set, and
testing using the previously left-out testing set.
It is important to note that the testing set was not used in training at all, and it never should
be. If the performance during the cross-validation is not satisfactory, the AI/ML model
needs to be modified until the performance is adequate, without testing on the testing set.
The reason is that once testing set is used to assess performance, that will be the end of the
training process. If changes to the model are made as a result of observing the performance
using the testing set, then the testing set is technically being used in the training, and that
violates the paradigm of keeping the testing set separate (not previously seen by the mode).
In this case the estimate of the performance becomes overly optimistic, as opposed to
pessimistic.
Some AI/ML designers may still end up taking this (erroneous) step but mingling testing
data in training is highly discouraged.
81
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

6.1.4.6 Field Implementation and Testing

Field implementation computing platforms can have multiple layers. The stored data from
pre-acquired dataset is generally located in the cloud which enables offline
training/learning. The cloud computing infrastructure necessitates all the different layers
of offline training as discussed in Section 6.1.3. The prediction or inference engine can
reside in the central location making the central computing cluster a decision-making
entity. The inference or prediction is then delivered to field devices via a communication
backbone (wired or wireless depending on the application and severity). Online learning
can be performed on the edge server and decision control at the edge level can be
augmented based on the streamed data and online learning.

Field deployment is generally performed after the offline training and thus online training
is further facilitated based on real-world data streaming. With online learning, the
architecture continues to learn such that the new information is used to extend the learning
Euclidean space. For offline learning, the pre-acquired dataset is used. The overall
framework is illustrated in Figure 6-10.

As with any protection and control scheme field implementations, the implemented AI/ML
computing infrastructure also needs to undergo similar field-testing processes before
putting into operations. This may include a period of running the scheme in a monitoring
mode, and performing field testing using standard test equipment and procedure, etc.

Figure 6-10: Generic Field Implementation Computing Infrastructure

82
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

6.1.5 User Training

As with any power system protection and control products, end users need to be adequately
trained to use the AI/ML based protection and control products properly.
The following is a short list of topics related to training. Training modules could be self-
paced or instructor led, hands-on interactive with the user.
Manufacturers: Prepare video-based (cloud-based) online training and/or in-person
training modules from basic to very advanced levels would be prepared for both end users
and supporting personnel. Training modules may include: AI/ML basics,
installation/configuration/verification of the product, problem analysis / self-diagnostics
and precision / pinpoint problem reporting, troubleshooting guidance, life cycle
maintenance hands-on training, upgrade procedures and steps, acceptance testing, etc. Also
offer both users group and individual self-paced training opportunities.
End-users: Identify line of businesses (LOB) to be trained (e.g., P&C, Engineering,
Maintenance, IT, Security, Operation) is the first task. Trainees are to meet certain skillset
requirements of the training modules. Types of training modules need to be selected based
on LOB of the users to be trained which always include common modules such as security,
life cycle maintenance process, etc., as well as how AI/ML P&C applications can supply
reports needed for regulatory compliance (operational, maintenance, security, performance
measures). Users could also learn using P&C database to track for firmware / multiple
firmware versions for the P&C AI/ML based application products, and to determine
metrics and methodologies to compare AI/ML applications with existing / conventional
product or system performance.

7 Use of AI/ML for Protection: Risks, Challenges and Acceptance

Criteria
7.1 Risks
Although there is clear evidence that the AI/ML technology could be successfully applied
to solve practical power system protection and control problems, the AI/ML technology
should never be considered as some type of “magic” technology that can be applied to
solve all types of protection and control problems. It will be applied with the same type of
care as many other technologies that have been applied for power system protection and
controls. Otherwise, the application of the technology could inadvertently introduce new
risks to reliable power system protection and controls by:

• Applying the technology to solve the problem that is not suitable for it
• Failing to recognize the major challenges in applying the technology
• Traditional protective relay testing methods using physics-based models involve
thousands of test cases. AI and ML based protection algorithms needs to be vetted
through many more test cases to make sure the algorithms perform well for all
anticipated fault and non-fault conditions.
83
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

• Taking shortcuts instead of going through rigorous training and testing and
validation processes
In the last few decades, software has become part of various components and systems, and
contributes directly to the success of their different functions and missions. It is an essential
asset for current operation of IEDs, and with wide-scale digitalization it will become the
critical central point of the future real-time analysis, prediction, prescription and decision-
making of the power plant and substation. That is why, over the time, the size of software
systems is continuously increasing, and they become more and more complex. So far, the
reliability of software has rarely been considered in utility reliability studies or in the asset
management. Based on the criticality incumbent to software, its risk is now requested to
be integrated into the future studies.
There is one major difference between software and hardware engineering. The
unreliability in software results in design faults which are mainly caused by human
(intellectual) failures. On the other hand, hardware unreliability is based on random
component failures. Here is a partial list of different characteristics between software and
hardware 1:
Wear-out Software does not have energy related wear-out phase.
Errors can occur without warning
Repairable system concept Periodic restarts can help fix software problems
Time dependency and life cycle Software reliability may not be a function of operational time
Environmental factors Do not affect software reliability, except it might affect
program inputs
Reliability prediction Software reliability can not be predicted from any physical
basis, since it depends completely on human factors in
design
Redundancy Can not improve software reliability if identical software
components are used
Interfaces Software interfaces are purely conceptual other than visual
Failure rate motivators Usually not predictable from analyses of separate statements
Built with standard components Well-understood and extensively-tested standard parts will
help improve maintainability and reliability. But in software
industry, we have not observed this trend. Code reuse has
been around for some time but to a very limited extent.
Strictly speaking there are no standard parts for software,
except some standardized logic structures

Reliability theories have been developed over the years and used successfully for hardware
systems. Previous work has been focused on extending the classical reliability theories

1
“Software Reliability”, Jiantao Pan. http://users.ece.cmu.edu/~koopman/des_s99/sw_reliability/
84
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

from hardware to software in such a way that we can establish software reliability
framework consistently from the same viewpoints as hardware. But some major drawbacks
were obvious :

• While hardware failures may occur independently (or approximately so), software
failures do not happen independently.
• The interdependency of software failures is also very hard to describe in detail or
to model precisely.
• Similar hardware systems are developed from similar specifications, and hardware
failures, usually caused by hardware defects, are repeatable and predictable.
Software systems are typically one-of-the-kind.
Thus, even though it is very difficult to apply at first the usual tools such as FMEA (Failure
Mode and Effect Analysis) and FMECA (Failure Mode and Effect Criticality Analysis) for
software reliability, it is on the other hand crucial that we put a lot of effort right away to
include the software reliability analysis in those studies since the software is omnipresent
in most of the hardware used in critical utility components and systems.

7.2 Challenges
A crucial aspect for the AI/ML based methods to succeed is the availability of reliable field
data. It is very hard to find large sets of field data, especially for protection, since these
data need to faithfully capture all disturbances and faults, so both desired dependability
and security of protection could be achieved .
For example, for fault location application, the field data need to capture all kinds of faults,
with different fault types, fault locations, fault inception angles and fault resistance; for
protection applications, these data need to faithfully capture not only extensive internal
faults but also all disturbances (external faults or other system transients). Take
transmission lines with voltage level higher than 220kV with the State Grid Corporation of
China (SCGG) as examples. Through year 2020, there were around 2000 transmission line
faults, and the overall length of transmission lines is around 620,000 kilometers [105]. That
is to say, on average there is only 1 fault per year for a transmission line with the length of
310 km. One can observe that the field data during faults are extremely sparse both spatially
and temporally. There are several possible ways to use the very limited amount of field
data, as shown in Figure 7-1. Details are elaborated below.

85
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

Very limited amount of field data for fault-related applications

Field data for training

Simulation
data
Field data of the for training
Field data of the
interested component and
interested component only
other components

High credibility of the data (data directly come from the field); Large training set;
Much improved coverage
Pros: Patterns within the data directly Slightly larger training set
of fault conditions and
correspond to the interested line system conditions

Training set may be too small to Training set may still be Low credibility of the
have reasonable results; relatively small; data (patterns of the
Patterns within the data may
Cons: simulation data may be
not correspond to the
component of interest far away from those of
Limited coverage of various fault inception angles, locations, types, the field data)
impedances; various system operating conditions, topologies, etc.

Figure 7-1: Pros and Cons of AI/ML based methods for fault-related applications

First, if the field data of only the interested component are adopted for training, since the
training set may be too small with insufficient coverage of fault conditions (such as fault
types, locations, inception angles, and resistances), it is unlikely to have acceptable results.
In addition, learning may be required every time there are changes in topology or other
operating conditions of the system, which may take a protective relay out of service and
leave the network unprotected or cause considerable delays of the fault
location/classification procedure.
Second, besides the field data of the interested component, if the field data of other
components are also adopted to enlarge the training set, the patterns within the data may
not correspond to the component of interest only. In this case, portability of data generated
in one component/network to train for other components/networks may be questionable.
Scalability of such solutions have been shown to be poor in published literature. In
addition, if an AI/ML based method uses data of multiple resolutions to create “richer”
patterns, such data would come from different types of sensors, each sensor potentially
having unique errors. This also brings challenges to the performances of AI/ML based
methods.
Finally, to solve the issues of insufficient coverage and small training set, large training
sets can be generated using the simulation software. Note that the generated dataset at least
needs to be “complete” in a sense that it has considered all aspects of fault conditions and
system conditions; otherwise, overfitting issues may occur [106]. Even so, the credibility
86
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

of the generated training data is still questionable: how to validate that the patterns of the
training data generated by the simulation software are consistent with those of the field
data? It is observed that faults often tend to have different signatures than those gleaned
from system-analytics or simulations. In fact, physics-based relays may perform better in
such cases, and even if they fail, they fail “predictably”; AI/ML based methods tend to
perform poorly when encountered with signatures not seen while training. It is also easier
to understand failures of physics-based methods than data-driven, learning based methods,
making the physics-based methods easier to correct. Nevertheless, the physics-based
methods may also suffer from modeling inaccuracy since the calibration of system models
in practical power systems are also challenging [107]. The bottom line is, before applying
an AI/ML based method into practical power systems, one would need to evaluate the
performance of the method using the field data, regardless of whether the AI/ML based
method is trained using simulation or field data.
Interpretation of reliability of AI/ML based methods also needs to be considered rationally.
Even with 99.9% success rate of an AI/ML based scheme on a protected network, dozens
of daily misoperations or nonoperations would result if the technology is widely deployed
across a major grid (thousands of networks). This goes to show that currently deployed
protection schemes, though they experience failures from time to time, have an extremely
high-performance baseline.
There are, however, problems in power system protection that do not yield to clear physics-
based modeling. Such problems would benefit from AI/ML based solutions if data are
available. AI/ML community has already recognized the issues with opaque models and
limited availability of field data. This has prompted research into physics-aware AI/ML,
which is expected to capture the best of both approaches in creating better solutions than
those that exist today.
Sometimes, staged-fault testing is done to collect field data to verify newer fault detection
techniques for various surfaces [108]. However, staged-fault testing is expensive and time
consuming and can impact system security due to failure of equipment/system during such
testing. Many utilities are moving away from such testing.

7.3 Criteria for Accepting AI/ML Applications in Field

In general, it is necessary to do a study/proof of concept before accepting an AI/ML
application in the actual utility power system. The study typically contains:

• The impact of a failure of the device, system or any element that is being replaced
or improved.
• Cost of implementation, operation and benefit gain
• Schedule from proof of concept to implementation date
• Feedback/review plan for adjustment of deployed system (maintenance cost)
• Personnel training plan to use the system.
• List of departments impacted and how they need to adapt to the new system
• Change management, plan for regularly review/audit documentation and results of
the AI/ML decision.
87
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

General recommendations to address the definition of a process for accepting AI/ML

applications in the field are as follows.
Main criteria points are planning, effective communication, progressive implementation,
process changing, main risk scenarios testing, redundancy in case of failure.

7.3.1 Planning
Implementing new technologies requires a good strategy because of two aspects, i.e.,
human nature has a resistance for changes and changing technology has risks, that need to
be properly analyzed and controlled. Stay high level during this stage and try not to get
down to the detail.
In the planning stage there need to be:
1. A problem to be solved; otherwise, it will be a very difficult path to move forward
in the AI/ML.
2. A benefit in savings of time and money.
3. An explanation of how AI/ML will solve this problem
4. A strategy on how to implement these successfully.
There needs to be a commitment from the decision maker that this is worth investing. Back
to the planning stage, the planning document that is convincing and wholesome is critical.

7.3.2 Effective Communication

Keeping people informed about the AI/ML implementation process, steps, benefits, testing
stages and risks and their managements is fundamental. Regular communication needs to
take place to provide updates on implementation progress. Users could establish an open
and direct communication to developers when finding and resolving issues. Users could
also communicate lessons learned during each stage.

7.3.3 Progressive Implementation

When starting with the operation of a new AI/ML solution, device, system, start with no
critical points, elements or substation, do not compromise the power system security.
Aim small and be nimble. Start small, review the results, update/improve and repeat.
Slowly increase the implementation plan or area. This method provides a place to fail safely
and allow the team to come back up and try again really quick. The key is to be honest
when things failed, to change and get going again.
A continuous or even periodic comparison against a non-AI/ML solution needs to happen
to update the learning and training process.

7.3.4 Process Changing

Taking the right time to change process and to update team training is important to avoid
misunderstandings when using a new AI/ML solution. Provide the right documentation to
88
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

understand the solution. People can use AI/ML better when they have a better
understanding about how the black box works.
Make sure management is included in the new process change and provide reports on how
the new AI/ML solution is improving and meeting the goals/intended purpose.
Provide forum to discuss changes and open communication to developer/maintainer of the
system if there are any problem. This has to be treated as continuous improvement.
Business process may change because of this AI/ML implementation and not once but a
few times throughout the life of the AI/ML system.

7.3.5 Main Risk Scenarios Testing

Main risk scenarios testing of the AI/ML solution needs to be carried out with minimum
two critical scenarios using real data. Depending on the complexity of applications, many
scenarios may need to be considered during testing phase. Testing stage is fundamental for
any solution or idea for changing conventional methods.
Future regulatory (for example, NERC PRC standards) compliance testing requirements of
AI/ML systems need to be taken into account.

7.3.6 Redundancy in Case of Failure

Having a plan B when applying a new AI/ML solution in critical points of the power system
where security could be compromised. Redundancy is desired. The following are examples
of redundancy that can be included:

• Computer systems
• Network
• Sensors
• Backups/Restoration

8 CONCLUSIONS AND RECOMMENDATIONS

This report demonstrates that the practical application of limited AI/ML technology for
power system protection and control has already started. The two practical use-case
examples discussed in this report are clear evidence that the technology could be applied
successfully to solve practical power system protection and control problems. The other
examples highlighted also illustrate that there are many more areas that AI/ML technology
has a great potential to help solve some of the power system protection and control
challenges.
It is also very clear that practical and widespread application and deployment of this
technology is still at a very early stage. Despite a large number of research and development
efforts that have been reported to date, many of such efforts are still a long way from being
applied mainstream. The development, implementation, testing, and user acceptance of
89
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

AI/ML technology-based applications still have some way to go before achieving

deployment maturity and wide-spread acceptance by utilities and practicing engineers.
It is the hope of the authors that this report will serve to help accelerate the adoption of the
AI/ML technology for practical power system protection and control applications. Given
the large number of on-going R&D efforts, the body of the knowledge in this area is
expected to grow and expand rapidly. Being an emergent area, it is not possible to attempt
addressing details of all aspects in this report within a limited timeframe, but care has been
taken so that this report will still be relevant in a few years after the body of the knowledge
has expanded based on more widespread deployments globally.
Therefore, this report may become a living document that needs updating on a regular basis
(e.g., every 3-4 years) to help incorporate new knowledges/insights and the latest progress
in the practical application of the AI/ML technology in power system protection and
control.

90
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

9 APPENDIX A – BIBLIOGRAPHY
Books
Jennie Si; Andrew G. Barto; Warren Buckler Powell; Don Wunsch, Handbook of Learning
and Approximate Dynamic Programming , IEEE, 2004,
Sutton, R. & Barto, A. Reinforcement Learning: An Introduction (MIT Press, 1998)
Sutton, R.S. Learning to predict by the methods of temporal differences. Mach Learn 3, 9–
44 (1988). https://doi.org/10.1007/BF00115009

Papers
[1] A. Mosavi, M. Salimi, S. Ardabili, T. Rabczuk, S. Shamshirband and A.
Varkonyi-Koczy, "State of the Art of Machine Learning Models in Energy
Systems, a Systematic Review," Energies, vol. 12, 2019.

[2] L. Wu, G. Kaiser, C. Rudin, D. Waltz, R. Anderson, A. Boulanger, H. Dutta and

M. Pooleery, "Evaluating Machine Learning for Improving Power Grid
Reliability," in ICML Workshop on Machine Learning for Global Challenges,
2011.

[3] C. B. Jones, A. Summers and M. J. Reno, "Machine Learning Embedded in

Distribution Network Relays to Classify and Locate Faults," in IEEE Innovative
Smart Grid Technologies (ISGT), 2021.

[4] S. Hossain-McKenzie, E. C. Piesciorovsky, M. J. Reno and J. C. Hambrick,

"Microgrid Fault Location: Challenges and Solutions," Sandia National
Laboratories, SAND2018-6745, 2018.

[5] A. C. Adewole, R. Tzoneva and S. Behardien, "Distribution Network Fault

Section Identification and Fault Location Using Wavelet Entropy and Neural
Networks," Applied Soft Computing, vol. 46, pp. 296-306, 2016.

[6] A. Prasad, J. B. Edward and K. Ravi, "A Review of Fault Classification

Methodologies in Power Transmission Systems: Part-I," Journal of Electrical
Systems and Informaton Technology, vol. 5, no. 1, pp. 48-60, 2018.

[7] Y. Wang, M. Lui and Z. Bao, "Deep Learning Neural Network for Power System
Fault Diagnosis," in Proceedings of the 35th Chinese Control Conference, 2016.

91
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

[8] S. K. Saha, P. Das and A. K. Chakrabothy, "An ANN based Relay Design for
Identification Faults of 400kv High Voltage AC Transmission Line,"
International Journal of Computer Applications, vol. 46, no. 20, 2012.

[9] V. Malathi, N. S. Marimuthu, S. Baskar and K. Ramar, "Application of extreme

learning machine for series compensated transmission line protection,"
Engineering Applications of Artificial Intelligence, vol. 24, no. 5, pp. 880-887,
2011.

[10] F. Wilches-Bernal, A. Bidram, M. J. Reno, J. Hernandez-Alvidrez, P. Barba, B.

Reimer, R. Montoya, C. Carr and O. Lavrova, "A Survey of Traveling Wave
Protection Schemes in Electric Power Systems," IEEE Access, 2021.

[11] R. Montoya, B. Poudel, A. Bidram and M. J. Reno, "DC Microgrid Fault

Detection Using Multiresolution Analysis of Traveling Waves," International
Journal of Electric Power & Energy Systems, 2022.

[12] M. Jimenez-Aparicio, M. J. Reno, P. Barba and A. Bidram, "Multi-Resolution

Analysis Algorithm for Fast Fault Classification and Location in Distribution
Systems," in IEEE International Conference on Smart Energy Grid Engineering,
2021.

[13] M. J. Aparicio, S. Grijalva and M. J. Reno, "Fast Fault Location Method for a
Distribution System with High Penetration of PV," in Hawaii International
Conference on System Sciences (HICSS), 2021.

[14] S. S. Venkata, M. J. Reno, W. Bower, S. Manson, J. Reilly and G. W. S. Jr.,

"Microgrid Protection: Advancing the State of the Art," Sandia National
Laboratories, SAND2019-3167, 2019.

[15] Z. Chen, L. Wu, S. Cheng, P. Lin, Y. Wu and W. Lin, "Intelligent fault diagnosis
of photovoltaic arrays based on optimized kernel extreme learning machine and I-
V characteristics," Applied Energy, vol. 204, pp. 912-931, 2017.

[16] M. Awaad, S. Mekhamer and A. Y. Abdelaziz, "Design of an adaptive

overcurrent protection scheme for microgrids," International journal of
engineering science and technology, vol. 10, no. 1, pp. 1-12, 2018.

[17] Q. Cui and S. Li, "A Microgrid Protection Scheme with Conventional Relay
Measurements," in 2018 IEEE Power & Energy Society General Meeting
(PESGM), 2018.

92
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

[18] A. Iqbal and . Pooja, "Intrusion Detection in Smart Grid Using Machine Learning
Approach," Journal of Computational and Theoretical Nanoscience, vol. 16, no.
9, pp. 3808-3816, 2019.

[19] H. Lin, J. M. Guerrero, C. Jia, Z.-h. Tan, J. C. Vasquez and C. Liu, "Adaptive
overcurrent protection for microgrids in extensive distribution systems," in
IECON 2016 - 42nd Annual Conference of the IEEE Industrial Electronics
Society, 2016.

[20] M. Manohar, E. Koley and S. Ghosh, "Microgrid protection under wind speed
intermittency using extreme learning machine," Computers & Electrical
Engineering, vol. 72, pp. 369-382, 2018.

[21] M. Mishra and P. K. Rout, "A comprehensive micro-grid fault protection scheme
based on S-transform and machine learning techniques," International Journal of
Advanced Mechatronic Systems, vol. 7, no. 5, p. 274, 2017.

[22] N. E. Naily, S. M. Saad, J. Wafi, A. Elhaffar and N. Husseinzadch, "Adaptive

Overcurrent Protection to Mitigate High Penetration of Distributed Generation in
Weak Distribution Systems," in 2017 9th IEEE-GCC Conference and Exhibition
(GCCCE), 2017.

[23] W. J. Tang and H.-T. Yang, "Data Mining and Neural Networks Based Self-
Adaptive Protection Strategies for Distribution Systems with DGs and FCLs,"
Energies, vol. 11, no. 2, p. 426, 2018.

[24] M. Mishra and P. K. Rout, "Detection and classification of micro-grid faults based
on HHT and machine learning techniques," IET Generation, Transmission, &
Distribution, vol. 12, no. 2, pp. 388-397, 2018.

[25] E. Casagrande, W. L. Woon, H. H. Zeineldin and D. Svetinovic, "A Differential

Sequence Component Protection Scheme for Microgrids With Inverter-Based
Distributed Generators," IEEE Transactions on Smart Grid, vol. 5, no. 1, pp. 29-
37, 2014.

[26] S. Afrasiabi, M. Afrasiabi, B. Parang and M. Mohammadi, "Integration of

Accelerated Deep Neural Network into Power Transformer Differential
Protection," IEEE Transactions on Industrial Informatics, pp. 1-1, 2019.

[27] S. Das and B. K. Panigrahi, "Online intelligent technique for preventing relay
maloperation under stressed conditions," in 2017 19th International Conference
on Intelligent System Application to Power Systems (ISAP), 2017.

93
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

[28] R. Dubey, S. Samantaray and B. Panigrahi, "An extreme learning machine based
fast and accurate adaptive distance relaying scheme," International Journal of
Electrical Power & Energy Systems, vol. 73, pp. 1002-1014, 2015.

[29] Y. Feng, B. Duan, C. Tan and Z. Yao, "A settings tracking and providing scheme
for differential protection based on machine learning," in 2015 IEEE Innovative
Smart Grid Technologies - Asia (ISGT ASIA), 2015.

[30] M. Tasdighi and M. Kezunovic, "Preventing transmission distance relays

maloperation under unintended bulk DG tripping using SVM-based approach,"
Electric Power Systems Research, vol. 142, pp. 258-267, 2017.

[31] S. Zhang, Y. Wang, M. Liu and Z. Bao, "Data-Based Line Trip Fault Prediction in
Power Systems Using LSTM Networks and SVM," IEEE Access, vol. 6, pp.
7675-7686, 2018.

[32] R. C. Matthews, T. R. Patel, A. Summers, M. J. Reno and S. Hossain-McKenzie,

"Per-Phase and 3-Phase Optimal Coordination of Directional Overcurrent Relays
Using Genetic Algorithm," Energies, 2021.

[33] T. Patel, S. Brahma, J. Hernandez-Alvidrez and M. J. Reno, "Adaptive Protection

Scheme for a Real-World Microgrid with 100% Inverter-Based Resources," in
IEEE Kansas Power and Energy Conference (KPEC), 2020.

[34] H. Lin, K. Sun, Z. Tan, C. Liu, J. M. Guerrero and J. C. Vasquez, "Adaptive

protection combined with machine learning for microgrids," IET Generation,
Transmission, & Distribution, vol. 13, no. 6, pp. 770-779, 2019.

[35] Y. Zhang, M. D. Ilíc and O. K. Tonguz, "Mitigating Blackouts via Smart Relays:
A Machine Learning Approach," in Mitigating Blackouts via Smart Relays: A
Machine Learning Approach, 2011.

[36] E. Bernabeu, J. Thorp and V. Centeno, "Methodology for a

Security/Dependability Adaptive Protection Scheme Based on Data Mining,"
IEEE Transactions on Power Delivery, vol. 27, no. 1, 2012.

[37] G. N. Korres and P. J. Katsikas, "Identification of circuit breaker statuses in WLS

state estimator," IEEE Transactions on Power Systems, vol. 17, no. 3, pp. 818-
825, 2002.

[38] C. Francis, R. D. Trevizan, M. J. Reno and V. Rao, "Topology Identification of

Power Distribution Systems Using Time Series of Voltage Measurements," in
IEEE Power and Energy Conference at Illinois (PECI), 2021.

94
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

[39] B. Poudel, D. R. Garcia, A. Bidram, M. J. Reno and A. Summers, "Circuit

Topology Estimation in an Adaptive Protection System," in IEEE North American
Power Symposium (NAPS), 2021.

[40] A. Rajendra-Kurup, M. Martinez-Ramon, A. Summers, A. Bidram and M. J.

Reno, "Deep learning based circuit topology estimation and fault classification in
distribution systems," in 2021, IEEE PES ISGT Europe.

[41] V. Kekatos and G. B. Giannakis, "Joint power system state estimation and breaker
status identification," in 2012 North American Power Symposium (NAPS),
Champaign, 2012.

[42] R. Dobbe, W. Westering, S. Liu, D. Arnold, D. Callaway and C. Tomlinar,

"Forecasting-based state estimation for three-phase distribution systems with
limited sensing," in arXiv preprint arXiv:1806.06024, 2019.

[43] J. Xie, A. P. S. Meliopoulos and B. Xie, "Transmission line fault classification

based on dynamic state estimation and support vector machine," in 2018 North
American Power Symposium (NAPS), Fargo, 2018.

[44] A. S. Zamzam, X. Fu and N. D. Sidiropoulos, "Data-driven learning-based

optimization for distribution system state estimation," IEEE Transactions on
Power Systems, To be published.

[45] O. Simeone, "A very brief introduction to machine learning with applications to
communication systems," IEEE Transactions on Cognitive Communications and
Networking, vol. 4, no. 4, pp. 648-664, 2018.

[46] T. Skripcak and P. Tanuska, "Utilisation of on-line machine learning for SCADA
system alarms forecasting," in 2013 Science and Information Conference,
London, 2013.

[47] M. Teixeira, T. Salman, M. Zolanvari, R. Jain, N. Meskin and M. Samaka,

"SCADA system testbed for cybersecurity research using machine learning
approach," Future Internet, vol. 10, no. 76, 2018.

[48] L. A. Maglaras and J. Jiang, "A novel intrusion detection method based on
OCSVM and K-means recursive clustering," EAI Endorsed Transactions on
Security and Safety, vol. 2, no. 3, pp. 1-10, 2015.

[49] M. Zhou, Y. Wang, A. K. Srivastava, Y. Wu and P. Banerjee, "Ensemble-based

algorithm for synchrophasor data anomaly detection," IEEE Transactions on
Smart Grid, vol. 10, no. 3, pp. 2979-2988, 2019.

95
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

[50] A. Ahmed, V. V. G. Krishnan, S. A. Foroutan, M. Touhiduzzaman, A. Srivastava,

Y. Wu, A. Hahn and S. Sindhu, "Cyber physical security analytics for anomalies
in transmission protection systems," in 2018 IEEE Industry Applications Society
Annual Meeting (IAS), Portland, 2018.

[51] W. Chen, S. Hsu and H. Shen, "Application of SVM and ANN for intrusion
detection," Computers & Operations Research, vol. 32, pp. 2617-2634, 2005.

[52] S. Pan, T. Morris and U. Adhikari, "A specification-based intrusion detection

framework for cyber-physical environment in electric power system,"
International Journal of Network Security, vol. 17, no. 2, pp. 174-188, 2015.

[53] U. Adhikari, T. H. Morris and S. Pan, "A cyber-physical power system test bed
for intrusion detection systems," in 2014 IEEE PES General Meeting |
Conference & Exposition, 2014.

[54] V. Gurevich, Cyber and Electromagnetic Threats in Modern Relay Protection,

2014.

[55] T. Matsumoto, T. Kobayashi, S. Katayama, K. Fukushima and K. Sekiguchi,

"Protection relay systems employing unconditionally secure authentication
codes," in 2009 IEEE Bucharest PowerTech, 2009.

[56] S. Rahman, H. R. Pota and J. Hossain, "Cyber vulnerabilities on agent-based

smart grid protection system," in 2014 IEEE PES General Meeting | Conference
& Exposition, 2014.

[57] S. Hossain-McKenzie, M. J. Reno, R. Bent and A. Chavez, "Cybersecurity of

Networked Microgrids: Challenges, Potential Solutions, and Future Directions,"
Sandia National Laboratories, SAND2020-13723, 2020.

[58] S. Hossain-McKenzie, M. J. Reno, J. Quiroz, P. Schulz, A. Summers and C.

Carter, "Cyber Security Issues and Solutions for Protective Relaying and Local
Monitoring," Sandia National Laboratories, SAND2018-0406, 2018.

[59] S. Ward, J. O'Brien, B. Beresh, G. Benmouyal, D. Holstein, J. T. Tengdin, K.

Fodero, M. Simon, M. Carden, M. Yalla, T. Tibbals, V. Skendzic, S. Mix, R.
Young, T. Sidhu, S. Klein, J. Weiss, A. Apostolov, D. Bui, S. Sciacca, C. Preuss,
S. Hodder and G. Seifert, "Cyber Security Issues for Protective Relays; C1
Working Group Members of Power System Relaying Committee," 2007 IEEE
Power Engineering Society General Meeting, Tampa, 2007.

[60] R. C. B. Hink, J. Beaver, M. Buckner, T. Morris, U. Adhikari and S. Pan,

"Machine Learning for Power System Disturbance and cyber-attack

96
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

discrimination," in 2014 International Symposium on Resilient Control Systems

(ISRCS), 2014.

[61] C.-W. Ten, J. Hong and C.-C. Liu, "Anomaly Detection for Cybersecurity of the
Substations," IEEE Transactions on Smart Grid, vol. 2, no. 4, pp. 865-873, 2011.

[62] Y. Chen and B. Lou, "S2a: Secure smart household appliances," in 2nd ACM
Conference Data Applicaiton Security Privacy, San Antonio, 2012.

[63] R. Mitchell and I.-R. Chen, "Behavior-Rule Based Intrusion Detection Systems
for Safety Critial Smart Grid Applications," IEEE Transactions on Smart Grid,
vol. 2, no. 4, pp. 865-873, 2011.

[64] Y. Yang, McLaughlin, S. Sezer, T. Littler, B. Pranggono, P. Brogan and H. F.

Wang, "Inrusion Detection System for Network Security in Synchrophasor
Systems," in IET Interational Conference, 2013.

[65] H. Hadeli, R. Schierholz, M. Braendle and C. Tuduce, "Leveraging determinism

in inudstrial control systems for advanced anomaly detection and relaible security
configuration," Emerging Technologies and Factory Automation, pp. 22-25, 2009.

[66] J. Hong, R. F. Nuqui, A. Kondabathini, D. Ishchenko and A. Martin, "Cyber

Attack Resilient Distance Protection and Circuit Breaker Control for Digital
Substations," IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, vol. 15,
no. 7, 2019.

[67] J. Mulder, M. Schwartz, M. Berg, J. R. V. Houten, J. M. Urrea, M. A. King, A. A.

Clements and J. Jacob, "WeaselBoard: Zero-Day Exploit Detection for
Programmable Logic Controllers," Sandia National Laboratories, Albuquerque,
2013.

[68] "PFP Cybersecurity," 2019. [Online]. Available: https://www.pfpcyber.com/.

[69] Y. Huang, J. Tang, Y. Cheng, H. Li, K. A. Campbell and Z. Han, "Realtime

detection of false data injection in smart grid networks: An adaptive cusum
method and analysis," IEEE Syst. J., vol. 10, no. 2, pp. 532-543, 2016.

[70] A. P. S. Meliopoulos, G. J. Cokkinides, P. Myrda, Y. Liu, R. Fan, L. Sun, R.

Huang and Z. Tan, "Dynamic state estimation-based protection: Status and
promise," IEEE Transactions on Power Delivery, vol. 32, no. 1, pp. 320-330,
2017.

[71] H. Lin, K. Sun, Z.-H. Tan, C. Liu, J. M. Guerrero and J. C. Vasquez, "Adaptive
Protection Combined with Machine Learning for Microgrids," IET Generation
Transmission & Distribution, 2019.

97
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

[72] Turing, A. "Computing Machinery and Intelligence" Mind, October 1950, LIX
(236): 433–460

[73] Silva I, Spatti D. H., Flauzino R., et al, Artificial Neural Networks, a Practical
Course, Springer 2017.

[74] Yang, Xin-She, et al, Artificial Intelligence, Evolutionary Computing and

Metaheuristics, in the Footsteps of Alan Turing, Springer, 2013.

[75] Cao L. Metasynthetic Computing and Engineering of Complex Systems,

Advanced Information and Knowledge Processing, Springer, 2015.

[76] Jones, M. T, Artificial Intelligence A Systems Approach, Computer Science

Series, Jones & Bartlett Learning, 2009.

[77] Kumar A., Great Mind Maps for Learning Machine Learning, Data Analytics,
Data Science, Machine Learning AI, Dec 28, 2020. Accessed on: May 20, 2021.
[Online]. Available: https://vitalflux.com/great-mind-maps-for-learning-machine-
learning/

[78] Grosan C, Abraham A., Intelligent Systems, A Modern Approach, Springer, Vol
17, 2011

[79] Y. Bengio, I. Goodfellow and A. Courville, Deep Learning, London: MIT Press,
2016.

[80] Y. Bengio, A. Courville and P. Vincent, "Representation Learning: A Review and

New Perspectives," 2014.

[81] J. Schmidhuber, "Deep Learning in Neural Networks: An Overview," Lugano,

2014.

[82] H. -. Jacobsen, "A generic architecture for hybrid intelligent systems," 1998 IEEE
International Conference on Fuzzy Systems Proceedings. IEEE World Congress
on Computational Intelligence (Cat. No.98CH36228), 1998, pp. 709-714 vol.1,
doi: 10.1109/FUZZY.1998.687575.

[83] Jeffrey Wischkaemper, Sukumar Brahma, “Machine Learning and Power System
Protection [Viewpoint]”, IEEE Electrification Magazine, March 2021, pp. 108-
112.

[84] Jörg Blumschein, Yilmaz Yelgin, Andrea Ludwig, “Adaptive Autoreclosure to

Increase System Stability and Reduce Stress to Circuit Breakers”, presented at the
70th Annual Conference for Protective Relay Engineers, April 3-6, 2017

98
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

[85] Piñeros J. “Intelligent Model to Determine and Verify Automatically Distance

Protection Settings”, Research work for M.Sc. Degree, Universidad de Antioquia,
Medellín Colombia, 2016.

[86] Peter B. SnowAlexander P. ApostolovJefferson D. Bronfeld, US Patent

5,537,327, “Method and apparatus for detecting high-impedance faults in
electrical power system”, July 16, 1996

[87] Peter B. SnowAlexander P. ApostolovJefferson D. Bronfeld, US Patent 5,734,575

“Method and apparatus for detecting high-impedance faults in electrical power
systems”, March 31, 1998

[88] Nikos Hatziargyriou et al., “Definition and Classification of Power System

Stability – Revisited & Extended”, IEEE Transactions on Power Systems, Vol.
36, No. 4, July 2021.

[89] K.. Dharmapala, A. Rajapakse, K. Narendra and Y. Zhang, "Machine Learning

Based Real-Time Monitoring of Long-Term Voltage Stability Using Voltage
Stability Indices," in IEEE Access, vol. 8, pp. 222544-222555, 2020.

[90] Power System Relaying Committee Report, “Transmission relay system

performance comparison,” 2005. [Online]. Available: http://www.pes-
psrc.org/Reports/I17ReportRev3.zip

[91] D. C. E. de la Garza, Masters Thesis - Hidden Failures in Protection Systems and

its Impact on Power System Wide-area Disturbances. Virginia Polytechnic
Institute and State University, 2000.

[92] O. P. Dahal, H. Cao, S. Brahma, and R. Kavasseri, “Evaluating performance of

classifiers for supervisory protection using disturbance data from phasor
measurement units,” in 2014 IEEE PES Innovative Smart Grid Technologies
Conference Europe (ISGT-Europe), Istanbul, Oct 2014.

[93] S. Brahma, R. Kavasseri, Huiping Cao, N. R. Chaudhuri, T. Alexopoulos, and Y.

Cui, “Real Time Identification of Dynamic Events in Power Systems using PMU
data, and Potential Applications – Models, Promises, and Challenges”, IEEE
Trans. Power Delivery – Special Issue on Innovative Research Concepts for
Power Delivery Engineering, Vol. 32-1, pp. 294 – 301, Feb. 2017.

[94] O. Dahal, S. Brahma, and H. Cao, “Comprehensive clustering of disturbance

events recorded by phasor measurement units,” IEEE Trans. Power Del., vol. 29,
no. 3, pp. 1390–1397, Jun. 2014.

99
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

[95] M. Biswal, Y. Hao, P. Chen, S. Brahma, H. Cao, and P. DeLeon, “Signal features
for classification of power system disturbances using PMU data,” Proc. Power
Syst. Comput. Conf., Jun. 2016, pp. 1–7

[96] M. Brown, M. Biswal, S. Brahma, S. Ranade, and H. Cao, “Characterizing and

quantifying noise in PMU data,” in Proc. IEEE Power Energy Soc. Gen. Meeting,
Jul. 2016, pp. 1–6.

[97] Y. Yin, S. Alaeddini and Y. Fu, "Automatic Fault Analysis and Visualization of
Digital Substation Event," 2020 IEEE Power & Energy Society General Meeting
(PESGM), 2020, pp. 1-5, doi: 10.1109/PESGM41954.2020.9281882.

[98] M. J. Aparicio, M. Reno, P. Barba, and A. Bidram, “Multi-resolution analysis

algorithm for fast fault classification and location in distribution systems,”
International Conference on Smart Energy Grid Engineering (SEGE), 2021.

[99] R. Montoya, B. Poudel, A. Bidram, and M. J. Reno, “DC microgrid fault

detection using multiresolution analysis of traveling waves,” International Journal
of Electrical Power & Energy Systems, vol. 135, 107590, Feb. 2022.

[100] R. Yousefian, R. Bhattarai and S. Kamalasadan, "Transient Stability Enhancement

of Power Grid With Integrated Wide Area Control of Wind Farms and
Synchronous Generators," in IEEE Transactions on Power Systems, vol. 32, no. 6,
pp. 4818-4831, Nov. 2017.

[101] R. Yousefian, A. Sahami and S. Kamalasadan, "Hybrid energy function based

real-time optimal wide-area transient stability controller for power system
stability," 2015 IEEE Industry Applications Society Annual Meeting, 2015.

[102] F. Gomez, A.D. Rajapakse, U.D. Annakkage and I.T Fernando, "Support Vector
Machine-Based Algorithm for Post-Fault Transient Stability Status Prediction
Using Synchronized Measurements," in IEEE Transactions on Power Systems,
vol. 26, no. 3, pp. 1474-1493, August 2011.

[103] A. D. Rajapakse, F. Gomez, K. Nanayakkara, P. A. Crossley and V. V. Terzija,

"Rotor Angle Instability Prediction Using Post-Disturbance Voltage Trajectories,"
in IEEE Transactions on Power Systems, vol. 25, no. 2, pp. 947-956, May 2010.

[104] Working Group on Centralized Substation Protection and Control, IEEE Power
System Relaying Committee, “Advancements in Centralized Protection and
Control within a Substation”, in Proc. IEEE Trans. Power Delivery,
doi:10.1109/TPWRD.2016.2528958, Vol. 31, No. 4, pp. 1945-1952, August
2016.

100
PES-TR112— Practical Applications of Artificial Intelligence and Machine Learning in Power System
Protection and Control

[105] Annual National Reliability Report of China, National Energy Administration,

2020.

[106] Y. Xing, Y. Liu, and X. Zheng, “A Systematic Framework Presenting Impact of

Dataset Completeness on Data Driven Approaches: A Transmission Line Fault
Classification Study”, IEEE Conference on Energy Internet and Energy System
Integration (EI2), 2020.

[107] Y. Liu, et. al, “Dynamic State Estimation for Power System Control and
Protection”, in IEEE Trans. Power Syst., vol. 36, no. 6, pp. 5909-5921, Nov.
2021, doi: 10.1109/TPWRS.2021.3079395.

[108] M. Carpenter, et. al, “Staged-Fault Testing for High Impedance Fault Data
Collection”, 58th Annual Conference for Protective Relay Engineers, Texas
A&M University, College Station, TX, April 5-7, 2005.

[109] IEEE Std 1782TM-2014, “IEEE Guide for Collecting, Categorizing, and Utilizing
Information Related to Electric Power Distribution Interruption Events”

[110] IEEE Std 1159TM-2019, “IEEE Recommended Practice for Monitoring Electric
Power Quality”

[111] IEEE Std C37.114TM-2014, “IEEE Guide for Determining Fault Location on AC
Transmission and Distribution Lines”
[112] Scikit-learn: Machine Learning in Python, Pedregosa et al., Journal of Machine
Learning Research 12, pp. 2825-2830, 2011. Available online at:
https://scikit-learn.org/stable/tutorial/machine_learning_map/index.html

[113] R. Das and S. A. Kunsman, “A Novel Approach for Ground Fault Detection”,
57th Annual Conference for Protective Relay Engineers, Texas A&M
University, College Station, TX, March 30-April 1, 2004.

[114] M. R. Lyu, "Software Reliability Engineering: A Roadmap," Future of Software

Engineering (FOSE '07), 2007, pp. 153-170, doi: 10.1109/FOSE.2007.24.

101