Machine Learning-Based Fault Detection And Classification In Power Systems

Abstract

The modern power systems are getting more complicated, closely coupled and confronted to various different 
faults and disturbances for which timely, fast, accurate and adaptive actions take placed for the protection. The 
existing protection schemes based on conventional relay, although widely used over the years, have inherent 
limitations regarding adaptability, fault classification speed, and coordination in smart grid environments. This 
review paper provides an extensive search and meta-analysis of literature from the previous decade of automated 
power system protection schemes designed employing machine learning (ML) and deep learning (DL) methods. 
This survey provides a systematic review of the published literature over the past 30 years covering fault detection, 
classification, location, and isolation methods using various types of algorithms such as Artificial Neural 
Networks (ANN), Support Vector Machines (SVM), Decision trees (DT), Random Forests (RF), k-Nearest 
Neighbours (k-NN), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and 
hybrid intelligent techniques. This meta-analysis assesses these approaches based on their performance on key 
metrics such as classification accuracy, response time, robustness to noise, scalability, and their use with modern 
grid topologies, e.g., those in which renewable energy sources are part of the grid, or microgrid systems. Abstract 
Key results show that deep learning models, here portrayed by Convolutional Neural Network (CNN)- LSTM 
hybrids, outperforms traditional ML strategies in many complex, high-dimensional fault scenarios, in many 
situations attaining classification accuracies higher than 99% in controlled experimental environments. 
Nevertheless, issues remain with respect to data imbalance, interpretability, cyber-security, and computational 
cost in their actual deployment. We conclude the paper by highlighting open research gaps and avenues towards 
fully autonomous self-healing power system protection frameworks.