A Machine Learning–Based Approach for Fault Detection in Power Systems

Authors: Aakanksha Upadhyay

Country: India

Full-text Research PDF File:   View   |   Download

Abstract: Detection of faults is a fundamental need to reliable and safe operation of modern electric power systems, where protective functions are required to properly discriminate between internal faults and non-fault disturbances within very strict time limits. Traditional protection techniques are still useful in most environments, but their operation may be challenged by varying operating conditions, enhanced penetration of power-electronics interfaced resources, measurement uncertainty and the increasing variety of transient events that are similar to faults. This paper is inspired by these issues and creates a machine learning (ML)-based system to detect faults quickly and classify them into different types using synchronized electrical signals. The proposed methodology is based on the existing principles of protection and the recent developments in the field of data-driven learning, where fault analysis is developed as a supervised learning problem with short sliding windows of multi-channel voltage and current measurements. It introduces a deep neural architecture that trains on minimally processed measurements and thus eliminates the need for manually engineered features, including low detection latency, robustness to noise, fault resistance variation and operating-point variation. The methodology is placed in comparison with classical wavelet- and feature-engineering techniques, contemporary deep-learning techniques of waveform-based fault analysis, and PMU-based wide-area fault analytics. The paper also defines an evaluation plan that focuses on realistic generalization by partitioning the dataset into scenarios, multi-metric measures of dependability and security, and interpretability of the results based on explainable AI methods. The resulting paper offers a strict and repeatable roadmap of ML-based fault detection applicable both to the transmission and distribution settings, and it includes clear instructions on how data should be generated, how the model should be trained, and how it should be validated to be used in practice.

Keywords: fault detection; power system protection; machine learning; deep learning; synchrophasors; convolutional neural networks; explainable AI.

Paper Id: 232875

Published On: 2025-07-11

Published In: Volume 13, Issue 4, July-August 2025

Share this