Machine Learning Techniques for Fault Detection in Electrical Power Networks

Abstract

Reliable fault detection is fundamental to the security and continuity of electrical power networks. Conventional protection schemes are fast and interpretable but may lose sensitivity when operating conditions change because of distributed generation, topology variation, high fault resistance, measurement noise, and non-stationary transients. This methodology paper develops a structured machine-learning framework for detecting and classifying power-network faults from three-phase voltage and current measurements. The framework combines signal segmentation, normalization, physically meaningful features, supervised classifiers, stratified validation, and protection-oriented performance metrics. Support vector machine (SVM), random forest (RF), and multilayer perceptron (MLP) models are compared because they represent margin-based, ensemble, and neural-learning paradigms that can be implemented with modest computational resources. A reproducible physics-informed synthetic benchmark of 5,400 operating cases is used only to demonstrate the evaluation procedure; it contains normal operation and five aggregated fault categories under variable loading, severity, and measurement perturbations. The illustrative test results show classification accuracies of 97.56% for SVM, 98.15% for RF, and 97.56% for MLP, with RF providing the strongest macro F1-score. The methodology emphasizes that accuracy alone is insufficient: relay-oriented deployment must also consider class recall, noise robustness, detection latency, data leakage, and validation on electromagnetic-transient or field measurements. The resulting workflow provides a concise, reproducible basis for future experimental work on intelligent protection of transmission and distribution networks.