Machine Learning-Based Fault Detection and Classification in Smart Electrical Power Systems

Abstract

Reliable protection of smart electrical power systems requires rapid discrimination between normal operating conditions and short-circuit faults despite changing loads, distributed generation, measurement noise, and network reconfiguration. This methodology paper develops a machine-learning framework for fault detection and classification using three-phase voltage and current measurements. A reproducible synthetic benchmark was generated to represent normal operation and ten common fault classes: single-line-to-ground, line-to-line, double-line-to-ground, and three-phase faults. The pipeline applies normalization and physically meaningful features based on phase RMS values, symmetrical-component proxies, unbalance indices, residual current, and voltage-current ratios. Five supervised classifiers—k-nearest neighbours (KNN), support vector machine (SVM), random forest (RF), gradient boosting, and multilayer perceptron (MLP)—were trained on an 80% stratified training set and evaluated on a held-out 20% test set. The best models achieved 99.80% test accuracy and macro F1-score of approximately 0.998, while KNN achieved 98.18%. The results show that compact engineered features can support accurate and computationally efficient fault classification, while also highlighting the need for validation on field measurements and inverter-dominated networks before deployment.