Machine Learning-Based Fault Detection and Classification in Smart Power Distribution Networks

Abstract

Smart power distribution networks are becoming increasingly observable through smart meters, distribution-level phasor measurement units, and intelligent electronic devices, creating an opportunity for rapid data-driven fault diagnosis. This methodology paper proposes a machine-learning framework for detecting and classifying distribution-network faults from synchronized electrical measurements. The procedure combines signal-window quality control, physically meaningful feature extraction, standardized model training, and multiclass validation. Five operating classes are considered: normal operation, line-to-ground (LG), line-to-line (LL), double-line-to-ground (LLG), and three-phase (LLL) faults. A reproducible demonstration dataset containing 5,000 labeled events was constructed from class-dependent voltage, current, sequence-component, transient-energy, and distortion features with variations representing load level, distributed generation, fault resistance, and measurement uncertainty. Support vector machine (SVM), random forest, k-nearest neighbors (KNN), and decision tree classifiers were evaluated on a stratified 75:25 split. The SVM achieved the best overall performance with 98.16% accuracy and a macro F1-score of 98.15%, while random forest and KNN also exceeded 97% accuracy. The methodology emphasizes robustness, transparent validation, and compatibility with practical smart-grid sensing. It provides a concise workflow that can be transferred to an IEEE test feeder, real-time digital simulator, or utility measurement archive for subsequent field validation.