AI-Based Fault Detection and Classification in Distribution Networks Using SVM and Random Forest

Abstract

Fault detection, classification, and localisation in radial distribution networks remain challenging. This is due to non-uniform loading, high resistance (R)/reactance (X) ratios, variable fault resistance, and measurement noise. These issues are especially pronounced in developing power systems such as Nigeria’s. To address these challenges, this study proposes a data-driven fault diagnosis framework. The framework integrates Support Vector Machines (SVMs) for fault-type classification and Random Forest (RF) models for faulted-line identification and fault-distance estimation. The models are benchmarked against a conventional impedance-based method on the IEEE 69-bus radial test system, with clustered loads and heterogeneous line parameters. Electrical characteristics, including phase voltages, phase currents, voltage imbalance, current imbalance, and harmonic distortion indices, were considered for feature extraction and model building. To increase the method’s robustness, data augmentation using Gaussian noise and z-score normalisation was applied, resulting in 4,896 samples. The experiments showed that the SVM classifier achieved a fault classification accuracy of 92.44%. The Random Forest model achieved a fault classification accuracy of 99.32%. The accuracy of fault zone determination reached 99.93%, and the accuracy of exact faulted-line identification reached 100%. For fault distance estimation, the Random Forest regression model achieved an RMSE of 11.78%, significantly outperforming traditional impedance-based methods. This model demonstrates significant robustness under noisy operational conditions. It offers a realistic way of managing faults intelligently in modern-day distribution networks

.