Machine Learning Approaches for Fault Classification in Nigerian 11kV Distribution Feeder Systems

This study develops and validates machine learning classification models for automatic fault type identification in Nigerian 11kV distribution feeder systems, addressing the persistent challenge of slow fault location and restoration that characterises Nigerian distribution network operations. Current fault identification at Nigerian distribution companies relies primarily on field patrol inspection following protective relay tripping, a process that extends outage duration and increases the cost of supply interruptions. A data-driven fault classification approach is developed using transient voltage and current waveforms recorded at feeder departure points during fault events. Fault event data comprising 320 recorded fault events from two EKEDC Lagos feeders and three EEDC Enugu feeders over 36 months were obtained with DISCOM cooperation. Signal processing of the waveforms was performed to extract 18 time-domain and frequency-domain features including wavelet coefficients, RMS current, negative sequence current magnitude, and rate of change of frequency. Four classification algorithms are trained and compared: random forest, support vector machine, k-nearest neighbours, and a convolutional neural network applied directly to the waveform time series. A synthetic fault event generator using PSCAD simulations was used to augment the training dataset with additional fault scenarios not well represented in the field data. Results show that the random forest model achieves a classification accuracy of 94.7 percent for distinguishing single-line-to-ground, line-to-line, double-line-to-ground, and three-phase fault types, outperforming the CNN model on the limited available field data. Recommendations include deployment as a distribution control room decision support tool. Keywords: fault classification, machine learning, distribution feeder, 11kV, power systems.