Empirical Analysis of Deep Learning Models for Automated Fault Diagnosis in Nigerian Power Distribution Networks

Automated fault diagnosis in power distribution networks is critical for reducing outage duration and improving supply reliability, yet the performance of deep learning models for fault classification using real-world Nigerian grid data has not been empirically characterised. This study empirically evaluated convolutional neural network, LSTM, and hybrid CNN-LSTM architectures for multi-class fault detection and classification using current and voltage waveform data from Nigerian distribution feeders. Waveform data were sourced from power quality analysers installed on 12 feeders operated by Abuja DisCo and Enugu DisCo over a 9-month period, capturing 1,847 fault events across five fault categories: single-line-to-ground, double-line, triple-line, transformer inrush, and capacitor bank switching. Feature extraction used wavelet transform coefficients as input to all three architectures. The CNN achieved 92.4 percent overall classification accuracy, the LSTM achieved 89.1 percent, and the hybrid CNN-LSTM achieved 94.7 percent, outperforming both single architectures. Training time on a GPU workstation was 4.2 hours for the hybrid model. Inference latency for a single fault event was 38 milliseconds, meeting real-time protection relay requirements. The hybrid model's superior performance was attributed to its ability to capture both local waveform patterns (CNN) and temporal fault progression (LSTM) simultaneously. The study fills an empirical gap in fault classification performance data using actual Nigerian distribution grid waveforms and recommends deployment of the hybrid model in digital protection relays for the South East and North Central distribution zones.

Keywords: deep learning, fault diagnosis, power distribution, CNN-LSTM, Nigeria electricity