Transfer learning for power system fault location using artificial neural networks

This paper investigates the application of transfer learning techniques to artificial neural networks (ANNs) for fault detection in power distribution systems, formulated as a classification problem. Comprehensive datasets are developed using multiple IEEE test feeders of varying complexity, including the 13-bus, 34-bus, 37-bus, and 123-bus test feeders. Various fault types are simulated across all three-phase buses in each system. Baseline performance is established by independently training ANNs on each feeder. Subsequently, knowledge learned from the complex 123-bus feeder is transferred to accelerate and improve fault location in simpler networks. The results demonstrate that transfer learning significantly improves both training efficiency and classification performance. Training convergence is accelerated by a factor of 1.68 to 2.56 across target feeders, corresponding to epoch reductions between 40.6% and 61.0%. Additionally, computational time is reduced by 24.0% to 49.5%, further enhancing the practical viability of the proposed approach. These findings suggest that transfer learning offers a powerful strategy to address data scarcity and computational challenges in fault location, enabling utilities to deploy accurate, efficient fault detection systems across diverse distribution networks with minimal retraining effort.