Fault detection and prediction in electrical distribution system using nonlinear dynamics and machine learning models

Electrical distribution systems (EDS) are highly vulnerable to various faults, including equipment failure, wildlife interference, vegetation contact and lightning strikes, due to their wide geographical spread and constantly changing operating conditions. We propose a novel fault detection and prediction framework that uniquely combines Hankel Alternative View of Koopman (HAVOK) analysis with Long Short-Term Memory (LSTM) networks, advancing beyond traditional methods by enabling both real-time detection and proactive prediction capabilities. Our innovative approach achieves superior fault detection accuracy of 99.2% across all fault types, with particularly exceptional performance in identifying high impedance faults (HIZ) at 97.8% accuracy, significantly outperforming conventional methods that typically achieve only 85–90% HIZ detection rates. Using a three-bus power system model, we demonstrate our method’s effectiveness across diverse fault scenarios (SLG, DLG, TLG, LL and HIZ) with an industry-leading average detection time of 15.3 ms, 35% faster than traditional approaches. The integration of HAVOK analysis through Hankel matrix construction and singular value decomposition (SVD) enables robust fault characterization with a remarkably low false alarm rate of 0.42%. Our LSTM-based prediction model achieves 96.5% accuracy in forecasting fault patterns up to 48 h in advance, enabling proactive maintenance strategies that have reduced system downtime by 53% and maintenance costs by 47% in simulations. This comprehensive approach not only advances the state-of-the-art in fault detection and classification but also establishes a new paradigm for predictive maintenance in power distribution systems, significantly enhancing both infrastructure reliability and human safety.