A Deep Learning Approach for the Detection of Surge Events in Metal Oxide Surge Arresters

This paper presents a deep learning-based approach using Long Short-Term Memory (LSTM) network to detect surge events in Metal Oxide Surge Arresters (MOSAs) through leakage current analysis. MOSAs protect electrical equipment from overvoltage by diverting surge energy; however, repeated surge events degrade their performance over time, often without clear warning signs of failure. Traditional methods for assessing MOSA health rely heavily on leakage current measurements, which provide an overall indication of arrester condition but do not pinpoint individual surge occurrences within the data. By accurately identifying surge events, the proposed method allows utilities to monitor the cumulative stress on MOSAs, enabling proactive maintenance and reducing the risk of equipment failure. Historical leakage current data from previous studies were preprocessed and used to train the LSTM model, chosen for its effectiveness in analyzing sequential time-series data. The model successfully distinguished normal operating patterns from surge-related anomalies, achieving a detection accuracy of 96.43%. These results highlight the potential of deep learning to revolutionize surge detection, providing a robust tool for predictive maintenance and improved power system reliability.