NEURAL NETWORK-BASED EARLY LIGHTNING STRIKE PREDICTION FOR DISTRIBUTION INFRASTRUCTURE RESILIENCE ENHANCEMENT

Abstract

Lightning-induced outages are a persistent threat to the resilience and reliability of power distribution networks, especially during convective weather seasons. This study presents a neural network-based early lightning detection framework aimed at enhancing distribution network resilience by integrating advanced deep learning and power system simulation. Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) architectures were optimized through hyperparameter tuning (grid search, Bayesian optimization) and evaluated using precision, recall and F1-scores. Results demonstrated the LSTM model’s superior performance, achieving 91% accuracy and 83% recall for lightning events, outperforming the CNN model (81% accuracy, 63% recall). While both models identified wet-season peaks (April–October) and afternoon maxima (15:00–20:00), the CNN exhibited anomalous dry-season predictions, highlighting its limitations in capturing long-term climatological trends. The LSTM’s robust handling of temporal dependencies yielded a realistic 365-day forecast for 2024, aligning with historical seasonality. The optimized LSTM model’s predictions are subsequently validated and fed into a MATLAB-based distribution network simulation environment. Where, predicted lightning events triggered adaptive operational strategies, such as dynamic relay setting adjustments and preemptive sectionalizing, informed by predicted lightning occurrence data. Simulation results demonstrated that integrating early lightning prediction reduces unnecessary breaker operations and power outages, with quantifiable improvements in key reliability indices such as System Average Interruption Frequency Index (SAIFI), System Average Interruption Duration Index (SAIDI) and Customer Average Interruption Duration Index (CAIDI). SAIFI improved from 1.82 to 0.92 (49% improvement), SAIDI improved from 3.5 to 1.15 (65% improvement), CAIDI improved from 1.92 to 1.25 (35% improvement). This also led to faster restoration times compared to conventional reactive approaches to lightning disturbances. This study underscores the value of combining deep learning, historical data, and power system simulation for modern grid resilience, and provides a scalable blueprint for utilities seeking to minimize weather-related disruptions through predictive analytics.