Neural Network-Based Early Lighting Strike Prediction for Distribution Infrastructure Resilience Enhancement

Abstract

Lightning-induced disturbances are a major cause of power outages and equipment failures in medium-voltage distribution networks, particularly across tropical regions such as Abuja, Nigeria. This study presents a neural network–based early lightning prediction framework for Abuja, Nigeria, integrating advanced deep learning techniques with power system simulation to support proactive grid management. Using hourly meteorological data from 2013 to 2023 obtained from the Nigerian Meteorological Agency (NiMet), a seven-phase methodology was employed, including data preprocessing, feature engineering, and exploratory data analysis to address class imbalance, missing values, and temporal dependencies. Key features included lag variables, rolling aggregates, and cyclic temporal encoding to capture diurnal and seasonal patterns. Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) models were optimized through hyperparameter tuning and evaluated using precision, recall, F1-score, and forecast skill metrics. The LSTM achieved 91% accuracy and 83% recall, outperforming the CNN. Predictions were integrated into a MATLAB-based distribution network simulation, where adaptive relay settings and preemptive sectionalizing reduced breaker operations and outage durations. Reliability indices, including SAIDI and SAIFI, improved compared to conventional reactive methods. Findings highlight LSTM-driven lightning forecasting as a scalable solution for enhancing power distribution network resilience through predictive analytics and automated operational strategies.