Thunderstorm Forecasting by Using Machine Learning Techniques: A Comparative Model Analysis Leveraging Historic Climatic Records of Bangladesh

Abstract

Accurate thunderstorm forecasting is essential for protecting communities and minimising disruptions to agriculture, infrastructure and human lives, particularly in Bangladesh. However, predicting thunderstorms remains challenging due to the complex interplay of meteorological factors, data limitations and regional variations. This study addresses these challenges by integrating historical meteorological data with advanced machine learning and deep learning techniques to improve prediction accuracy. Using data from the Bangladesh Meteorological Department, we compare various models, evaluating their performance based on the coefficient of determination (R²) and root mean squared error (RMSE). Among traditional machine learning models, the Support Vector Machine (SVM) Regressor performed best with an R² of 0.658 and RMSE of 3.65. Among deep learning models, Convolutional Neural Networks (CNNs) achieved superior accuracy with an R² of 0.743 and RMSE of 3.17, effectively capturing spatial patterns in thunderstorm occurrences. Additionally, deep learning models such as CNN, ANN and LSTM successfully detected annual trends and fluctuations, improving prediction reliability. These findings highlight the potential of deep learning in enhancing thunderstorm forecasting, contributing to more effective disaster preparedness and risk management in thunderstorm-prone regions.