A Combined Modeling Framework for Accurate Thunderstorm Forecasting in Low-Risk Areas of Bangladesh.

Abstract

Thunderstorms can devastate buildings, crops and human life even in "low risk" areas like Sitakunda, Bangladesh. This conundrum underscores the need for timely and accurate forecasting to safeguard vulnerable, resource-poor populations. But traditional models often struggle to account for the nonlinear and seasonality of thunderstorms, and little research has been done on predicting the daily occurrence of thunderstorms in low-risk regions. This research seeks to enhance thunderstorm frequency forecasting by extensively evaluating and comparing statistical, machine learning, and deep learning approaches (ETS, ARIMA, STL-ETS, STL-ARIMA, NNAR, TBATS, Prophet, GARCH, SVM, and LSTM). An innovatively combined model is also developed. Applying daily thunderstorm counts from Sitakunda (1981-2023) and forecasting three years ahead (2024–2026), and measuring performance using RMSE and MAE, the findings show the combination model achieves the lowest prediction errors and most reliable forecasts. Although Prophet and SVM show comparable results, GARCH has the lowest RMSE (1.3137). By contrast, LSTM and NNAR exhibit larger errors, suggesting that they struggle to model the unpredictable nature of thunderstorms. These results demonstrate the benefits of combination forecasting in improving thunderstorm predictions for lowrisk areas. Incorporation of these models into early warning systems and local governance can enhance preparedness, minimize disruptions and injuries, and foster resilience to thunderstorm risks in the long term.