High-Frequency Thunderstorm Prediction via Seasonally Decomposed Hybrid Models: Toward Data-Driven Early Warning and Climate Risk Mitigation

Thunderstorms are frequent and destructive mesoscale phenomena in tropical and subtropical regions, posing major hazards through intense rainfall, lightning, strong winds, hail, and occasional tornadoes. In Bangladesh, premonsoon Kalbaishakhi (Nor’westers) are particularly damaging and remain difficult to forecast at local scales using computationally intensive numerical weather prediction approaches. This study aimed to develop and comparatively evaluate high-frequency data-driven forecasting frameworks for daily thunderstorm occurrence in Mymensingh, a thunderstorm-prone district in north-central Bangladesh and to assess their applicability for operational early warning. Daily thunderstorm frequency observations obtained from the Bangladesh Meteorological Department covering January 1, 1981, to December 31, 2024, were analyzed using a chronological training–testing split (80%–20%). Classical statistical models (ARIMA, ETS, TBATS, and GARCH), machine learning models (SVR, ANN, random forest, and Prophet), deep learning models (LSTMs), and sequential hybrid frameworks were evaluated using MAE, RMSE, MASE, and MAPE. Seasonal-Trend Decomposition using Loess (STL) was applied to address strong seasonality and intermittent zero-inflated behavior prior to hybrid modeling. Exploratory analyses revealed pronounced annual seasonality, minimal winter thunderstorm activity, and peak occurrences during April–June with substantial interannual variability. Among the nonseasonally adjusted models, ETS and ARIMA demonstrated comparatively reliable performance, while hybrid models generally improved forecasting accuracy. STL decomposition substantially enhanced predictive performance, with STL–LSTM achieving the lowest MAPE (2.25%) among individual models, whereas the STL–ARIMA–SVR hybrid model demonstrated the most balanced overall forecasting performance (MAE = 0.0274, RMSE = 0.4830, MASE = 0.2071, MAPE = 3.52%). For operational early warning assessment, STL–ARIMA–SVR achieved a probability of detection (POD) of 0.8435 and a critical success index (CSI) of 0.4157, indicating strong thunderstorm-event-detection capability under intermittent conditions. The findings demonstrate that seasonally decomposed hybrid forecasting frameworks can substantially improve localized thunderstorm prediction and support climate risk mitigation and operational early warning systems in Bangladesh. Future studies should incorporate multistation observations, atmospheric predictor variables, and real-time deployment evaluation to further improve forecasting robustness and operational applicability.