Station-specific thunderstorm forecasting using climate predictors with seasonal decomposition analysis in...

Thunderstorms pose significant threats to life, agriculture, and infrastructure in Bangladesh, particularly in the northeastern regions of Sylhet, Sreemangal, and Mymensingh. This study develops and rigorously evaluates localized thunderstorm frequency forecasting models for these high-risk stations using 40 years (1985-2024) of monthly data and thirty six modeling approaches encompassing classical time series (SARIMA/SARIMAX, ETS/ETSX), machine learning (ANN/ANNX, SVR/SVRX, XGBoost/XGBoostX), and deep learning (LSTM/LSTMX) architectures, applied to raw and seasonally adjusted (X-11, STL) data under both univariate and exogenous frameworks incorporating five climatic variables (temperature, relative humidity, cloud cover, rainfall, atmospheric pressure). Correlation analysis identified cloud cover (r = 0.83-0.88), atmospheric pressure (r = -0.79 to -0.85), temperature (r = 0.77-0.79), and rainfall (r = 0.74-0.78) as dominant meteorological drivers of thunderstorm activity. Seasonal pre-adjustment using STL substantially improved forecast accuracy, with STL-XGBoost achieving the lowest univariate errors (Sylhet: RMSE = 2.53; Sreemangal: 2.09; Mymensingh: 2.08). Inclusion of exogenous variables dramatically enhanced predictive skill with tree-based and kernel methods: STL-XGBoostX with climate predictors proved optimal for Sylhet (MASE = 0.22), while STL-SVRX with climate predictors was optimal for Sreemangal and Mymensingh (MASE = 0.17), representing over 80% improvement compared to univariate benchmarks. Seasonal forecasts (2025-2028) project peak monsoon activity reaching 100.5 frequency at Sreemangal and identify April-August as the critical preparedness window for these stations. This study provides the Bangladesh Meteorological Department with a validated, operational framework for station-specific thunderstorm frequency forecasting, demonstrating that explicit seasonal decomposition combined with tree-based ensemble and kernel methods offers robust accuracy for early warning systems. The methodology can be extended to additional stations and integrated with socio-economic data to enhance disaster preparedness and climate adaptation strategies for Bangladesh's most thunderstorm-vulnerable communities.