Application of seasonal-adjusted hybrid models for forecasting Discomfort Index in a heat-prone region of Bangladesh

Abstract

Extreme heat and humidity pose an increasing threat to human health, labor productivity, and overall well-being, particularly in heat-vulnerable and rapidly urbanizing regions such as Rajshahi, Bangladesh. As rising temperatures and elevated humidity levels intensify exposure to heat stress, accurate forecasting has become essential for effective early warning systems and climate-resilient urban planning. However, modeling thermal discomfort is challenging due to the need to analyze long-term, high-frequency meteorological time series data with complex seasonal and nonlinear structures. Therefore, this study applies and evaluates seasonal-adjusted, machine-learning-based hybrid models to forecast thermal discomfort using 40 years (1985–2024) of daily temperature and humidity data. The Thom’s Discomfort Index (DI), a standard measure of thermal stress, was calculated and decomposed using the STL (Seasonal-Trend decomposition based on LOESS) method to separate trend, seasonality, and residual components. A total of 128 hybrid model combinations were implemented by integrating traditional time series models (ARIMA, TBATS, ETS, and GARCH) with machine learning techniques (ANN, Prophet, SVR, Decision Trees, Random Forests, XGBoost, LSTM, and GRU). Among all models, the STL-TBATS-LSTM hybrid achieved the best performance, with MAE = 0.4810, MAPE = 2.1230, RMSE = 0.6381, and MASE = 0.6644, followed closely by STL-TBATS-DTR. Historical analysis from 1985 to 2024 revealed strong seasonal peaks in discomfort from June to August, along with a clear long-term increase in both the frequency and intensity of high-discomfort days. Forecasts for 2025–2027 project a substantial rise in thermal stress, with approximately 39.8% of days falling under “High Discomfort” and 1.2% under “Severe Discomfort.” These findings highlight the escalating burden of heat stress in Bangladesh and underscore the urgency of STL-based hybrid forecasting models in supporting climate adaptation strategies and enhancing public health preparedness.