Improving WRF Forecasts of Bay of Bengal Cyclones Using LSTM-Based Time-Series Bias Correction with BMD Observations

Accurate prediction of tropical cyclones in the Bay of Bengal is essential for disaster preparedness in Bangladesh. This study evaluates the Weather Research and Forecasting (WRF) model initialized with free Global Forecast System (GFS) data in simulating cyclone track, intensity, and rainfall for cyclone Remal (2024), validated using freely available observations: wind, pressure, and rainfall from the Bangladesh Meteorological Department (BMD), satellite rainfall from MOSDAC, and cyclone track records from the IBTrACS database. WRF simulations reveal systematic biases, including track deviations, underestimation of maximum wind speeds and central pressure, and errors in rainfall time series. To address these, we employ a Long Short-Term Memory (LSTM) neural network to model and correct temporal biases in track, intensity, and rainfall forecasts. LSTM has previously achieved strong performance in cyclone intensity prediction over the North Indian Ocean (Biswas et al., 2021) and in correcting rainfall biases compared with traditional methods such as quantile mapping (Seo & Ahn, 2023). The network is trained using paired WRF forecasts and observed datasets, with performance assessed through root mean square error, mean absolute error, and temporal/spatial correlation. Results indicate track error reductions of 5–9% and intensity error decreases of 15–30%, consistent with recent ANN-based cyclone forecast improvements (Kim et al., 2023). Rainfall time series accuracy also improves substantially, aligning more closely with observed sequences. This streamlined LSTM-only post-processing framework, based entirely on free-access data, demonstrates a practical and reproducible approach for improving cyclone forecasting over the Bay of Bengal. The method has direct applicability to operational forecasting and early warning systems in Bangladesh.