A Hybrid Bi‐ LSTM ‐ AM Framework for Improving Extreme Rainfall Prediction and Bias Correction

ABSTRACT Reliable forecasting of extreme rainfall remains a critical challenge for flood management in climate‐sensitive regions. This study develops and evaluates an attention‐augmented, bidirectional long short‐term memory framework (Bi‐LSTM‐AM) used for gauge‐based prediction and raster‐based bias correction for extreme rain events. The resulting prediction and bias‐correction models are trained by station observations and reanalysis data during 1980–2019. Nine extreme rainfall events in 2022–2024 are extracted as the independent verification set. Among all the bias correction models, the WRF + Bi‐LSTM‐AM combination achieves the highest accuracy across lead times of 1–5 days. The most efficient size of lookback window is found to be 7 days that well balances predictive accuracy and data demands. An event‐clustered paired bootstrap comparison further proves a consistent advantage of Bi‐LSTM‐AM over QM. From the perspective of spatial dependence, the accuracy of the bias‐correction models seems to correlate less with local terrains but more significantly with the rain centres. The future research need of improving the tracking of trajectories and intensities of rain centres is highlighted for extreme rainfall forecasting.