Beyond Traditional Hydrological Models: Transformer‐Based Deep Learning for Future Streamflow and Flood Risk Assessment in the Brahmaputra River Basin

ABSTRACT This study presents an integrated framework combining process‐based hydrological modelling with advanced deep learning techniques to improve climate‐driven streamflow prediction and flood risk assessment in the Brahmaputra River Basin (BRB) at Bahadurabad. Unlike conventional comparative studies, this work explicitly evaluates the added value of Transformer‐based architectures against traditional hydrological modelling (HEC‐HMS) and recurrent neural networks (LSTM and BiLSTM) under bias‐corrected CMIP6 climate projections. Historical discharge data from 1981 to 2014 were used for calibration, validation, training and testing. Among the evaluated models, the Transformer demonstrated superior predictive capability, achieving an R 2 of 0.94 during testing with PBIAS within ±1%, whereas HEC‐HMS showed comparatively lower accuracy ( R 2 = 0.72) and consistent underestimation exceeding 20%. Future projections for the 2030s, 2050s and 2080s reveal a systematic shift towards earlier monsoon onset and a 10%–25% increase in seasonal discharge, with mean annual flow rising from approximately 22,000 m 3 /s in the baseline period to nearly 30,000 m 3 /s by the late century. By coupling data‐driven predictions with flood frequency analysis using the Gumbel distribution, this study further demonstrates a pronounced intensification of extreme flood events under future climate scenarios. The findings highlight the enhanced capability of Transformer‐based models to capture complex hydroclimatic dynamics and underscore their potential as a robust tool for climate‐resilient water resources planning in large transboundary river basins.