ABSTRACT Accurate river water stage forecasting is essential for flood management in vulnerable deltaic regions, yet optimal lag window selection remains inadequately explored. This study systematically evaluates how historical data window configurations (1, 3, 7, 15, and 30 days) impact next-day forecasting accuracy using six algorithms: traditional machine learning (SVR, RFR, LGBMR) and deep learning (ANN, LSTM, GRU) approaches. Utilizing 26 years (1999–2024) of daily hydrological data (9,497 observations) from Jaganathganj Station, Old Brahmaputra River Basin, Bangladesh, model performance was assessed through nine statistical metrics and principal component analysis for comprehensive evaluation. Results demonstrate that deep learning architectures consistently outperformed traditional methods, with the 30-day GRU configuration achieving optimal performance (MAE: 0.1255 m, R2: 0.9903, MAPE: 1.1545%), representing 13.21% improvement over 1-day baseline. All models exceeded the Nash–Sutcliffe efficiency of 0.97, confirming operational reliability. Despite increased computational demands, GRU maintained real-time inference capability (0.265 ms/sample), enabling deployment in resource-constrained environments. These findings provide evidence-based guidance for implementing operational flood early warning systems, where the 30-day GRU configuration can deliver accurate next-day predictions critical for disaster risk reduction in monsoon-influenced deltaic regions, potentially saving lives and minimizing economic losses through timely evacuation alerts and improved water resource management decisions.
Islam et al. (Tue,) studied this question.