A Study on Advancing Runoff Prediction Through Peak Protection Selective Learning in U.S. CAMELS Basins

Accurate prediction of extreme runoff events is crucial for flood control and disaster mitigation and water resource risk management. In recent years, deep learning has emerged as a key method for runoff simulation. However, its training strategy applies uniform optimization across all time steps, which encourages the model to fit normal hydrological processes while reducing its sensitivity to extreme runoff events. To address this issue, this study introduces Selective Learning and proposes a runoff prediction model with a flood peak protection mechanism, namely PPSL-DI-LSTM. In this framework, sample importance is dynamically regulated through uncertainty filtering, anomaly filtering, and flood peak protection, thereby reducing the influence of unstable or non-generalizable samples while enhancing the learning of high-flow processes. Based on this strategy, a dynamic weighted loss function, termed Selective Weighted RMSE (SW-RMSE), is designed to optimize the training process. Experimental results on the CAMELS dataset show that, compared to the baseline DI-LSTM model, the proposed model achieves improvements of approximately 4% and 6% in median Nash-Sutcliffe Efficiency (NSE) and median Kring-Gupta Efficiency (KGE), respectively. Simultaneously, the Flow Duration Curve High Flow Variability (FHV) was significantly reduced by approximately 84% and the overall percent bias (PBIAS) improved from −3.80% to 1.15%. Additionally, when integrated into a standard LSTM, PPSL maintains comparable median NSE while improving median KGE (from 0.751 to 0.778) and substantially reducing FHV (from −12.86% to −3.34%). Our method overcomes the limitations of uniformly treating all training samples in complex hydrological simulations, and has strong potential to substantially improve the accuracy and reliability of future runoff modeling studies.