Daily runoff simulation in humid regions using the entropy-weighted ensemble learning models

Abstract

Traditional hydrological models struggle to meet the accuracy requirements for runoff simulation under climate change and anthropogenic interventions. To address this limitation, we propose ensemble learning models (ELMs) that integrate optimal process-driven and data-driven models for daily runoff simulation in two typical humid basins in China: the Xiangjiang River Basin (XJRB) and Minjiang River Basin (MJRB). Model performance is evaluated by a newly developed comprehensive index CI based on entropy weight method. Our results reveal that the Xin'anjiang model outperforms other process-driven models with NSE values of 0.795 (XJRB) and 0.765 (MJRB), while the Long Short-Term Memory model outperforms other data-driven models (NSE: 0.945 and 0.955, respectively). Furthermore, hybrid ELMs surpass all single models, reducing MAE and RMSE by 15 % and 21 % in XJRB, and improving the NSE by 0.157 in MJRB. This framework enhances simulation accuracy and operational robustness, demonstrating strong potential for flood risk mitigation.