A coupled SWATPlus and BiLSTM tuning model for improved daily scale hydroclimate simulation in typical loess hilly areas of China

Global climate change and land use alterations are contributing to more frequent and severe extreme events globally. This trend is especially pronounced in the Loess Plateau's hilly regions of China. In order to more accurately simulate runoff, especially peak flows, to enhance the planning and management of water resources in China's loess hilly areas, as well as to mitigate the impacts of global climate change, we propose a novel hydrologic model. The model employs an integrated approach by combining the Soil and Water Assessment Tool (SWATPlus), a process-based conceptual hydrological model, with a deep learning model known as the Bi-Directional Long Short-Term Memory (BiLSTM). Our primary objective is to enhance runoff simulation performance in the Zuli River Basin (ZRB), a key tributary of the upper Yellow River. We designed two coupled models, SWATPlus-BiLSTM-D and SWATPlus-BiLSTM-T. In SWATPlus-BiLSTM-D, all influential parameters of SWATPlus were kept at their default values; while in SWATPlus-BiLSTM-T, we calibrated multiple SWATPlus parameters. By comparing the daily runoff simulation results, we find that SWATPlus-BiLSTM-T consistently outperforms SWATPlus-BiLSTM-D, the stand-alone SWATPlus, and the BiLSTM model throughout the simulation period. Of particular note, SWATPlus-BiLSTM-T significantly outperforms the other three models in the simulation of daily peak flows. In the testing process, the error values of SWATPlus-BiLSTM-T reached relatively excellent levels, with an NSE of 0.88, an R2 of about 0.9, and an RMSE of 2.63 m3/s. The assessment and management of hydrological and environmental conditions in river basins can be significantly optimized through the development of such model tuning methods.