Enhancing the WAS physical hydrological model and integrating machine learning and residual decomposition to improve runoff simulation performance in the cold region of the Taoerhe River Basin, Northeast China

Abstract

Study region: The Taoerhe River Basin, located in Northeast China. Study focus: Physical hydrological models still have limitations in representing runoff processes in seasonally frozen semi-arid basins. This study improved runoff simulation through physical model modification and hybrid residual correction. First, a temperature-responsive regulation factor was introduced into the water allocation and simulation model (WAS). Subsequently, a multi-level hybrid framework combining a physical model and machine learning (ML) algorithms was developed: Hybrid 1 and Hybrid 2 employed RF to correct WAS residuals; Hybrid 3 and Hybrid 4 integrated Variational Mode Decomposition with multiple ML algorithms (RF, XGBoost and CatBoost) to capture nonlinear signals, and used a stacking strategy to integrate residual predictions. New hydrological insights for the region: The improved WAS model performed well, with NSE above 0.80 and RMSE and MAE reduced by approximately 40%. Hybrid 4 achieved the best overall performance. During testing, R 2 and NSE exceeded 0.90, while RMSE and MAE were 47.99 m 3 /s and 26.50 m 3 /s, respectively. Compared with the improved WAS model, R 2 and NSE increased by more than 0.05, whereas RMSE and MAE decreased by more than 20%. Climate perturbation experiments showed that precipitation mainly determined total runoff, while temperature perturbations mainly altered the partitioning of basin runoff. Thus, small changes in annual runoff may mask seasonal redistribution, offering insight for similar seasonally frozen semi-arid basins.