Multiple spatio-temporal scale runoff forecasting and driving mechanism exploration by K-means optimized XGBoost and SHAP

Hydrological simulations have seen extensive use of machine learning (ML) models. However, the existing ML models face challenges in effectively handling temporal and spatial heterogeneity of the driving features and transparency of features. To improve the runoff prediction ability of ML and the confusion of runoff generation mechanism interpreted by ML, this study combined K-means with XGBoost (Extreme gradient boosting,) and SHAP (Shapely additive explanations,) to develop an interpretable ML-based hydrological model (KXGBoost) across the Continental United States as a case study. The results show that K-means clustering based on the interpretation of SHAP can effectively capture the temporal and spatial heterogeneity of runoff-driven features. And the performance and interpretability of data-driven hydrological models in runoff simulation can be significantly improved by KXGBoost. KXGBoost yields an NSE (Nash-Sutcliffe Efficiency) of 0.803 and 0.596 during the training and testing, respectively, representing an improvement of 0.089 and 0.029 as compared to the XGBoost model. KXGBoost has demonstrated its ability to identify multiple runoff generation mechanisms under multiple spatio-temporal perspectives. This study provided a novel perspective for understanding hydrological processes and improving runoff simulation and demonstrates the potential of ML-based hydrological models in water resources management.