耦合可解释机器学习的水文模型在洪水模拟误差校正中的应用OA

With global climate change,extreme weather events and floods occur frequently.The research on high-precision hydrological forecasting is of great significance for the prevention and control of rainstorms and flood disasters.To address the issues of poor adaptability and unclear parameter interpretation of traditional physical models in complex terrains,this study took the Gaotian River Basin in Qingyuan City,Guangdong Province as an example,and adopted the TOPMODEL semi-distributed hydrological model,combined with machine learning XGBoost and interpretable technology to correct the flood simulation error.Parameter sensitivity analysis based on SHAP values was conducted through the random forest model,and the results of SHAP value analysis were coupled with the model parameter calibration to guide the optimization of model parameter calibration through SHAP values.The data calibrated and verified by the evaluation criteria was input into the XGBoost correction module to predict and correct the errors of the original simulation values.Finally,the corrected runoff sequence was output,and the error correction results of XGBoost were compared with traditional error correction methods such as Kalman filter correction method and the multiple regression correction method.The results show that:① SHAP value analysis is more suitable for explaining model behavior rather than directly guiding parameter optimization.In this paper,the calibration results of model parameters guided by SHAP value analysis are not as good as those of empirical value calibration.The reason might be that the SCE-UA algorithm itself is designed to automatically identify the effective combination of key parameters through global exploration,while pre-screening based on SHAP values might prematurely exclude parameters that play an important role only under specific parameter interactions.This limits the search space and optimization ability of the algorithm.② The XGBoost correction module has a considerable degree of interpretability and transparency when fitting the TOPMODEL.The correction performance of XGBoost for the input flood sequence is superior to that of traditional correction methods,with an NSE value of 0.98.Meanwhile,the correction module effectively captures the fluctuation pattern of errors,verifying the ability to correct the dynamic deviation of the TOPMODEL.It directly captures and reproduces the complex nonlinear dynamic characteristics of the hydrological system from historical data,which well solves the overfitting and complex system modeling problems existing in traditional methods.③ The model error of TOPMODEL has obvious nonlinearity,peak dependence,and time series dependence.However,the XGBoost module can effectively capture this nonlinear error structure and demonstrate strong correction ability,indicating that machine learning methods have advantages in dealing with systematic deviations of hydrological models.The coupling of interpretable machine learning and physical processes provides new ideas for flood simulation and forecasting research.