基于CatBoost的水库入库径流预测及可解释分析OA

The operational regulation of newly constructed upstream reservoirs induces structural changes in the inflow formation mechanisms of downstream reservoirs due to artificial flow regulation.Conventional hydrological models,predominantly designed for characterizing natural rainfall-runoff processes,are inadequate for addressing runoff prediction requirements under dual natural-human regulated drivers.Focusing on the Jiangya Reservoir in the Lou River Basin,this study proposes a runoff prediction framework that integrates data-driven modeling and interpretability analysis.A CatBoost-based inflow prediction model is developed,combined with parameter sensitivity analysis and SHAP(Shapley Additive Explanations)interpretability methods,to unravel the impact mechanisms of upstream reservoir operation on the inflow of the downstream.The results demonstrate that:① The CatBoost model outperforms comparison models in predicting Jiangya Reservoir inflows,with a 6.3%～15.8%improvement in the Nash-Sutcliffe Efficiency(NSE)coefficient and a 19%～33%enhancement in prediction stability(measured by the Bias-Variance metric,BV).These quantitative advancements verify its superior capability in modeling complex hydrological correlations through the symmetrical tree structure,ordered target statistics,and ordered boosting mechanism.②Parameter sensitivity analysis optimizes the inflow prediction model by identifying the optimal combination of CatBoost hyperparameters.This process reveals the mechanisms by which the learning rate,number of iterations,tree depth,and regularization constraints affect the model,while also validating parameter synergy effects and strategies for enhancing model robustness.③ SHAP interpretability analysis quantitatively verifies the model's decision logic while revealing operational thresholds governing the transition between natural precipitation-dominated and artificial regulation-dominated inflow generation modes.This study validates the applicability of the CatBoost model for runoff prediction under dual natural-anthropogenic drivers,and the proposed parameter optimization framework and interpretability methodology provide technical references for inflow forecasting in cascade reservoir systems.