水电能源科学

为有效应对珠江河口区咸潮上溯日趋加重的问题，利用磨刀门水道逐时观测数据，基于梯度提升决策树(GBDT)、极端梯度提升(XGBoost)、轻量级梯度提升(LightGBM)、类别提升(CatBoost)模型框架，结合可解释性的SHAP模型进行咸潮上溯逐时预报。结果表明，4种机器学习算法模型均具有较好的预报效果，其中CatBoost模型表现最好，24 h预见期模型纳什效率系数为0.738 5;基于SHAP模型特征重要性排序进一步优选输入因子，可以提高模型精度，优化后的CatBoost模型纳什效率系数、相关系数分别提升了0.30%、0.13%;对咸潮上溯预报不同特征进行SHAP分析可提高模型可解释性，分析发现盐度特征对咸潮上溯预报呈线性正相关影响，单一特征的SHAP分布图散点的分布越集中，特征重要性越大。

To address the increasingly severe issue of saltwater intrusion in the Pearl River Estuary, based on the hourly observational data of the Modaomen Waterway and machine learning frameworks including Gradient Boosting Decision Tree(GBDT), Extreme Gradient Boosting(XGBoost), Light Gradient Boosting Machine(LightGBM), and Categorical Boosting(CatBoost), hourly saltwater intrusion was predicted by the SHapley Additive exPlanations(SHAP) model. The results show that all four machine learning models have good predictive performance, with the CatBoost model achieving the best results, yielding a Nash-Sutcliffe efficiency coefficient of 0.738 5 for a 24-hour lead time. Feature importance ranking based on the SHAP model enabled further optimization of input factors, improving model accuracy. The Nash-Sutcliffe efficiency coefficient and correlation coefficient of the optimized CatBoost model increased by 0.30% and 0.13%, respectively. The SHAP analysis of different features enhanced model interpretability, revealing a linear positive correlation between salinity characteristics and saltwater intrusion predictions. The more concentrated the SHAP distribution of a single feature, the greater its importance is.