【摘要】 基于水质指标、气候指标、湿地运行参数3个方向,收集以往研究文献数据,通过3种机器学习模型预测人工湿地出水氨氮（NH₄⁺-N）、COD、磺胺甲噁唑（SMX）以及部分重金属的浓度.结果表明,随机森林（Random Forest）在整体性能上略优于XGBoost和LightGBM,其决定系数（R²）和均方根误差（RMSE）的表现更为稳定,尤其是在NH₄⁺-N和SMX的预测上取得更高精度（NH₄⁺-N预测的R²分别为0.93、0.89和0.87）.相比之下,在COD的预测中,3种模型的表现相对较弱,R²分别为0.71、0.61、0.64.通过引入SMOTE数据扩充技术,模型的预测性能和精度得到了显著的提升,尤其是对COD的预测性能提升幅度达7.04%～26.23%.本研究将数据分析与机器学习算法相结合,可为实际工程应用提供可行方法.更多还原

【Abstract】 Based on water quality indicators, climate indicators, and wetland operation parameters, data from previous studies were collected to predict the effluent concentrations of ammonia nitrogen（NH₄⁺-N）, COD, sulfamethoxazole（SMX）, and some heavy metals in constructed wetlands using three machine learning models. The results showed that the Random Forest model slightly outperformed XGBoost and LightGBM in overall performance, demonstrating more stable R² and RMSE values. In particular, it achieved higher accuracy in predicting NH₄⁺-N and SMX concentrations, with R² values of 0.93, 0.89, and 0.87, respectively, for NH₄⁺-N. In contrast, the models performed relatively weaker in COD predictions, with R² values of 0.71, 0.61, and 0.64, respectively. By incorporating the SMOTE data augmentation technique, the prediction performance and accuracy of the models were significantly enhanced, especially for COD, where improvements ranged from 7.04% to 26.23%. This study combines scientific data analysis with machine learning algorithms, providing a feasible approach for practical engineering applications.更多还原