利用改进XGBoost模型预测和分析湿地潜流带地下水中硝态氮含量  

Using an improved XGBoost model to predict and analyze nitrate nitrogen content in groundwater of wetland hyporheic zones

作　　者：周念清[1] 夏明锐 陆帅帅 郭梦申 王在艾 赵文刚 ZHOU Nianqing;XIA Mingrui;LU Shuaishuai;GUO Mengshen;WANG Zaiai;ZHAO Wengang(Department of Hydraulic Engineering,Tongji University,Shanghai 200092,China;Hunan Institute of Water Resources and Hydropower Research,Hunan Changsha 410007,China)

机构地区：[1]同济大学土木工程学院水利工程系,上海200092 [2]湖南省水利水电科学研究院,湖南长沙410007

出　　处：《上海国土资源》2024年第2期41-47,共7页Shanghai Land & Resources

基　　金：国家自然科学基金项目(42077176,42272291,42242202)。

摘　　要：湿地潜流带是地下水中氮循环的重要场所,以洞庭湖湿地潜流带为研究对象,探讨地下水中氮素迁移转化影响因素与作用机制。在湘江入湖口湿地设置了4个剖面共16口监测孔,进行了为期一个水文年的地下水取样与测试分析。研究选定的特征参数包括氧化还原电位(Eh)、溶解氧(DO)、水温(T)、地下水位(H)及埋深、酸碱度(pH)以及溶解有机碳(DOC)等。利用XGBoost方法建立机器学习模型,用于预测硝态氮的相对浓度,并通过贝叶斯优化(BO)、麻雀搜索算法(SSA)、粒子群算法(PSO)分别对XGBoost预测模型进行超参数优化,得到最佳XGBoost预测模型(BO-XGBoost)。在此基础上,采用SHAP(SHapley Additive exPlanations)方法对BO-XGBoost模型进行可解释性分析。研究结果表明,BO-XGBoost模型的性能最好,在训练集与测试集的决定系数均超过0.90;可解释性分析结果和相关分析都揭示了Eh、DO、T、H、pH和DOC等影响因子对湿地潜流带地下水中硝态氮含量的影响是逐渐降低的规律。The hyporheic zone in wetlands is an important area for nitrogen cycling in groundwater.The hyporheic zone of Dongting Lake wetlands is taking as the research object,this study explores the influencing factors and mechanisms of nitrogen migration and transformation in groundwater.4 profiles and a total of 16 monitoring wells were set up in the wetland at the entrance of the Xiangjiang River,and groundwater samples were tested and analyzed for one hydrological year.The selected characteristic parameters for the study include redox potential(Eh),dissolved oxygen(DO),water temperature(T),groundwater level(H)and burial depth,pH,and dissolved organic carbon(DOC).An XGBoost machine learning model is established to predict the relative concentration of nitrate nitrogen.The optimal XGBoost prediction model(BO XGBoost)is obtained by using Bayesian Optimization(BO),Sparrow Search Algorithm(SSA),and Particle Swarm Optimization(PSO)algorithms to optimize the hyperparameters of the XGBoost prediction model.Based on this,the SHAP(Shapley Additive exPlans)method is used to analyze the interpretability of the BO-XGBoost model.The research results indicate that the BO-XGBoost model has the best performance,with determination coefficients exceeding 0.90 in both the training and testing sets.The interpretability analysis results and correlation analysis reveal that the impact of factors such as Eh,DO,T,H,pH,and DOC on the nitrate nitrogen content in groundwater in wetland hyporheic zone gradually decreases.

关 键 词：湿地潜流带 硝态氮 机器学习 XGBoost SHAP 

分 类 号：P641[天文地球—地质矿产勘探]