机器学习-多元线性回归预测铜的水生态基准  

Predicting the water ecological criteria of copper using machine learning and multiple linear regression approaches

作　　者：杨晓玲[1] 王梦晓 李晓娟 袁雅文 邵美晨 穆云松 白英臣[2] 吴丰昌[2] YANG Xiao-ling;WANG Meng-xiao;LI Xiao-juan;YUAN Ya-wen;SHAO Mei-chen;MU Yun-song;BAI Ying-chen;WU Feng-chang(School of Environment and Natural Resources,Renmin University of China,Beijing 100872,China;State Key Laboratory of Environmental Criteria and Risk Assessment,Chinese Research Academy of Environmental Sciences,Beijing 100012,China)

机构地区：[1]中国人民大学化学与生命资源学院,北京100872 [2]中国环境科学研究院,环境基准与风险评估国家重点实验室,北京100012

出　　处：《中国环境科学》2024年第7期3976-3985,共10页China Environmental Science

基　　金：国家自然科学基金资助项目(42277235,41521003)。

摘　　要：以铜和我国代表性水生生物作为研究对象,在原有生物配位体模型(BLM)理论框架基础上,通过梯度提升决策树算法筛选和评估环境要素(硬度、pH值和溶解有机碳)的特征重要性,建立多变量耦合的多物种水生毒性预测模型;在此基础上对毒性预测值进行物种敏感度分布(SSD)分析,构建适合我国水环境特征的铜淡水生态基准预测模型.研究发现,基于3门5科水生毒性数据建立的三参数耦合预测模型的预测准确率(RFx,2.0)比BLM模型提升了42%;以Sigmoidal-logistic1模型拟合SSD曲线的效果最佳(0.922<R2<0.991,0.0267<RMSE<0.0767,P>0.05),进而计算出我国流域水环境中铜的短期水生态基准推荐阈值为0.07350~15.38µg/L;基于机器学习的特征重要性分析定量识别了DOC在金属水生态基准制定中的关键作用,也为众多环境影响因素的集约化处理提供直接证据.与以BLM模型为代表的国外技术相比,研究工作在开发适用于中国水环境特征和管理需求的金属“原位”水质基准制定技术方面做了有益尝试,不仅节约了环境监测和管理成本,也体现了水环境管理的区域化和精准化.In this research,copper and representative aquatic organisms in China were investigated as a case study.Based on the theoretical framework of the biotic ligand model(BLM),the key environmental factors(hardness,pH and dissolved organic carbon)were screened by a gradient boosting decision tree algorithm,and multivariate coupled predictive models were established for predicting acute toxicities of different aquatic organisms.And then,the species sensitivity distribution(SSD)analysis was performed to predict the WQC of copper for protecting aquatic organisms,which was suitable for the characteristics of water environment in China.It was found that the prediction accuracy(RFx,2.0)of a three-variable model based on aquatic toxicity data of 3phylum and 5families was 42%higher than that of the BLM.The SSD curves for the nine organisms were best fitted using a sigmoidal-logistic model(0.922<R2<0.991,0.0267<RMSE<0.0767,P>0.05),and the threshold of short-term water ecological criteria of copper is recommended as 0.07350~15.38µg/L in the river basin of China.Based on the feature importance analysis from machine learning,the key role of DOC in the formulation of WQC for metals was quantitatively identified,and it also provided direct evidence for intensively treating multiple environmental factors.Compared with existing technologies including the BLM,our finding makes a beneficial attempt to develop an"in situ"WQC predictive model to meet water environment characteristics and management needs in China.It will reduce the costs for environmental monitoring and management,and enhance the regionalization and precision of water environment management.

关 键 词：铜 DOC 水质基准 模型校正 物种敏感度分布 机器学习 

分 类 号：X52[环境科学与工程—环境工程]