机器学习耦合受体模型揭示驱动因素对PM_(2.5)的影响  

Machine Learning Coupled with Receptor Model to Reveal the Effect of Driving Factors on PM_(2.5)

作　　者：许博 徐晗 赵焕 张忠诚 高洁 李岳 冯银厂[1,2] 史国良 XU Bo;XU Han;ZHAO Huan;ZHANG Zhongcheng;GAO Jie;LI Yue;FENG Yinchang;SHI Guoliang(State Environment Protection Key Laboratory of Urban Particulate Air Pollution Prevention,College of Environmental Science and Engineering,Nankai University,Tianjin 300350,China;CMA-NKU Cooperative Laboratory for Atmospheric Environment-Health Research,Tianjin 300074,China;College of Computer Science,Nankai University,Tianjin 300350,China)

机构地区：[1]南开大学环境科学与工程学院,国家环境保护城市空气颗粒物污染防治重点实验室,天津300350 [2]中国气象局-南开大学大气环境与健康研究联合实验室,天津300074 [3]南开大学计算机学院,天津300350

出　　处：《环境科学研究》2022年第11期2425-2434,共10页Research of Environmental Sciences

基　　金：国家自然科学基金项目(No.42077191);中央高校基本科研业务费专项(No.63213072)。

摘　　要：PM_(2.5)主要受排放源、大气化学、气象条件等驱动因素的非线性影响,了解驱动因素对PM_(2.5)浓度的影响十分重要.本研究基于南开大学大气环境综合观测超级站的逐时在线观测数据,耦合机器学习方法和受体模型,揭示了驱动因素的重要性以及对PM_(2.5)浓度的影响.结果表明:①2018年11月—2020年10月观测地点的PM_(2.5)浓度范围为3.21~291.80μg/m^(3),采暖季PM_(2.5)浓度和化学组分均高于非采暖季.②使用受体模型解析PM_(2.5)的来源及其贡献,发现观测期间二次源的贡献率(44.7%)最高,其他依次为燃煤源(23.6%)、机动车排放源(11.0%)、扬尘源(9.9%)、生物质燃烧源(7.2%),工业源的贡献率(3.6%)最小.③利用随机森林-SHAP模型量化排放源、大气氧化能力、气象条件等驱动因素对PM_(2.5)浓度的影响,发现观测期间排放源对PM_(2.5)浓度的影响程度为54.3%,高于其他驱动因素;气象条件对PM_(2.5)浓度的影响程度次之,为32.4%;大气氧化能力对PM_(2.5)浓度的影响程度相对较低,为13.3%.在采暖季和非采暖季,各驱动因素对PM_(2.5)浓度的重要性在排序上没有变化,然而驱动因素对PM_(2.5)浓度的影响程度有所不同.采暖季排放源对PM_(2.5)浓度的影响程度高于非采暖季,采暖季大气压对PM_(2.5)浓度的影响程度低于非采暖季.研究显示,排放源对PM_(2.5)的影响相对较大,气象条件和大气氧化能力对PM_(2.5)浓度的影响也不容忽视.There is a non-linear relationship between PM_(2.5) and driving factors,such as emission sources,atmospheric chemistry,and meteorological conditions.Hence,it is important to understand the effects of driving factors on PM_(2.5) concentration.Based on the hourly online observation data of the Atmospheric Environment Comprehensive Observation Superstation of Nankai University from November 2018 to October 2020,this study combined the machine learning method with the receptor model to reveal the importance of driving factors and their impact on PM_(2.5) concentration.The results showed that:(1)PM_(2.5) concentration was 3.21-291.80μg/m^(3) at the observation site during the measurement campaign,and PM_(2.5) concentration and chemical species in heating season were all higher than those in non-heating season.(2)The receptor model identified the source and contribution of PM_(2.5),and the contribution of secondary sources during the observation period was the highest(44.7%),followed by coal-fired sources(23.6%),vehicle emission sources(11.0%),dust sources(9.9%),and biomass combustion(7.2%),and the contribution of industrial sources was the lowest(3.6%).(3)This paper also explored the effects of driving factors such as emission sources,atmospheric oxidation capacity,and meteorological conditions on PM_(2.5) concentration through random foreat-SHAP model.The effect of emission sources was 54.3%,which was much higher than other factors during the measurement campaign,the effect of meteorological conditions was the second(32.4%),and the effect of atmospheric oxidation capacity was lowest(13.3%).In the heating season and non-heating season,the importance ranking of driving factors on PM_(2.5) concentration didn't change,but the influence of the driving factors on PM_(2.5) concentration changed significantly.The impact of emission sources on PM_(2.5) in the heating season was significantly higher than that in the non-heating season,and the impact of atmospheric pressure on PM_(2.5) in the heating season was lower than that

关 键 词：颗粒物 受体模型 随机森林 驱动因素 影响 

分 类 号：X511[环境科学与工程—环境工程]