Leveraging Explainable Machine Learning to Understand the Impacts of Meteorology on PM2.5 Sources during a Household Energy Transition in China

COLLIN BREHMER, Christopher Barrington-Leigh, Jill Baumgartner, Sam Harper, Martha Lee, Xiaoying Li, Guofeng Shen, Talia Sternbach, Shu Tao, Xiang Zhang, Yuanxun Zhang, Ellison Carter, Colorado State University

     Abstract Number: 662
     Working Group: Identifying and Addressing Disparate Health and Social Impacts of Exposure to Aerosols and Other Contaminants across Continents, Communities, and Microenvironments

Abstract
Quantifying the impacts of energy and environmental policies on fine particulate matter (PM_2.5) air pollution is challenging due to multiple contributing sources and non-policy related factors like meteorology. Variations in meteorology over time could mask the impacts of an air pollution policy on the specific source it is targeting. Accounting for time-varying meteorological impacts on PM and its sources is critical for evaluating the impacts of an air pollution policy but has been difficult to incorporate in field-based impact evaluations of air pollution policies, which themselves are scarce. We collected outdoor and personal exposure PM2.5 measurements in 50 villages in rural Beijing across three winter seasons (2018-2022). Samples were analyzed for elemental concentrations and water-soluble ions. Source apportionment by Positive Matrix Factorization identified common sources in outdoor and personal exposure samples as nitrite, secondary ions, dust, transported dust, coal combustion, construction dust (outdoor only), and biomass combustion (personal exposure only). Our goal was to determine which sources were impacted by meteorology and which meteorological variables had the highest impact. Random forest models predicted source contributions using temperature, wind speed, boundary layer height, and relative humidity. The sources with the highest cross-validated predictability were dust for outdoor samples and secondary ions for outdoor and personal exposure samples (CV r²: 0.31-0.40). We used variable importance plots to determine the meteorological variables driving changes in PM_2.5 sources (relative humidity) and partial dependence plots to determine how they are driving changes in sources (positive relationship with secondary ions). By comparing source concentrations and meteorology across years we found higher PM_2.5 in winter 2019-2020 due to increased secondary ion concentrations related to higher relative humidity. These findings enhance our understanding of how PM_2.5 concentrations change over time in relation to meteorology, providing insights into the real-world effectiveness of air pollution policies.