PM2.5 pollution is a complex process mainly affected by emission sources and meteorological conditions. However, it is hard to accurately assess the effects of emission sources and meteorological conditions on the variation of PM2.5 concentrations in the complex atmospheric environment. In this study, the Random Forest model with Shapley Additive exPlanations (RF-SHAP) and Partial Dependence Plot (RF-PDP) was combined with Positive Matrix Factorization (PMF) to evaluate the impacts of various factors on PM2.5 pollution. The results show that anthropogenic emissions and meteorological conditions contributed about 67% (40.5 μg/m3) and 33% (19.7 μg/m3) to variation in PM2.5 concentrations, respectively. Specifically, secondary nitrate (SN) had the greatest impact among all sources (about 45%). Hence, we further explore the impacts of the primary sources and meteorological conditions on SN formation. Coal combustion and vehicle emissions significantly contribute to the formation of SN by providing a large number of precursor NOX. Additionally, the RF-PDP method was further employed to estimate the synergistic effects of primary sources and meteorological conditions on SN formation. The results help reveal strategies to simultaneously reduce SN by controlling primary emissions under suitable meteorological conditions. This work also suggests that the machine learning model can utilize online datasets well and provide a reliable approach for analyzing the causes of PM2.5 pollution.
Keywords: Machine learning; PM(2.5); PMF; Random forest; Secondary nitrate.
Copyright © 2022 Elsevier Inc. All rights reserved.