Spatiotemporal change of PM2.5 concentration in Beijing-Tianjin-Hebei and its prediction based on machine learning

Abstract

For decades, PM_2.5 (a type of fine particulate matter) in large urban areas has had a profound negative impact on human health. In this study, spatiotemporal analysis and four machine learning methods (XGBoost, ANN, CNN and MLR) were used to assess the changes and drivers of PM_2.5 concentrations in the Beijing-Tianjin-Hebei (BTH) from 2016 to 2019 based on 68 stations. The results indicated a significant decrease in PM_2.5 concentrations in BTH region (average decrease of 7.69 μg/m³/yr), especially in the southwest region where pollution is the most serious, and the overall annual average still exceeded the national standard. In spatiotemporal modeling, XGBoost effectively captured the spatial characteristics of PM_2.5 pollution and achieved the most robust prediction in general (RMSE = 22.11 μg/m³, MAE = 15.18 μg/m³, R² = 0.64). The SHapley Additive exPlanations (SHAP)-based global and local driving analyses revealed that CO had the greatest relative impact on PM_2.5 (52.46 %), while NO₂ and SO₂ were also important driving factors, with variable importance values of 10.68 % and 6.01 %, respectively. Moreover, temperature and surface humidity are key meteorological drivers of the formation and development of PM_2.5 pollution. It is also worth noting that topography is an important geographic background for the formation of haze in the BTH region, which may induces air pollution under unfavorable meteorological conditions and hindering the improvement of air quality under favorable meteorological conditions. The results of this study deepen our understanding of air pollution and its driving factors in important urban agglomerations in China.