The coupling model of random forest and interpretable method quantifies the response relationship between PM2.5 and influencing factors

Abstract

Ambient fine particulate matter (PM2.5) is affected by many factors, such as source emissions, meteorological conditions, and chemical reactions. Revealing the effects of these factors on PM2.5 is essential to understand the causes of PM2.5 pollution. The machine learning method can establish the non-linear relationship between influencing factors and PM2.5. Here, a coupling model of machine learning and interpretation method was constructed to comprehensively quantify the importance of influencing factors to PM2.5 from multiple dimensions and analyze the sensitivity of influencing factors. Among the primary indicators of influencing factors, the importance of emission, meteorological conditions, and atmospheric chemical reaction to PM2.5 is 49%, 29%, and 22%, respectively. In the secondary indicator of influencing factors, the transmission effect is the most important meteorological condition, with an important degree of 15%. The liquid phase reaction is the most important atmospheric chemical reaction, with an importance of 7%. Among the three levels of influencing factors, emission, transport distance, liquid phase reaction coefficient, aerosol acidity, and accumulation promotion coefficient are important factors. The sensitivity of a single factor is complex and changeable, and the interaction between emission and other important factors is the strongest among the two factors. Of which the interaction between transmission distance and emission during the observation period is the strongest, and the interaction coefficient is 1.82. Our study focuses on the effect of influencing factors on PM2.5, provides a basis for the analysis of the causes of PM2.5 pollution, and technical support for the treatment of PM2.5.