Analysis Of Influencing Factors And Prediction Of PM_2.5 In Dalian Based On Ensemble Learning

In recent years,with the rapid development of China’s economy,the environmental problems,especially the air pollution,have attracted more and more attention.With the increasing concentration of PM_2.5 in the air,the frequent occurrence of haze weather not only brings many influences on people’s normal work and life,but also harms people’s health and restricts the rapid development of China’s economy.The analysis of local PM_2.5 pollution factors and the accurate prediction of PM_2.5 concentration can make people do well in advance in protection work,reduce exposure in haze environment,and help the government to formulate pollution control measures.Compared with the traditional method,this paper proposes a PM_2.5 modeling method based on machine learning.According to the monitoring data of monitoring station,the influencing factors analysis and short-term prediction of PM_2.5 can be realized.This paper collected all the air monitoring data and meteorological data in Dalian in 2018,and processed and characterized these data.It was found that they had linear and nonlinear relationship with PM_2.5 data.At the same time,PM_2.5 distribution is also affected by human activities and natural factors in different regions or different periods.In order to measure whether these factors have effect on PM_2.5modeling,this study uses Random Forest model to test these factors,including air pollutants,meteorological factors,zoning modeling and quarterly modeling.The results show that the above factors can improve the accuracy of the model and have a significant effect on PM_2.5modeling.During this period,it was found that the best effect was using kmeans partition,R²could reach 0.919;In different seasons,the impact of O₃ and NO₂ on PM_2.5 can be found by using the characteristic importance index of random forest;Finally,using the above factors,a stacking model which is composed of Random Forest and Xgboost is constructed for short-term prediction of PM_2.5.Through the test of the prediction performance,the model can accurately predict the PM_2.5 concentration after 48 hours,and R² can reach 0.784 at least.The results show that the Kmeans-Stacking model based on temporal and spatial factors has higher prediction accuracy,and can accurately predict the PM_2.5 concentration in the next 48 hours.