Research On Method Of PM_2.5 Concentration Estimation Based On Mixed Model In Yangtze River Delta

In recent years,the problem of air pollution has attracted more and more people’s attention.PM_2.5 is a serious hazard to human health due to its small particle size,strong adsorption capacity,long residence time,and long-distance transmission.Near-surface air quality monitoring stations are limited in number,and can only provide PM_2.5 concentrations at a certain spatiotemporal scale.Therefore,large-scale dynamic monitoring of PM_2.5 has become a hot issue in the field of atmospheric environment research.The current research is mainly focused on using the aerosol optical depth（AOD）data provided by polar orbit satellites to estimate the PM_2.5 concentration near the ground,but polar orbit satellites can only observe the same area 1-2 times a day.It cannot meet people’s demand for real-time information.Therefore,this study took advantage of the high time resolution of the new generation of geostationary satellite Himawari-8 launched by Japan,and used its AOD data combined with the European Centre for Medium-Range Weather Forecasts（ECMWF）meteorological data.Firstly,the correlation between PM_2.5,AOD and meteorological factors was analyzed.On this basis,the linear model（multiple linear regression）and nonlinear（random forest,gradient boost regression tree,support vector regression）were used to construct regional PM_2.5concentration estimation model.Finally,wavelet decomposition and machine learning were used to complete the construction of the mixed model in the Yangtze River Delta Urban Agglomeration（YRDUA）,and the accuracy of the model was evaluated.The main research contents and conclusions of this article are as follows:（1）The concentration distribution of PM_2.5 in the Yangtze River Delta shows significant diurnal,monthly,seasonal and spatial characteristics.In 2016 and 2017,the daily average values ranged are8.22μg/m³～149.47μg/m³,12.10μg/m³～191.27μg/m³,and the over-standard rates are12.56%and12.05%.The concentration of PM_2.5 in Zhoushan is the lightest,and the air pollution in Chuzhou is the most serious.The average monthly value of PM_2.5 in various cities generally reaches the lowest value of PM_2.5 concentration of 25μg/m³ in August and peaks in December.The PM_2.5 concentration shows the characteristics of high in summer and low in winter.The order of seasonal pollution degree is winter（66.41μg/m³）>spring（44.62μg/m³）>autumn（41.59μg/m³）>summer（30.35μg/m³）.The spatial distribution of PM_2.5 is high in the inland area in the east and low in the coastal area in the west.The PM_2.5 value in Huzhou is higher than other areas in the same period.（2）There is a significant linear correlation between PM_2.5 and meteorological factors.On the whole,temperature,boundary layer height,relative humidity,albedo,horizontal wind speed and precipitation are negatively correlated with PM_2.5 concentration,and vertical wind speed and air pressure are positively correlated.（3）In order to reveal the advantages and disadvantages of linear regression and non-linearity in the estimation of PM_2.5 concentration,two PM_2.5 concentrations obtained by estimation methods were compared and analyzed.The results show that the PM_2.5 mass concentration estimation model based on machine learning in YRDUA is superior to the multiple linear regression model.The optimal models in Shanghai,Hangzhou,and Nanjing are gradient boost regression tree,and the RMSE are 13.511,12.686,14.079,respectively.The RMSE of the optimal model（random forest）in Hefei is 24.529.（4）Using wavelet function to decompose the PM_2.5 time series of major cities in the YRDUA,a mixed model of PM_2.5 concentration estimation in the Yangtze River Delta region was constructed,and the PM_2.5 concentrations obtained by various estimation methods were compared and analyzed.The results show that the WT-GBRT models in Shanghai and Hefei perform best,and the WT-RF models in Hangzhou and Nanjing have the highest accuracy.Compared with the single model,the proposed mixed estimation model of fused wavelet decomposition has an average decrease of about 23.17%and 25.96%.