Research On PM_2.5 Prediction In The Three Northeastern Provinces Based On Multi-source Data

Long-term exposure to particulate matter≤2.5μm in size(PM_2.5)can seriously endanger human health,therefore,the monitoring of PM_2.5concentration is of great significance.China’s PM_2.5ground monitoring stations started late,lack of historical data,and ground monitoring stations are not fully covered and unevenly distributed.The long-term and high spatial resolution Aerosol Optical Depth（AOD）generated by Moderate Resolution Imaging Spectroradiometer（MODIS）Multi-angle Atmospheric Correction（MAIAC）algorithm can be used to predict PM_2.5concentration.The machine learning model is able to better explore the nonlinear coupling characteristics between PM_2.5and other variables,providing a high-performance prediction model with nonlinear variables for PM_2.5prediction.Correlation analysis has been conducted for PM_2.5,satellite remote sensing AOD,air pollutants and meteorological factors in three northeastern provinces from 2016 to 2020,and a univariate regression model has been established to quantitatively explore the correlation between PM_2.5and AOD.The results show that PM_2.5is positively correlated with AOD,carbon monoxide,sulfur dioxide,nitrogen dioxide,ground pressure and longitudinal wind speed,and negatively correlated with ozone,air temperature,relative humidity,latitudinal wind speed and sunshine hours,and the optimal one-way regression model of PM_2.5and AOD is a power function regression model.Through the spatial and temporal analysis of PM_2.5and AOD,the results show that the PM_2.5concentration and AOD values in three northeastern provinces show an overall decreasing trend from 2016 to 2020,but the decrease of AOD value is not obvious.The spatial distribution of PM_2.5concentration and AOD value is influenced by the industrial layout and population density,and the spatial distribution characteristics of both are very similar and have strong spatial correlation.Support Vector Machine Regression（SVR）and Random Forest Regressor（RFR）PM_2.5prediction models have been developed using AOD,air pollutants and meteorological factors.To improve the prediction accuracy of the models,Particle Swarm Optimization（PSO）has been used to optimize the SVR and RFR,and comparative model analysis has been performed.The results show that the mean R²values of SVR and RFR PM_2.5prediction models are 0.86and 0.88,the mean RMSE values are 11.01μg/m3and 9.25μg/m3and the mean MAPE values are 36.17%and17.16%,respectively.The mean R²values of the PSO-SVR and PSO-RFR PM_2.5prediction models optimized using the PSO algorithm are 0.88 and 0.89,the mean RMSE values are 10.30μg/m3and 8.97μg/m3and the mean MAPE values are 19.01%and16.90%,respectively.Comparing the four models,both SVR and RFR can achieve short-term or long-term PM_2.5concentration prediction,but RFR has better prediction effect compared with SVR.PSO has improved the performance of the base model for both,and has better optimization effect for SVR.Among the four models,PSO-RFR has the best prediction effect and is more advantageous for PM_2.5concentration prediction,which provides technical support for pollution warning and information management of air quality monitoring system in three northeastern provinces.