Influencing Factors Analysis And Forecast Of PM_2.5 In Fushun City Based On Multi-Site Data

Air quality is related to the production and life of people,which is a key factor in evaluating the comprehensive strength of a city.PM_2.5 is an important air quality evaluation index.Analyzing the influencing factors of PM_2.5 concentration and predicting it can provide early warning information for production and life,and help the government and relevant departments to propose timely and effective prevention and control measures.Most of the existing PM_2.5 prediction methods considered only the historical data of the target station,and ignored the spatial correlation of PM_2.5.Therefore,this thesis focused on the data of 24 monitoring stations in Fushun city to find the influencing factors of PM_2.5 concentration,meanwhile,to investigate the prediction performance of the proposed model on PM_2.5 concentration at the target site with single-site information and multi-site information considering the spatial correlation of PM_2.5 concentration over the study area.Firstly,the spatial and temporal distribution characteristics of PM_2.5 concentrations were described based on the data from 24 monitoring stations in Fushun city in 2020 to build a foundation for PM_2.5concentration prediction.Secondly,to prevent redundancy of features in prediction,the Generalized Additive Model was used to analyze the influencing factors of PM_2.5concentrations.Finally,considering the spatial correlation of PM_2.5,the prediction models with single-site and multi-site information were constructed and compared.The research results are as follows:（1）The PM_2.5 concentration in Fushun had obvious seasonal differences,with the highest value in winter and the lowest in summer,and the PM_2.5 concentration in spring,summer and autumn was significantly different from that in winter.On the monthly characteristics,the PM_2.5 concentration reached the highest in January,and the concentration in January was more than twice as high as that in the rest of the months.On the hourly characteristics,from 21:00 at night to 9:00 the next day,the PM_2.5concentration was more stable,about 54,and from 10:00 am to 20:00 pm,the PM_2.5concentration trend showed a"U"shape.Particularly,the PM_2.5 concentration reaches the lowest in the whole day at 16:00 pm.The PM_2.5 concentration in Fushun had spatial autocorrelation,and the analysis of its spatial distribution characteristics showed that the PM_2.5 concentration in the central area was higher than that in the border area.（2）The Generalized Additive Model can effectively describe the relationship between each factor and PM_2.5 concentration.Atmospheric pressure showed a linear relationship with the variation of PM_2.5 concentration,while PM₁₀,NO₂,CO,O₃,humidity,wind speed,wind direction and PM_2.5 concentration showed a non-linear relationship.In addition to the influence of individual factors,the interactions between O₃ and SO₂,atmospheric pressure significantly affected variation of PM_2.5concentration at the α=0.05 level.Finally,the effect of each factor and its interaction on PM_2.5 concentration was quantitatively analyzed by model visualization.（3）Considering the spatial autocorrelation of PM_2.5 concentrations over the research area,this thesis established Random Forest Regression（RFR）and Support Vector Regression（SVR）models to predict PM_2.5 concentrations at the target site based on single-site and multi-site information.Before prediction,the model parameters were optimized by Particle Swarm Optimization（PSO）and Grid Search method,and it was found that the model optimized by PSO had better prediction performance and faster convergence rate.Compared with single-site information,the model with multi-site information had less error more accuracy when used to forecast the PM_2.5 concentration of the target site.Meanwhile,the fit of the SVR model under the same optimization method was about 5%better than that of the RFR model for both single-site and multi-site information.Finally,when predicting PM_2.5 concentrations of the target site,the PSO-SVR model should be selected for prediction based on multi-site information if the data are more adequate.