Regional Prediction Of Daily PM_2.5 Concentration Based On Parametric And Nonparametric Methods

Research on air quality and carbon emission has important strategic significance for adjusting industrial production plan and urbanization speed.Improving the scientificity and accuracy of quantitative prediction of PM_2.5 concentrations has always been a hot issue of concern to governments at all levels,environmental protection and scientific research departments.On the one hand,the use of multi-source and multi-element meteorological observation data,emphasizing the regional structure and regional impact;On the other hand,the advanced research results of modern statistical methods are applied to the establishment of prediction models.In order to improve the shortcomings of repeated modeling of single-site models and considering only local influences,this paper studied parametric and nonparametric regional prediction models of PM_2.5 concentration based on the daily data of PM_2.5concentration and surface routine meteorological elements in 13 cities in Jiangsu Province from2013 to 2018.All the established regional prediction models are data-driven and have little dependence on PM_2.5 emission sources,physico-chemical transition mechanism and meteorological influence mechanism.The main research contents are as follows:（1）The PM_2.5 concentration,temperature,air pressure,relative humidity and wind speed in 13 cities in Jiangsu Province are analyzed by empirical orthogonal function（EOF）,which are decomposed into spatial mode EOF_S and time expansion coefficient EC_S.Predictors are selected based on the first time expansion coefficient EC1 of each variable.The selection method reflects the regional synergistic effect of small and medium-sized meteorological fields on PM_2.5,and highlights the physical significance of the influence of meteorological transmission on PM_2.5.（2）Based on EOF analysis and principal component regression（PCR）,a new regional prediction model of PM_2.5 concentration parameters,EOF PCR,is established.The results show that the average forecast accuracy of the model is 64.99%in four seasons.EOF-PCR region model has the advantages of high efficiency and effectiveness.On the one hand,regional PM_2.5concentration can be predicted simultaneously with less computing power and speed.On the other hand,the prediction results are accurate.（3）Based on a new window width factor and the adaptive N-W kernel regression estimator（ANWKRE）,an improved adaptive N-W kernel regression estimator（A*NWKRE）based on variable window width is proposed.The results of example analysis show that the improved A*NWKRE is better than the N-W kernel regression estimator（NWKRE）with fixed window width and the ANWKRE with variable window width.（4）Based on EOF analysis and A*NWKRE,a new nonparametric regional prediction model of PM_2.5 concentration EOF-A*NWKRE is established.The results show that the average forecast accuracy of the model is 74.56%in four seasons.The advantage of EOF-A*NWKRE domain model is to avoid the influence of high-dimensional data on non-parametric methods.Compared with the parametric region model,the prediction accuracy is higher with less data input.