Research On Prediction Of PM2.5 Concentration Based On Spatial Dimension

PM2.5 is one of the main air pollutants.As air quality is getting worse,people have been paying more attention to prevention and control of PM2.5 in daily life.So it is significant to improve the prediction accuracy of PM2.5 concentration value.In this thesis,the methods of traditional machine learning and deep learning are used to explore the feasibility of PM2.5 concentration prediction based on spatial dimension.That reduces the dependence of PM2.5 concentration on historical data.This thesis mainly complete the following three works:1.For the lack of PM2.5 concentration data,this thesis first derived the PM2.5 concentration data of Xi’an in the past year.Then,it analyzes the correlation of PM2.5 concentration distributioin between monitoring stations by the method of correlation coefficient.Experimental results suggested there is high autocorrelation between the same station,and a medium high correlation with other monitoring stations.Further,this thesis analyses the data distribution correlation and the distance between stations,it finds the stations and the distance between them is approximately negative correlation.The farther the distance is,the weaker the correlation is.2.In view of the lack of space dimension features in the current PM2.5 concentration prediction model,a PM2.5 concentration prediction model based on the combination of distance factor and space-time dimension is proposed.First,based on the law of PM2.5 concentration distribution between monitoring stations,this study builds baselines by using linear regression and support vector regression in both time and space dimensions.Then,a prediction model based on the combination of time-space dimensions is proposed.It considers not only the historical data in time dimension,but also the relevance in spatial dimension.Besides,on the basic of space-time model,the distance factor is proposed to optimize the spatial dimension characteristics,and the correlation between PM2.5 concentration value and the distance between stations is fully considered.The findings indicate the prediction model based on space-time dimension can effectively improve the prediction accuracy of it.The introduction of distance factor further improves the prediction accuracy of the model.3.Traditional prediction model relies too much on the historical data in time dimension.Besides,the accuracy of prediction model based on space dimension is low.This thesis presents a PM2.5 concentration prediction model based on the optimization of long and short distance in space.In this model,neural network is introduced to predict PM2.5 concentration value.The nonlinear operation of activation function improves the fitting ability of the model.In addition,the optimization strategy of long and short distance in space is put forward to divide the surrounding monitoring stations according to the distance between the stations.With the deepening of neural network model,the influence of the further monitoring stations on the predicted value becomes smaller and smaller.Its weight is adjusted according to the training loss adaptively,and it implicitly learns the regular distribution between the correlation of the concentration value and the distance between the stations,which fully mines the inherent information of the data.The results indicate the prediction model of PM2.5 concentration value based on spatial long-short distance optimization significantly improves the prediction accuracy.