Research On PM2.5 Concentration Prediction Model Based On Deep Learning

With the rapid development of urbanization and industrialization in China,the problem of urban air pollution is becoming more and more serious.As one of the main pollutants of air pollution,PM2.5 is seriously endangering human health,so PM2.5 concentration prediction has become an indispensable task.Reasonable and accurate prediction of PM2.5 concentration has important guiding significance for government air defense work,social activity arrangements and urban residents' travel.First of all,this paper introduces the research background and significance of PM2.5 concentration prediction,and domestic and foreign research status,so that readers have a certain understanding of PM2.5 concentration prediction.Secondly,this paper analyzes the characteristics of PM2.5 concentration data and preprocesses the data.Mastering the change law and development trend of PM2.5 concentration data,and understanding the inherent characteristics and external influence factors of PM2.5 concentration data are the basis for accurate prediction of PM2.5 concentration.Therefore,this paper uses statistical methods to carry out autocorrelation and periodic analysis of the intrinsic characteristics of PM2.5 concentration,and analyzes the correlation of the main external factors affecting PM2.5 concentration.In order to make PM2.5 concentration data easy to be learned by the model,this paper numerically encodes the original data,processes missing values,constructs periodic features,and normalizes.Thirdly,this paper introduces the LSTM neural network and proposes GRA-LSTM and GRA-CLSTM on this basis.Existing machine learning-based prediction models can only use the historical data of the target prediction site without considering the regionality of PM2.5 when predicting the PM2.5 concentration at a single site,and cannot fully consider the regional effects of pollutants.For this question,GRA-LSTM uses gray correlation analysis to build spatial features and improves ordinary LSTM prediction models.GRA-CLSTM aims at the problem of information loss and gradient disappearance caused by ordinary LSTM for long time series,and introduces CNN,which is good at feature extraction,to improve the model.Finally,in order to verify the effectiveness and superiority of GRA-LSTM and GRA-CLSTM,this paper uses 13 air quality monitoring stations in Xi'an from June 1,2018 to July 6,2019 for a total of 9,624 hours of historical air quality Data and historical meteorological data of Xi'an City for simulation experiments,and the experimental results proved that GRA-LSTM and GRA-CLSTM have better prediction performance.