PM2.5 Concentration Analysis And Prediction Of LSTM Network Model

The main station of China National Environmental Monitoring Station reported the national air quality situation in 2018.The report shows that the air quality improvement in Xining City is the worst,and PM2.5 has an important impact on air quality.In order to predict and control the PM2.5 concentration,It is necessary to research and analyze PM2.5.Firstly,this article to the time span from January 1,2014 to December 31,2018separate sequence of PM2.5 concentration was analyzed in xining city,and found that PM2.5 sequence has obvious cyclical trend is not obvious,however,in order to better study of sequence,on the average daily concentrations of PM2.5 data resampling get monthly PM2.5 concentrations,and in turn to the two data sets SARIMA model is established and the model of Prophet,SARIMA model for the super parameter adjustment,joined the holiday effect in the Prophet model,On January 1,2019 to November 30,2019,the average daily PM2.5 concentrations to forecast,the results show that short-term prediction,SARIMA model is superior to the Prophet model,when the sequence of trend and have periodic,Prophet model is superior to the SARIMA model,when one of the less obvious,Prophet model is suitable for long-term forecast,root mean square error is 16.707,the effect of long-term forecasts.Secondly,this article analyzes the influence factors of PM2.5 concentrations in xining city,the grey correlation analysis was carried out on the numerical factors,and got the influencing factors of the correlation of sorting for the AQI>PM10>SO₂>total rainfall>wind speed>CO>wind scale>O₃>NO₂>humidity>the lowest temperature>visibility average temperature>the highest temperature,done to type influencing factors and concentrations of PM2.5 virtual variable regression,the results show that the two quality levels and weather classification variables impact on PM2.5concentrations,Snow and weather have a certain effect on reducing PM2.5concentration for other weather.,Finally,this article attempts to use the deep learning library Keras to construct an LSTM model with 26 influencing factors as a multivariate input to predict the PM2.5concentration,using the relevant data from 2014 to 2018 as training data,and the relevant data for 2019 As the test data,the root mean square error of the model is 13.553,and the model's long-term prediction effect is better than that of the Prophet model.