Research On PM2.5 Monitoring Concentration Prediction Based On LSTM

The current concept of smart city is gradually spreading out,and many monitoring systems have been deployed in the city,and the air monitoring system is an important one.In response to the problem of compound air pollution in Chinese cities,especially PM2.5 pollution,a very important function of the air monitoring system is to carry out real-time effective monitoring and accurate prediction of PM2.5 concentration.However,existing prediction models can only use the history of the target prediction site when predicting the PM2.5 concentration of a single site.The data does not consider the time and space characteristics of PM2.5,which will make the prediction accuracy insufficient.To this end,this article mainly focuses on the study of PM2.5 monitoring concentration prediction based on LSTM(Long Short-term Memory),and studies the impact of different feature screening methods on PM2.5 concentration prediction.In addition,an in-depth analysis of the long-term dependence between the sites of the PM2.5 monitoring network is carried out,and based on this,a P-LSTM prediction model based on the Pearson two-way weighted network is proposed.The main content of this article includes:(1)Analyze the time-space characteristics of PM2.5-related influencing factors of the Fuzhou air monitoring network,and found that the time-series evolution trend of PM2.5 pollutant concentration is similar to the time-series evolution trend of other atmospheric pollutants,and discovered that there is a mutual influence between the changes in PM2.5 concentration among the monitoring stations in the monitoring network.In order to obtain actual PM2.5 data,we built a PM2.5 data collection site.The structure of the collection site includes photovoltaic power generation circuits,charge and discharge power management,communication systems,PM2.5 collection equipment,and monitoring management systems.(2)We use grey relational analysis and principal component analysis to analyze PM2.5-related influencing factors,and select the input features in the PM2.5 prediction model through the analysis results,and select those that have a greater impact on the PM2.5 concentration factors and establish an LSTM(Long Short-term Memory Neural Network)prediction model.The prediction results show that the input features selected by the principal component analysis method have a better prediction effect in the PM2.5prediction model.(3)This paper uses a two-way weighted network matrix to conduct a more in-depth analysis of the spatial correlation characteristics between the monitoring stations.By using the Pearson correlation coefficient to assign the weights of the two-way weighted network matrix,to mine the long-term dependence of the spatial relationship in the monitoring network,and constructed the spatial characteristics X(t)of monitoring network.Based on this,a P-LSTM prediction model is proposed,we take the spatial characteristics X(t)and the influencing factors screened by principal component analysis as the input of the P-LSTM model to predict the PM2.5 concentration.By comparing the prediction results of different prediction models,it is shown that the P-LSTM model has better prediction accuracy.In addition,the P-LSTM model is used to predict the actual PM2.5 concentration,which further verifies the superiority of the P-LSTM model.