PM_2.5 Prediction Model Research Based On Multi-site

In recent years,with the development of society,haze and other air pollutions still pose a threat to human health.PM2.5 is an important component of haze,and it is crucial to grasp the change trend of PM2.5 concentration in advance for haze control.However,PM2.5 prediction is a complex and nonlinear problem because of the influence of climate,spatial location and other factors.How to predict the PM2.5precisely and efficiently is the key issue nowadays.In the most of papers,PM2.5 prediction models only take the time dependence into account while regardless of the spatial dependence.In this paper,we propose both PM2.5 prediction model based on dynamic mode decomposition and PM2.5prediction model based on multi-site long short-term memory.The specific work of this paper is as follows:(1)Considering that the previous PM2.5 prediction models only predict a single site based on the historical data,we propose a PM2.5 prediction model based on dynamic mode decomposition(DMD).The input of dynamic mode decomposition is PM2.5 historical data of all sites in Beijing.It can predict the PM2.5 data of all sites at one time,which greatly improves the efficiency.Furthermore,the prediction model based on dynamic mode decomposition has low requirement on data volume,which greatly saves the cost of data acquisition.(2)Considering that PM2.5 diffuses regionally and PM2.5 of target site is affected by PM2.5 of other sites,we propose a PM2.5 prediction model based on multi-site long short-term memory(MS-LSTM)to predict PM2.5 of target site.The model introduces PM2.5 historical data of multiple sites,uses cosine similarity to calculate the spatial correlation of target site with the other sites,and gives different weight to the relevant sites.The model can extract both time dependence and spatial dependence,which improves the prediction accuracy of PM2.5 for target site.The data set used in this paper is PM2.5 historical data of 36 air quality monitoring sites in Beijing.Our goal is to forecast PM2.5 concentration of the next hour.In order to verify the effectiveness of the method,SVR model,DMD model and LSTM model are selected as the benchmark model.In the experiment,MAE is9.2645 and RMSE is 14.6551.The results show that MS-LSTM is better than other models.