Research On Prediction Method Of PM2.5 Concentration Based On CNN-LSTM

In recent years,air pollution which has seriously affected human health and daily life,and has caused a negative impact on the sustainable development of society.It brought haze days,the most direct manifestation of the deterioration of air quality,during which fine particulate PM2.5 is the main cause of haze formation and has caught increasing attention.Although many air quality monitoring stations have set up nationwide for real-time monitoring,the release of air quality information itself can not meet the public’s demand.With the advent of the era of big data,a high priority is set to forecast PM2.5 accurately with massive historical data.Under the background above,PM2.5 concentration prediction based on deep learning algorithm is studied in this thesis.The main research contents are as follows:(1)This thesis summarizes the research methods of PM2.5 concentration prediction at home and abroad.In view of complexity of mechanism model prediction and inadequacy of statistical prediction model,a hybrid deep learning algorithm based on convolution neural network and long-term memory network is used to predict PM2.5 concentration.(2)The data set of Qingdao from 2014 to 2018 was obtained by reptile technology from open channels.The model used the data of the first 24 hours to predict the PM2.5 concentration of the next 1 hour.Then the changing trend of PM2.5 concentration and the correlation between PM2.5 concentration and 11 influencing factors including concentrations of 6 pollutants(PM2.5,PM10,NO2,SO2,CO,O3)and 5 meteorological factors(temperature,dew point,relative humidity,wind speed,sea level pressure)were analyzed in this thesis.(3)The model was developed and designed with using the open source Pytorch deep learning framework of Facebook with which a concentration prediction model of PM2.5 based on CNN-LSTM was built.The main process is as fllow:Input 11 kinds of influence factors as models and the feature extraction is carried out by convolutional neural network module,then PM2.5 concentration is predicted by LSTM network.(4)The five prediction methods:deep learning prediction model based on CNN-LSTM,deep RNN,deep LSTM,DRT and SVR were compared with three different indexes,MAE,RMSE and Corr.The experimental results show that the deep neural network prediction model based on convolutional neural network and long short-term memory network constructed in this thesis has the highest precision and the best prediction effect.