Research Haze Prediction Based On Deep Learning

With the rapid economic development of China,more attention has been paid to the process of social industrialization and urbanization,air quality has continued to deteriorate,haze pollution has become more frequent,and people’s daily lives and production processes have been seriously affected.The primary factors that determine haze pollution are PM2.5 and PM10 concentrations.And the higher the concentration,the more severe the smog pollution.Therefore,the prediction of PM2.5/PM10 concentration is important.This paper use the deep neural network to predict the concentration of PM2.5 and PM10,multi-node structure can learn the complex relationship between nonlinear features better,ensuring the model’s feature learning ability,and improving the prediction model’s prediction accuracy.The specific researchse of this paper are as follows:(1)In order to study the periodic variations of PM2.5 concentration sequence in different time scales and the distribution in the time domain,continuous complex morlet wavelet was used to transform the PM2.5 concentration sequence to obtain complex wavelets at different time scales.According to further calculations and analysis,when the PM2.5 was analyzed on the hourly time scale,it was found that the 24-hour period of change was obvious during periods of severe haze pollution;when the PM2.5 was analyzed on daily time scale,it was found that the 362-day period of change was obvious。(2)The paper construct deep confidence neural network model to predict the PM2.5/PM10 concentration in Chengdu city,which consists of RBMs neural network and BP neural network.The model use the t_th moment data to predict the data at the(t+6)_th moment,the results show the more number of hidden layers,the higher the prediction accuracy,after achieving 5 layers and 6 layers respectively,the accuracy of PM2.5 and PM10 changes little.In the same network,the accuracy of PM2.5 prediction is obviously higher than that of PM10,4 hidden layers and 5 hidden layers of DBNs has the best prediction result of PM2.5 and PM10 respectively.The prediction result of DBNs-BP is better than the deep BP neural network.(3)The paper construct DRNN composed of recurrent neural network and deep DLSTM composed of long and short time memory neural network respectively to predict the concentration PM2.5/PM10 at the moment of t+6 by using the data of previous t moments and t=24 according to the time cycle characteristics of PM2.5.The results show that the more number of hidden layers,the higher the prediction accuracy.When the number of layers reaches a certain value,the accuracy does not change much.The prediction result of DLSTM neural network is better than that of DRNN and the prediction result of DRNN is better than that of deep BP neural network.The DRNN with seven hidden layers and DLSTM with five hidden layers predict PM2.5/PM10 with best results.