Research On The Prediction Of PM_2.5Hourly Concentration Based On RNN-CNN Ensemble Deep Learning Model

Haze invasion not only seriously hinders the normal traffic trips,but also seriously endangers people's health.PM2.5 is the main component of haze so that controlling PM2.5 is the key to dealing with haze pollution.Therefore,it is of great significance to comprehensively understand spatio-temporal evolution of PM2.5 concentrations and to predict PM2.5 concentrations efficiently and accurately for traffic planning and air pollution control.Most of the current PM2.5 prediction models are generally used in a small area,the prediction effect is unstable and the generalization ability is weak.Based on a large number of research on the spatio-temporal characteristics and prediction models of PM2.5 concentrations at home and abroad,this paper designs a prediction model based on Recurrent Neural Network(RNN)which has strong memory using Gated Recurrent Unit(GRU)and a prediction model based on Convolutional Neural Network(CNN)which has strong feature expression ability for air quality time series in mainland China under deep learning framework,and chooses Stacking,an ensemble learning technique,to combine RNN and CNN so that we can take full advantage of both.Finally,the ensemble deep learning model RNN-CNN is proposed for PM2.5 hourly concentration prediction.The main contents of this paper are as follows:(1)Design an ensemble prediction model under deep learning framework.In PM2.5 hourly concentration prediction,considering 1)it is necessary to predict the future concentration using the contextual information on the timeline,which can be well achieved by RNN's excellent memory ability for time-series data;2)it is necessary to extract different levels of essential features from the high dimensional features for prediction,which can be well achieved by CNN's good feature expression ability for high-dimensional sequential data.Therefore,we construct a RNN prediction model using GRU and build a CNN prediction model for time series,then propose an emsemble model RNN-CNN combining RNN and CNN based on Stacking,taking full advantage of strong memory ability and good feature expression ability in the forecast.(2)Analyze the spatio-temporal characteristics of PM2.5 concentrations in mainland China in spatio-temporal multi-scale.Compare and analyze seasonal,monthly,weekly and intra-day variations of PM2.5 concentrations in seven urban agglomerations of China using mathematical statistics;then discovery the spatial distribution pattern,spatial agglomeration pattern and spatial movement rule of PM2.5 concentrations at different spatio-temporal scales using nearest neighbor effects,Kriging interpolation,global and local spatial autocorrelation analysis,and centroid model in spatial analysis.(3)Choose mainland China as the study area,and verify the validity of RNN-CNN prediction model.First,extend the predictors at different spatio-temporal scales according to the spatio-temporal characteristics of PM2.5 concentrations analysis;then use three prediction modes designed before to forecast the PM2.5 hourly concentrations of each station based on the air quality data of 1466 monitoring stations in mainland China in 2016;finally,compare and analyze the prediction error and fitting degree of three forecasting models in each region at different regional scales,and visualize the evaluation indexes in 3D scene,verifying the effectiveness and flexibility of the RNN-CNN prediction model.