A Multi-step Prediction Method For PM_2.5 Concentration Considering Spatio-temporal Correlation Under Deep Learning Framework

In recent years,China’s air pollution problems are frequent,and the haze has seriously affected people’s life and health.PM_2.5 is the main component of haze and an important indicator to measure the quality of atmospheric environment.Using historical data to effectively and accurately predict the PM_2.5 concentration over a long period of time in the future is of great guiding significance to formulate economic policies and control air pollution,etc.In view of the fact that most current studies on PM_2.5 concentration prediction considering spatio-temporal correlation are limited to single-step prediction of PM_2.5 concentration,and the spatial feature extraction method involved in the prediction involves much manual intervention,this paper proposes a multi-step prediction model for PM_2.5 concentration considering spatiotemporal correlation under the framework of deep learning.First of all,in terms of time dimension,this paper refers to the sequence generation method in the field of natural language processing,and bases on the encoder-decoder model,integrates attention mechanism and regularization training strategy,and proposes a set of time series multi-step prediction methods for prediction accuracy maintenance.Then,taking time and space dimensions into comprehensive consideration,this paper combines the time series multi-step prediction method and the modeling ability of the graph convolutional neural network to spatiotemporal dependence,and proposes the multi-step prediction model for PM_2.5 concentration considering spatiotemporal correlation,which effectively improves the prediction accuracy.The main research contents of this paper are as follows:（1）A set of time series multi-step prediction methods is proposed to maintain the prediction accuracy.According to continuous time series multi-step prediction problem,the paper bases on the encoder and decoder in the field of natural language processing,coalesces the mechanism of attention and regular training strategy,implements time series multi-step prediction "end-to-end" and "step",and improves the predictive accuracy while minimizing prediction accuracy of each step.The experimental results show that the attentional mechanism has a better effect on decreasing the decay rate of the prediction accuracy than the regular training strategy,the two have similar effect on improving the accuracy of the sequence prediction,and the combination of the two on the encoder-decoder model can obtain the best effect of the sequence prediction.（2）A multi-step prediction method for PM_2.5 concentration considering spatiotemporal correlation is proposed.Considering the spatio-temporal dependence of PM_2.5 concentration,this paper integrates the spatial distribution characteristics extracted by the graph convolutional neural network in the time series multi-step prediction model,and proposes a multi-step prediction method for PM_2.5 concentration considering spatiotemporal correlation.Experimental results show that the combination of spatial and temporal characteristics can further improve the accuracy of PM_2.5 concentration sequence prediction.In this paper,from the perspective of spatial and temporal correlation,the graph convolutional neural network and the time series multi-step prediction method oriented to the maintenance of prediction accuracy are organically integrated,and a multi-step prediction method for PM_2.5 hourly concentration under deep learning framework is constructed,which can realize efficient and accurate prediction of PM_2.5 concentration within dozens of hours.The research results of this paper can provide method support for urban air quality prediction.