Research On Air Quality Prediction Model Based On Spatial-temporal Data

In recent years,air pollution has become one of the hot issues of social concern.Air quality data can provide scientific basis for the accurate treatment of air pollution.In order to keep track of air pollution,many local governments have set up air monitoring stations to monitor air quality in their areas in real time.Therefore,research on air quality prediction can improve the government’s ability to accurately control air pollution.Artificial intelligence methods such as neural network and deep learning have been widely used in many fields.Some prediction methods have the following shortcomings:firstly,due to the diffusivity of the air,the air quality of the spatial location to be predicted is easily affected by the air quality of adjacent areas,and the existing research has not considered the spatial and temporal factors affecting the air quality.Secondly,when the traditional recurrent neural network model is directly applied to air quality prediction,the gradient disappearance and error accumulation will be caused by the increase of training sequence length.Thirdly,short-term data modeling is used to predict air quality without considering the periodicity of air quality data.Aiming at the above problems,the main research contents of this paper are as follows:Firstly,aiming at the spatial and temporal factors of air quality,a method of spatial division and evaluation for air quality prediction is proposed.Training on multiple sets of data to get the optimal partition radius and the layer number,to space zoning of monitoring sites,get the area of the monitoring site has a spatial correlation,the space on the sparse discrete spatial interpolation method is adopted to improve the air quality monitoring data of null value area filling,converts it to gather the regional data,get a fine-grained air quality data,to solve the air quality forecast in time and space factors to consider when the problem of inadequate.Secondly,to solve the problem of gradient disappearance and error accumulation,an ASTPN（ASTPN,Attention based on Spatial and Temporal Periodic Neural Networks）air quality prediction model was proposed.ASTPN model incorporates the circulation mechanism of neural network and attention,based on the factors of time and space,on the basis of comprehensive consideration of air quality data of short-term,long-term,and the characteristics of the cyclical trends,to solve the deficiency of the existing modeling methods rely on short-term trends,as well as the traditional cycle neural network model for training sequence length increases with the gradient of the existence of disappear and error accumulation problem.Finally,in order to verify the performance of ASTPN model,this paper took PM2.5concentration and SO₂concentration data of Beijing air quality monitoring station as the experimental data set,and compared ASTPN model with ARIMA model,SVR model,LSTM model and Geo MAN model for the experimental analysis of 24-hour air quality prediction.The experimental results show that the ASTPN model proposed in this paper has small prediction error and higher prediction accuracy.