Research Of Air Quality Prediction Based On Spatiotemporal Deep Neural Networks

With the development of technology and urban industrialization,air pollution has gradually come into people’s view,air pollution has a significant impact on plants,climate,and in serious cases,can also endanger human health.However,due to the complexity and variability of factors affecting air quality,air quality prediction still faces many challenges and difficulties.Currently,most air quality prediction models use approximate time series prediction methods.Although these models can effectively utilize the temporal trends of air quality,they often ignore the spatial correlation of air quality data or do not explore the deep spatial and temporal patterns,and in addition,the complex change patterns of air quality data are also an important factor that hinders the accuracy of prediction.To address the above problems,this paper uses spatio-temporal deep learning methods to improve the accuracy of air quality prediction by combining the effects of meteorological and geographical location factors,and the main work is as follows:(1)In order to fully explore the factors affecting air quality,this paper processes and analyzes multiple sources of data affecting air quality.Air pollution data,meteorological data,and spatial and temporal data are effectively integrated into the air quality prediction task through a targeted pre-processing approach and an analysis of the correlation between air quality and various influencing factors.(2)In order to better extract the spatio-temporal features of multi-source data,a3DCNN-Bi LSTM prediction model is proposed in this paper.The model first receives tensor inputs and extracts multidimensional spatial information through a 3D convolutional network;then,a bidirectional long and short-term memory network is used to model temporal information and extract the global features of the input sequence;Finally,the global features are input to the linear regression layer to obtain the prediction results of future air quality.The experimental results show that the model outperforms traditional prediction models such as ARIMA and deep learning prediction models such as LSTM and GRU.(3)In response to the shortcomings of the 3DCNN-Bi LSTM model,the VMD-GAT-Bi LSTM spatio-temporal prediction model is further proposed in this paper.First,the predicted data are decomposed into a certain number of components with relatively uniform fluctuations through variational modal decomposition;second,each component is combined with other features to construct several graph structures based on the location and correlation relationships among air monitoring stations;then,these graph structure features are extracted through graph attention networks and two-way long-and short-term memory networks at different levels and the prediction results are output;finally,the prediction results of each component are combined at the output layer to achieve the air quality prediction for the next 24 hours in Beijing.Comparison experiments with existing methods show that the model can better exploit the spatio-temporal correlation among monitoring stations and has better short-term and long-term air quality prediction effects.