Air Quality Prediction Based On Spatio-temporal Deep Learning

With the rapid development of modern industry and transportation,rapid population expansion and urbanization,air pollution has become a global problem.Therefore,improving the scientificity and accuracy of air quality analysis and prediction can help the government to take effective prevention and control measures in a timely manner to protect public health and improve people’s quality of life.The formation and diffusion of air pollutants are affected by a variety of factors,which poses many challenges to the accurate prediction of air quality: 1)Air quality is affected by meteorological and other time factors as well as road network structure,land use and other spatial factors,which is the result of the complex interaction process between air pollutant particles and spatio-temporal factors.2)Due to equipment maintenance and software errors,there are many anomalies and missing values in the collected air pollutant data and related spatio-temporal characteristic data and some data have a very high proportion indeed.3)When predicting air quality over multiple time steps,performance decreases as the time step increases.It is necessary to further explore the correlation between the predictive sequence and historical sequence in order to improve prediction performance.Research in this area is still ongoing.Secondly,3D Convolutional Networks(C3D)are used for joint convolution of spatio-temporal information to extract spatio-temporal features.LSTM)to solve the long-term data series feature extraction,and finally,the extracted features through the fully connected network to achieve the prediction of air quality.The validity of the proposed model was verified based on the comparison of three different indicators with the mainstream single model and mixed model in air quality prediction.A spatio-temporal hybrid deep learning model called C3D-LSTM has been proposed to address the problem of predicting future air quality values at a target monitoring site using historical monitoring data collected from the target site and its nearby monitoring stations.Firstly,the Improved Inverse Distance Weight(IDW)and Simple Exponential Smoothing(SES)models were used to fill in missing data for all monitoring sites.Secondly,a 3D Convolutional Networks(C3D)is used to perform joint convolution on temporal and spatial information to extract spatio-temporal features.Long Short-Term Memory neural network(LSTM)is used to handle long-term data sequence feature extraction.The extracted features are passed through a fully connected network to predict air quality.Finally,the effectiveness of the proposed model is validated by comparing it with mainstream single and hybrid models for air quality prediction based on three different indicators.To solve long-term prediction of air quality for multiple stations,a model called SSF is proposed based on spatial data-assisted Seq2 seq learning.Firstly,from the perspective of two different types of features,the K-Nearest Neighbor(KNN)method is used to fill and embed missing values for all station data.Secondly,based on the temporal and spatial correlations among air quality data,a Seq2 Seq structure with fusion attention mechanism is used to characterize time dependence.An autoencoder with self-attention mechanism is added to capture spatial correlations among stations.Finally,based on a real dataset of 21 stations in Beijing,three different experimental scenarios are designed.Comparative experiments with baseline models and ablation experiments all demonstrate the effectiveness and superiority of the proposed model.