Research On Air Quality Prediction Algorithm Based On Deep Learning

With the improvement of people’s living standards and health awareness,air pollution is receiving increased public attention.Air pollutants not only hinder social development,but also seriously damage people’s health.Accurate air quality prediction is of great practical significance.However,due to the influence of many potential factors and the complexity of its spatio-temporal variation patterns,the task of air quality prediction is still very challenging.The existing studies do not fully consider the temporal and spatial variation of air quality,resulting in insufficient effective ways to capture the spatio-temporal dependence of air quality.In addition,the lack of consideration of a variety of potential factors affecting air quality reduces the prediction accuracy of the method.Based on the research and analysis of the spatio-temporal variation regularity of air quality,combined with urban multi-source data sets,this thesis uses deep learning method to effectively capture the spatio-temporal variation pattern of air quality,and improves the accuracy of future air quality prediction.The main work and contributions of this thesis are as follows:(1)In this thesis,the spatial correlation of air quality is analyzed.The spatial adjacency matrix of air monitoring station is established by calculating Pearson correlation coefficient matrix and geographical distance between different stations.The spatio-temporal model is established by using graph attention network and long-term and short-term memory network to learn the spatio-temporal change pattern of air quality.The experimental results show that the deep spatio-temporal network has better performance than the time series prediction methods such ARIMA,LSTM.(2)This thesis considers the influence of various factors on air quality,collects and processes fine-grained real-time meteorological data,weather forecast data,road network structure data,point of interest data and altitude data,analyzes the relationship between various factors and air quality,and integrates them into spatio-temporal network to achieve more accurate air quality prediction.(3)In this thesis,a spatio-temporal convolution network based on multi-source data is designed,which uses the adaptive spatial graph convolution to capture the hidden spatial correlation of the air quality data.Considering the periodic change of air quality in temporal dimension,the strategy of data segmenting is adopted.The spatio-temporal module is composed of temporal convolution and spatial graph convolution to learn the spatio-temporal dependencies of each data segment.The spatio-temporal modules are fused by a Encorder-Decoder structure with attention mechanism which realizes sequentially predicting.The experimental results show that the proposed spatio-temporal convolution network can make a more accurate prediction of the future air quality compared with existing methods,and the ablation experimental results also verify the effectiveness of the components of the spatio-temporal convolution model.