Research On Air Quality Prediction Based On Deep Learning

In recent years,with the continuous advancement of industrialization and urbanization in China,the environmental pollution problems caused by air pollution induced factors such as industrial emissions and car exhaust have become increasingly prominent,and even multiple air pollution incidents have occurred.Therefore,it is of great significance to identify air pollution problems in advance,take effective measures to predict and study air quality.The current air quality prediction models mostly use RNN models such as LSTM and GRU for time series analysis,and there is still significant room for improvement in their prediction performance.In addition,most models only use data from a single monitoring station for research,but ignore the correlation between surrounding monitoring stations and the predicted stations under atmospheric circulation,which greatly limits the accuracy of predictions.In response to the above issues,the main work of the thesis as follows:(1)Analyze and explore the factors that affect air quality,collect hourly monitoring data from air quality monitoring stations and meteorological stations in Shandong Province in recent years,clean the data,fill in missing values,and fuse them.Analyze and validate the correlation between factors such as carbon monoxide,nitrogen dioxide,particulate matter(PM2.5 and PM 10),temperature,air pressure,wind speed,precipitation,and AQI,and select data from strongly correlated factors for subsequent research.(2)Propose the MSR-Informer(Multi Scale Residuals Informer)model to analyze the temporal data of the site and improve the accuracy of long-term prediction.Use Informer model to obtain long-term dependence between data,use multi-scale residual network to increase the Receptive field of the model,improve the accuracy of long series prediction,and verify the prediction effect of MSR-Informer model by comparing with LSTM,RNN and other models through experiments.(3)Propose the M-STGCN(Multi scale Spatial Temporary Graph Convolution Network)model to explore the correlation between monitoring sites.Use the graph convolution module to extract data features,obtain dynamic spatial associations between monitoring stations,and establish an air quality prediction model based on spatial features.(4)Based on the MSR-Informer model and the M-STGCN model,a hybrid model is further proposed.The purpose is to synthesize the information of two single models through the hybrid model,and aggregate the result of time prediction and space prediction,so as to finally obtain more accurate air quality prediction results.(5)Building an air quality prediction system based on hybrid models and MSR-Informer models,combined with front-end and back-end development techniques.