Prediction Of Urban Air Quality Based On Deep Learning

The air quality issue in the process of rapid urbanization has always been of great concern,which relates to urban ecological safety and the healthy life of residents.Therefore,prediction of urban air quality attracts much attention.In this thesis,the data of hourly historical air quality and meteorological data from Shanghai Air Quality Monitoring Network were collected as research samples.The correlation between urban air quality index（AQI）and major atmospheric pollutants was analyzed.The change trend of surface ozone（O₃）concentration and meteorological factors in different seasons was also analyzed.Deep learning related technologies such as Bidirectional Long Short Term Neural Network（BiLSTM）were combined with modal decomposition methods such as CEEMDAN to predict urban air quality index and surface O₃concentration.The results showed that there were differences in the daily average trends of AQI during the four seasons,with higher AQI in winter compared to other seasons.There was difference of change trends between the concentration of nitrogen dioxide（NO₂）and the concentration of surface O₃.The concentration of NO₂in winter was significantly higher than that in other seasons.There was a certain correlation between surface O₃concentration and relative humidity of air.Using GAT to extract urban air quality monitoring network information,combined with BiLSTM with the ability to fit temporal data,the predictive performance of the proposed model and its variants was analyzed and compared.It was found that as the prediction time increased,the performance significantly improved.By utilizing the correlation coefficient and the nonlinear characteristics of the proposed model,it was found that AQI has a certain nonlinear relationship with surface O₃concentration,which has a significant impact on the air quality of Shanghai.By utilizing the CEEMDAN algorithm in modal decomposition related techniques,combined with sample entropy（SE）and KMeans algorithm,the complexity of time series was reduced,and the information extraction efficiency of deep learning models was improved.The prediction performance of the proposed model and its variants was analyzed and compared,and it was found that the proposed model had high accuracy in predicting single site surface O₃concentration.Through horizontal validation,it was found that BiLSTM had certain advantages in processing temporal data.