Research On Urban Air Quality Prediction Model Based On Factor Correlation

With the development of urbanization and industrialization,air pollution has become an increasingly serious problem,and air quality has gradually become a hot topic of concern for urban residents.From real-time data monitoring and publishing to air quality prediction,it is a major transition from stating the current situation of air environment to stating the trend of air environment changes.In real world situations,air quality data is not only temporal,but also dynamic in terms of the association between various factors,which brings technical challenges to air quality prediction.Most of the existing air quality prediction methods model air quality data from traditional time series perspective or spatio-temporal perspective,and fail to adequately consider the influence of the correlation between various factors on the prediction.To address the above problems,this thesis models the correlation between air quality factors and further considers the dynamics of the correlation of factors,and proposes two single-site based air quality prediction models.The main research contents include:(1)Research and experiments on the application of traditional time series models and deep learning-based models to air quality prediction tasks.Experimental results on two publicly available datasets show that the temporal convolution model based on dilated causal convolution can capture the long-term dependence of time series and obtain more accurate prediction results;the spatio-temporal model based on dynamic graph convolution can significantly reduce the prediction error by modeling the spatial correlation of dynamics among the factors,which provides a basis for subsequent research.(2)To address the correlation of air quality factors,this thesis proposes the air quality prediction model C2-Guard based on factor correlation.The factor correlation module based on multiple correlation correction is constructed in this model.Firstly,we integrate the temporal dimension information of each factor through pooling operation,then we construct the factor correlation relationship through nonlinear full connection layer to obtain the correlation vector,and finally the correlation vector is multiplied with the original factor matrix to map the correlation weights into the factor matrix and realize the learning of the factor correlation.By superimposing the above operations,multiple corrections of correlation are then realized.Compared with the baseline model,the RMSE error values of C2-Guard are reduced by 0.78%and 2.99%on the two real datasets,respectively.(3)To address the fact that the correlation of air quality factors changes over time,this thesis proposes an air quality prediction model C2-Guard~+that incorporating dynamic correlations.The model incorporates dynamic graph generation and dynamic graph convolution on the basis of C2-Guard to model the factor correlation at each moment in the historical series,so as to capture the dynamics of factor correlation more accurately in prediction.The comparison results with the baseline model show that the RMSE error values of C2-Guard~+are reduced by 2.27%and 4.03%on the two real datasets,respectively.