Research On Analysis And Prediction Of Urban Air Pollution Network Based On Intensive Data

The core task of the fine management of the urban environment is to accurately and timely discover the key pollution areas of air quality,which is of great significance for improving the accuracy and efficiency of prevention and control.How to accurately dig out the distribution characteristics of air pollution,and predict air pollution concentration,has become the realistic question that managers needs to solve urgently.However,due to the sparse air monitoring sites and data deficient,there are few spatial characteristics analyses and related research on air pollution in small-scale areas in urban areas.Based on the data collected by the intensive air quality monitoring stations deployed in Lanzhou,this research analyzes and studies the spatial network characteristics,transmission paths,and concentration prediction of PM_2.5pollution in urban areas.The details of the researches are listed as follows:（1）Construction of urban air pollutants spatial network and analysis of topological characteristics.Take the monitoring station as the node,and use the Pearson correlation coefficient of the PM_2.5concentration between the nodes in the time series as the connecting edge to construct the network.Utilize the network characteristics of degree centrality,closeness centrality,and betweenness centrality to analyze the characteristics of urban air pollution.The research results show that the network constructed by choosing 0.89 as the edge threshold is the optimal network.In the analysis of central characteristics,the top-ranked Chengguan District and Xigu District have the most serious pollution,and the lower-ranked Anning District and Qilihe District have lighter pollution.By selective removal and random removal of nodes in the network,it is verified that nodes with high centrality have a greater impact on network transmission efficiency.This shows that controlling the nodes with high centrality can improve the governance effect of urban PM_2.5pollution.The results of the analysis are basically the same with the Lanzhou special air pollution control area,indicating that the air pollution transmission model based on network analysis is reliable and feasible.（2）Research on Air Pollution Transmission Path Prediction Based on GraphSAGE.By designing the initialization node layer,message propagation layer,attribute-structure feature fusion layer,and prediction layer,a pollutant transmission path prediction model PM_2.5-link is constructed.The six main air pollutants(SO₂,NO₂,CO,O₃,PM₁₀,PM_2.5)and temperature,humidity,and wind speed are used as the initial attribute feature of the node,and all neighbors of the node are aggregated as the spatial structure feature of the node through GraphSAGE.Then concat the two features as the final feature vector of the node into the prediction layer for training and learning.The research results show that the vector that concat node attribute features and spatial structure features can improve the prediction accuracy of the model.In the precision evaluation metrics,the PM_2.5-link model is 4.84%higher than the graph convolutional neural network.In recall,F1-score and NDCG,the PM_2.5-link model is 3.07%,3.91%,and 3.48%higher than the GraphSAGE model.Besides,the area under the ROC curve of this model is all the largest in the baseline model,this indicates that the prediction performance of this model is the best.The proposal of this method provides a new idea for accurately predicting the transmission path of air pollution.（3）Research on PM_2.5concentration prediction based on graph attention network.The GAT-Cent model is constructed by designing the temporal feature extraction layer,the neighborhood pollution sensing layer,and the central information fusion layer to predict the concentration of PM_2.5in the next day.The Bi LSTM is used to extract the attribute feature of each node in the time series as the initial representation of the node.Use the graph attention mechanism to learn the attention weight between a node and its neighbor nodes.The multi-head attention mechanism is used to aggregate the embedding representation and the concentration value of the node to obtain the concentration prediction value.Incorporate the degree centrality,closeness centrality,and betweenness centrality of the node into the last layer of the prediction model to adjust the concentration prediction value.The research results show that the GAT-Cent model is higher than the FNN,GAT,GCN,and other classic models in the three evaluation metrics.Among them,this model is better than GAT in the three metrics of RMSE,MAE,and Correlation Coefficient.The model has been improved by 16.73%,14.9%,and 17.83%.In the ablation experiment,the prediction performance of this model is better than that of the model without Bi LSTM and the three centrality adjustment mechanisms.It proves the validity and applicability of time series characteristics and centrality information in PM_2.5concentration prediction.