Study On Spatio-temporal Correlation Prediction Of Pollutants In Waiting Area Of Subway Station Platform

As an important way for people to travel and one of the core parts of urban traffic,due to the structure,ventilation equipment,people,and other factors,some subway stations are prone to serious air pollution and harm people’s travel safety and health.Therefore,it is of great significance to accurately obtain the information on pollutant concentration in subway stations and establish a prediction model with excellent performance.At present,the traditional method is mainly based on the time-series data of the current measurement point.The influence of the data of other measurement points on the change of pollutant concentration at the target point is ignored.In this thesis,combined with the basic principle of spatiotemporal modeling,a new spatiotemporal correlation prediction model of pollutants in the waiting area of the subway station platform is designed.In addition,to realize the practicability of the research results,this thesis also built a pollutant prediction visualization platform and embedded the proposed model into the visualization platform.It not only achieves theoretical innovation in the field of air pollution prediction,but also has important practical significance for improving the air quality of subway stations.The main work of the thesis is as follows:(1)Different from traditional statistical models,machine learning models,and neural network models,this thesis proposed GAT-SRU and GAT-TCN networks as the main predictors.On the one hand,SRU and TCN fully combine the advantages of CNN and RNN,which effectively improves the training effect of the model and achieves better PM2.5prediction results.On the other hand,the GAT algorithm based on graph structure and attention mechanism can fully analyze the spatiotemporal correlation between different subway stations and fully extract feature data.Experimental results show that the GAT algorithm can improve the predictive performance by more than 3%.(2)This thesis used the deep reinforcement learning as the main ensemble decision algorithm to effectively combine GAT-TCN and GAT-SRU.The proposed ensemble learning model further improved the adaptability and stability of the spatiotemporal correlation prediction model of pollutants in subway stations.Experimental results show that ensemble learning can improve the prediction accuracy of single predictors by more than 5%.Based on the characteristics of each model component,the GAT-SRU-TCN-DDPG model was proposed.Based on the comparison of the GAT-SRU-TCN-DDPG model with benchmark models and advanced models in these fields,it was proved that the proposed model has excellent performance.(3)Based on the B-S model architecture,a set of pollutant concentration prediction and visualization systems for subway stations was developed.The GAT-SRU-TCN-DDPG model was embedded into the system to improve the practical application value of the model.After testing and experiment,it was proved that the software can run stably and meet the basic functions required.72 Figures,14 Tables,151 References...