Research On Urban Traffic Flow Prediction Method Based On Transformer And GNN

Traffic flow prediction is an important research direction in the fields of smart cities and intelligent transportation,providing crucial guidance for urban traffic management,urban planning,and travel prediction.In the field of traffic flow prediction research,how to better capture the Spatio-temporal information contained in traffic flow data and more accurately predict statistical indicators is the focus of researchers.Traditional machine learning methods have difficulty predicting traffic flow data with complex Spatio-temporal information,so researchers are increasingly focusing on how to use suitable deep learning models to train traffic flow data of specific formats.this thesis thoroughly studies how to design high-performance traffic prediction models to address traffic flow prediction tasks under different traffic flow formats.1.To address the shortcomings of traditional Transformer models in fitting Spatiotemporal traffic flow data,this thesis proposes a novel Transformer-based traffic prediction model called TST-Trans,which is combined with grid-based traffic flow prediction tasks.TST-Trans uses ST-embedding to obtain the spatial dependency relationship of gridbased traffic flow data.The Spatial-block module reduces the number of model parameters by nearly 86%.By adjusting the position encoding position,the PE can better assist the model in predicting efficiency,which is improved by nearly 2%,and reduces the performance fluctuations.By adopting the MSA mechanism and ST-embedding method,the Transformer-block module can more comprehensively consider static information,capture changing Spatio-temporal correlations,and improve the model’s prediction accuracy by nearly 7%.2.To address the shortcomings of traditional graph convolutional models that depend on real spatial topology relationships,this thesis proposes a data-driven approach based on multiple correlations and KNN for constructing adjacency matrices to extend spatial relationships.In addition,this thesis constructs a novel ST-GTNN model that includes spatial graph attention and graph convolution modules,as well as an Attention-2D-TCN module,which respectively combines spatial and temporal attention mechanisms.From the perspective of reducing information loss,the models are designed with concatenation dimensions and residual encoding to minimize information loss caused by dimension conversion when transmitting Spatio-temporal data features internally in the model.Compared with existing methods,the prediction accuracy of the ST-GTNN model has been improved by 4.8%.3.Based on the two previous work components,the integrated predictive algorithm and big data visualization platform,named UTFPS,were designed and implemented in this thesis.The system consists of three main parts: user interface,business middle platform,and data backend.The platform includes traffic flow query function modules,model training function modules,and traffic flow prediction function modules.Additionally,this thesis provides a detailed process design for each functional module of UTFPS and implements the UTFPS system based on the functional design and current mainstream technologies in the industry.The final UTFPS has features such as small size and lightweight,more complete functions,strong system framework scalability,and a user-friendly and convenient interactive interface.