Traffic Flow Prediction Model Based On Graph Learning And Attention Mechanism

In recent years,intelligent transport systems have been developed around the world to manage traffic effectively.Traffic flow forecasting can assist traffic managers in preparing for possible traffic jams in advance,and is important for improving the overall operational efficiency of urban road networks and improving people’s travel experience.However,the complex,dynamic and temporal correlation of the traffic flow data itself poses a great difficulty to traffic flow forecasting.The various traffic flow prediction algorithms that have been developed are also each inadequate in mining their temporal and spatial characteristics.To address these problems,this project proposes to investigate two types of traffic flow prediction methods based on attention mechanisms and graph convolutional neural networks,with the objective of mining the spatio-temporal correlations of traffic flows in real road networks.At the same time,to address the shortcomings and deficiencies of the existing models combining graph convolutional neural networks and attention machine mechanisms,variants of graph convolutional neural networks and attention mechanisms are utilised to further improve the performance of the models.The paper combines efficient channel attention(ECA)with graph convolutional neural networks(GCN)on the one hand,and builds an ECA-ASTGCN-based traffic prediction model.The ability of the attention mechanism weight assignment is enhanced to better capture the potential spatio-temporal correlation of detection points in real traffic networks.On the other hand an adaptive graph learning algorithm called AdapGL is used to optimise the adjacency matrix of the input patterns.Faced with the uncertainty of the graph structure in real traffic networks,the algorithm can automatically extract the implicit correlations between nodes by adaptively capturing the hidden spatial correlations between nodes with a new parametric graph learning module.Finally,the performance of the above two traffic flow prediction models is evaluated experimentally on a real traffic dataset.Compared with some classical baseline models,the traffic flow prediction models proposed in this paper both achieve good prediction performance.