Short-term Traffic Flow Prediction Based On Graph Neural Network With Attention Mechanism

As the level of modern urban construction continues to improve,the number of cars has also increased,leading to increasingly severe traffic congestion problems.Developing intelligent transportation systems is an important means of alleviating urban traffic pressure,improving traffic efficiency and safety.Among them,traffic flow prediction,as one of the core technologies of intelligent transportation systems,can provide real-time road traffic flow forecasting,provide convenience and guidance for public travel,and provide effective support for road planning and traffic control by traffic management departments.The complexity of traffic flow data is reflected in its dynamics and strong spatiotemporal correlation,making it very difficult to obtain accurate prediction results in traffic flow prediction tasks.However,most existing prediction models only consider the time dimension of traffic flow data,ignoring the spatial dimension of traffic flow data.Therefore,building more accurate and reliable traffic flow prediction models requires consideration of both time and space dimensions,and modeling in conjunction with the topological structure of the traffic network.The main research content of this article is as follows:(1)Based on the urban traffic network,traffic flow matrix and adjacency matrix are constructed.Graph Attention Network(GAT)is used to capture the spatial correlation between traffic flow data of observation points,and residual connection is used to stack multiple layers of GAT to capture deeper spatial relationships between data.Attention mechanism is used to allocate weights to historical data.Bidirectional Gated Recurrent Unit(GRU)is used for time feature extraction of traffic flow data.Finally,the final prediction result is output through the fully connected layer.The results show that the Res GAT-ABi GRU combination model has higher prediction accuracy and its performance is better than baseline models such as SVM,GRU,and GCN-GRU models.(2)After preprocessing the original data,the traffic flow matrix and adjacency matrix are obtained.The original road flow information is clustered into multiple classes using the Kmeans clustering algorithm,and the similarity between nodes is calculated.According to whether two nodes are in the same class,weight allocation is performed to construct a weighted adjacency matrix.Pearson correlation analysis is used to construct a Pearson adjacency matrix based on the road flow data.Integrating different adjacency matrices can capture hidden interactions between nodes.Using the local clustering coefficient to replace the traditional unit matrix selfloop weighting can better reflect the connection relationship between nodes.Multiple adjacency matrices are used together with the traffic flow matrix for GNN spatial feature extraction.Bidirectional GRU is used to extract time correlation between traffic flow data,and selfattention mechanism is used to capture mutual dependencies between data.The final prediction result is obtained through residual connection and fully connected layers.The results show that the combination model has higher prediction accuracy than other baseline models and the GCNGRU model.