Traffic flow matrix-based graph neural network with attention mechanism for traffic flow prediction

Traffic flow forecasting is of great importance in intelligent transportation systems for congestion mitigation and intelligent traffic management. Most of the existing methods depend on deep learning to extract the spatial–temporal correlation of traffic nodes but ignore the traffic flow characteristics. In this paper, we design three traffic congestion indexes to reflect the operational status of nodes based on traffic flow theory and design a traffic flow matrix to better represent the relationship between nodes. We also design a novel graph convolution network with attention mechanisms called TFM-GCAM to better capture the spatial–temporal features and dynamic characteristics of nodes. A novel Fusion Attention mechanism is proposed to effectively fuse the dynamic characteristics and the spatial–temporal features for improvement. Experiments and ablation studies on the public dataset show the superiority of TFM-GCAM. We also discuss it with our previous works for a better understanding. Our research proposes to better integrate traffic flow theory into deep learning models and to better combine the respective strengths of attention mechanisms and graph neural networks for more effective traffic flow prediction.