Traffic Speed Prediction Based on Dynamic Graph Convolutional Networks with Multi-Head Attention Mechanism

Accurate traffic speed prediction is crucial for accelerating smart transportation infrastructure and promoting the development of intelligent transportation systems. However, traffic networks exhibit complex spatial structures and dynamically random temporal characteristics, and many existing graph convolutional network models overlook the dynamic spatiotemporal changes in traffic flow data, failing to accurately capture their hidden correlations. To further explore the dynamic spatiotemporal features hidden in the data and improve prediction accuracy, this paper proposes a traffic speed prediction model called ST-MAGCN, based on a multi-head attention mechanism and dynamic graph convolutional networks. The model combines graph convolutional networks to capture spatial features of the traffic network while introducing a temporal convolutional network with a gated mechanism and dilated causal convolution to capture temporal features across multiple time domains. Multi-head attention mechanisms are integrated between spatiotemporal convolutional blocks to enhance the representation of spatiotemporal correlations, and residual connections are used to improve the model's generalization ability. The proposed method was tested on the PeMSD7 dataset for 15, 30, and 45-minute traffic speed predictions. Using three evaluation methods, including root mean square error, the experimental results demonstrate that the ST-MAGCN model excels in feature extraction and prediction performance for traffic speed sequences compared to most existing benchmark models.