A traffic speed prediction algorithm for dynamic spatio-temporal graph convolutional networks based on attention mechanism

In intelligent transportation systems (ITS), accurate traffic speed prediction is critical for the timely detection of congestion, traffic flow optimization, enhanced travel experiences, and providing decision-making support for traffic management authorities. However, current deep learning-based prediction methods face limitations, such as the inability to fully capture complex spatio-temporal dependencies, and their prediction accuracy needs improvement. To address these challenges, this paper proposes a dynamic spatio-temporal graph convolutional network (Att-DSTGCN) incorporating an attention mechanism. This algorithm utilizes a temporal convolutional network and dynamic graph convolution to extract spatio-temporal dependencies, constructs a spatial graph generation function, and extends it to dynamic graph convolution for modeling node-to-node changes. Furthermore, traffic flow data are integrated through feature fusion to enhance prediction accuracy. The attention mechanism enables the model to focus on critical spatio-temporal correlations, allowing for more accurate aggregation of neighbor information. Experimental results on three public datasets demonstrate that Att-DSTGCN achieves superior performance in traffic speed prediction compared to existing baseline models, particularly in complex scenarios.