STFDGCN: Spatial–Temporal Fusion Time Delayed Simulation Graph Convolutional Network for Traffic Flow Forecasting

Accurate prediction of traffic flow is fundamental to building intelligent transportation systems. However, most existing research methods fail to fully account for the potential spatiotemporal correlations, heterogeneity, and time delays present in traffic flow data. In addressing these concerns, this work introduces a novel Spatio-Temporal Fusion Delay Graph Convolutional Network (STFDGCN) model. The STFDGCN model features a Spatio-Temporal Delay Graph Convolution Module (STDGCM) that simulates time delays in traffic flow through displacement operations following local graph convolutions, and captures spatiotemporal heterogeneity by employing multiple STDGCMs. The inclusion of a self-attention mechanism allows the model to effectively capture global spatiotemporal correlations. Additionally, the model incorporates an Information Fusion Module (IFM) that integrates local information with global information captured by the self-attention mechanism, offering a more comprehensive analysis. Finally, the model’s comprehensive capability is further improved by stacking multiple STFDGCL layers. Extensive tests on three real-world datasets validate the usefulness of the proposed STFDGCN model by demonstrating its superior prediction accuracy compared to other baselines, highlighting the model’s robustness and its potential to enhance traffic flow forecasting in various practical scenarios.