Abstract:In complex traffic scenarios, traffic flow exhibits strong spatiotemporal dependencies, is significantly influenced by external factors, and contains long-range temporal dependencies that are difficult to capture. However, existing methods based solely on graph convolutional networks (GCNs) or Transformers are unable to address these issues simultaneously. Therefore, this study proposes a traffic flow prediction model that integrates a temporal graph convolutional network (TGCN) with a Transformer architecture. TGCN layers based on dual dynamic adjacency matrices are designed for the model to adaptively update spatial relationships among nodes, overcoming the limitations of static topological structures in complex dynamic environments. The feature-wise linear modulation (FiLM) mechanism is adopted to achieve conditional fusion of multimodal features, thus enhancing the robustness of the model in non-stationary traffic flows. Finally, a Transformer architecture is employed, where the self-attention mechanism facilitates multi-scale modeling. Experimental results on PeMS04 and PeMS08 show that, for different prediction horizons, the proposed model maintains stable performance in noisy traffic scenarios and outperforms several mainstream baseline methods, demonstrating its strong generalization ability.