A spatiotemporal correlation model for traffic flow based on dynamic graph attention and multi-scale feature fusion

With the accelerated urbanization, traditional traffic management is confronted with severe challenges, as existing models fail to fully capture the spatiotemporal and nonlinear characteristics of traffic flow. To tackle this problem, this paper proposes a Multi-Scale Feature Fusion Graph Convolutional Neural Network (MSFGCN), which integrates multi-scale feature fusion, locally random feature augmentation, dynamic graph attention and an adaptive auxiliary module. The model extracts fine-grained spatial information via maximum mutual information coefficient matrices and adaptive matrix mining, and effectively captures spatiotemporal correlations by combining global spatial attention, temporal attention and periodic information embedding. Moreover, it adopts an interaction-aware attention network with dynamic spatiotemporal embedding, coupled collaborative network prediction and dynamic random graph attention modules, while optimizing the fusion adjacency matrix. Experimental results show that MSFGCN achieves the lowest prediction errors (PeMSD3: MAE 8.08, RMSE 11.52, MAPE 12.86%; PeMSD7: MAE 8.87, RMSE 12.86, MAPE 14.24%) and outperforms all baseline models. It also exhibits low computational complexity, generalization and robustness, especially during weekday peak hours, with relatively low computational complexity, providing a reliable theoretical basis for efficient traffic management and smart city construction.