DASTFormer: Delay-aware adaptive sparse transformer for traffic flow prediction

With the acceleration of urbanization, traffic flow prediction has become a core problem in intelligent transportation systems. However, existing models generally overlook delay states, making it difficult to capture the dynamic evolution of traffic networks, and they also fall short in modeling the nonlinear relationships and spatial dependencies among nodes. To address these issues, this paper proposes a traffic flow prediction model named Delay-aware Adaptive Sparse Transformer (DASTFormer). Specifically, the model first employs a delay-enhanced feature fusion layer to capture propagation delays among nodes. Then, it dynamically updates the correlation matrix via mutual information and integrates it into the self-attention mechanism, thereby improving the modeling capability for complex spatiotemporal dependencies. Finally, an adaptive sparse spatial attention mechanism is introduced to focus on key nodes, enhancing prediction accuracy. Experiments conducted on two real-world traffic datasets demonstrate that DASTFormer outperforms mainstream Transformer-based baselines. Compared with the delay modeling method of the same type, PDFormer, DASTFormer reduces MAE and RMSE by 2.283% and 1.489%, respectively, on the PeMS08 dataset, validating the effectiveness of the proposed architecture.