A traffic flow forecasting model based on dynamic graph learning and temporally adaptive attention

Accurate traffic flow forecasting is essential for ensuring transportation safety and advancing intelligent transportation systems. Static graph–based methods fail to capture the dynamic characteristics of traffic networks, leading to limitations in joint spatiotemporal modeling and multi-step prediction tasks. To address these challenges, this study proposes a Dynamic Spatiotemporal Interaction Model (D-STIM) for traffic flow forecasting. The model comprises three core modules: Efficient Adaptive Spatiotemporal Learning (EASL), Progressive Interactive Learning (PIL), and Temporally Adaptive Attention (TAA). EASL leverages low-rank factorization to model dynamic graph structures, thereby reducing computational complexity and enhancing structural adaptability. PIL establishes bidirectional interaction through spatial-guided temporal aggregation and temporal-guided spatial aggregation, enabling deep spatiotemporal fusion. TAA integrates positional encoding and temporal bias into the attention mechanism to effectively mitigate information degradation in long-horizon forecasting. Extensive experiments on four real-world traffic datasets demonstrate that D-STIM consistently outperforms mainstream baselines in both prediction accuracy and computational efficiency. Moreover, the proposed model provides practical safety benefits by supporting congestion mitigation, reducing accident risks, and informing proactive traffic management strategies • D-STIM unifies dynamic graph learning and adaptive attention for traffic forecasting. • Adaptive adjacency learning enhances dynamic spatial modeling and efficiency. • Bidirectional spatiotemporal interaction improves expressive feature extraction. • Experiments on four datasets show D-STIM achieves superior forecasting accuracy.