Traffic flow prediction is a fundamental component of intelligent transportation systems and plays a critical role in traffic management and autonomous driving. However, accurately modeling highway traffic remains challenging due to dynamic congestion propagation, lane-level heterogeneity, and non-recurrent traffic events. To address these challenges, this paper proposes an improved attention-mechanism spatio-temporal graph convolutional network, termed AMSGCN, for highway traffic flow prediction. AMSGCN introduces an adaptive adjacency matrix learning mechanism to overcome the limitations of static graphs and capture time-varying spatial correlations and congestion propagation paths. A hierarchical multi-scale spatial attention mechanism is further designed to jointly model local congestion diffusion and long-range bottleneck effects, enabling an adaptive spatial receptive field under congested conditions. To enhance temporal modeling, a gating-based fusion strategy dynamically balances periodic patterns and recent observations, allowing effective prediction under both regular and abnormal traffic scenarios. In addition, direction-aware encoding is incorporated to suppress interference from opposite-direction lanes, which is essential for directional highway traffic systems. Extensive experiments on multiple benchmark datasets, including PeMS and PEMSF, demonstrate the effectiveness and robustness of AMSGCN. In particular, on the I-24 MOTION dataset, AMSGCN achieves an RMSE reduction of 11.0% compared to ASTGCN and 17.4% relative to the strongest STGCN baseline. Ablation studies further confirm that dynamic and multi-scale spatial attention provides the primary performance gains, while temporal gating and direction-aware modeling offer complementary improvements. These results indicate that AMSGCN is a robust and effective solution for highway traffic flow prediction.
Zhang et al. (Mon,) studied this question.