DSTT-HF: Dynamic Spatio-Temporal Transformer with Hierarchical Fusion for Traffic Flow Prediction

Abstract Accurate traffic flow prediction plays a vital supporting role in urban traffic management and emergency response. Existing spatio-temporal prediction models suffer from limited capability in capturing time-varying topological dependencies and fine-grained multi-granularity feature aggregation, which degrades prediction performance especially under unexpected traffic fluctuations. To address these limitations, this paper proposes a Dynamic Spatio-Temporal Transformer with Hierarchical Fusion (DSTT-HF). The model constructs time-varying dynamic adjacency matrices from real-time node features and periodic traffic trends to mitigate the defects of static topology. A temporal attention module incorporating first-order difference features is designed to capture both long-term periodic patterns and short-term traffic variations. Moreover, a multi-scale hierarchical fusion module with learnable weighted summation is developed to enhance model robustness for multi-step prediction. Experiments conducted on public PEMS04 and PEMS08 datasets show that DSTT-HF achieves consistent performance improvement over several state-of-the-art baselines. Ablation experiments further validate the effectiveness of core components including feature interaction enhancement, dynamic graph generation and multi-scale fusion. Quantitatively, DSTT-HF achieves MAE of 17.80 and 13.57 on PEMS04 and PEMS08 for the one-step horizon, representing reductions of 8.2% and 9.8% over STGCN, respectively, with all metrics reported as mean ± std over 10 independent runs. The proposed DSTT-HF can effectively model complex dynamic spatio-temporal correlations and provides a feasible solution for short-term traffic flow forecasting within a three-step horizon