Spatial‑Temporal Graph Diffusion Network for City Traffic Flow Forecasting

City traffic flow forecasting is a critical task for intelligent transportation systems, aiming to predict traffic volumes across different urban regions to enable congestion control, traffic scheduling, and public safety. The presented work proposes a Spatial‑Temporal Graph Diffusion Network (ST‑GDN) that leverages hierarchical graph neural networks and multi‑granularity attention mechanisms to model both temporal dynamics and global spatial relationships.

Model Design

For temporal modeling, the approach encodes traffic sequences at hourly, daily, and weekly granularities using self‑attention modules that learn relationships among time steps and fuse the resulting representations. For spatial context, a Graph Attention Network updates region embeddings by propagating information across a constructed region graph that includes both geographically adjacent nodes and nodes with high attention‑based similarity, thereby capturing global spatial dependencies.

Experiments

The model is evaluated on several public datasets, including New York City taxi and bike flows and Beijing taxi flows. Using RMSE and MAPE as evaluation metrics, ST‑GDN consistently outperforms existing traffic forecasting baselines. Additional ablation studies confirm the effectiveness of the self‑attention encoder, the graph‑attention based spatial updater, and the diffusion module. Experiments also explore the impact of different temporal granularities (hour‑only, hour‑day, hour‑week, hour‑day‑week) and hyper‑parameters, showing that multi‑scale temporal representation and appropriate model capacity further improve prediction accuracy.

Conclusion

The proposed ST‑GDN demonstrates that hierarchical graph diffusion combined with multi‑scale temporal attention can accurately capture complex spatio‑temporal patterns in urban traffic, leading to superior forecasting performance and providing valuable insights for smart city applications.