uTransformer: unified spatial-temporal transformer with external factors for traffic flow forecasting

Traffic flow forecasting is essential for applications such as route planning and vehicle communication. However, dynamic traffic patterns, diverse spatial-temporal dependencies, and external factors pose significant challenges. Existing methods often model temporal, spatial, and external factors separately, failing to capture their joint effects across different spatial-temporal scales and external influences. To address these, we propose uTransformer, a unified transformer model that captures joint spatial-temporal dependencies and integrates external factors like weather and holidays. Unlike existing models, uTransformer constructs a three-dimensional spatial-temporal correlation map (STC-MAP) to model spatial-temporal interactions across various time scales, effectively capturing long-range dependencies. Additionally, it features a scalable encoding scheme that flexibly incorporates multiple types and quantities of external factors. Experimental results on NYC-BIKE and NYC-TAXI datasets show uTransformer achieves high accuracy, with RMSE improved by 10.97 \(\%\) $\%$ and MAPE by 16.06 \(\%\) $\%$ , outperforming baseline models and enhancing prediction reliability.