A spatio-temporal graph transformer for multi-horizon urban traffic speed forecasting

Precise multi-horizon traffic-speed forecasting is challenging because road sensors are spatially connected and traffic patterns vary over time. This study presents a Spatio-Temporal Graph Transformer (ST-GT) that combines distance-based graph aggregation with Transformer self-attention to model spatial and temporal dependencies, respectively. A single non-autoregressive model uses 12 historical five-minute observations to predict the next 12 steps simultaneously; performance is reported at 15, 30, and 60 min on the METR-LA dataset. Under the common masked evaluation protocol used in this study, ST-GT achieved MAE values of 3.280, 3.651, and 4.094 mph; RMSE values of 5.855, 6.710, and 7.658 mph; and MAPE values of 8.374%, 9.665%, and 11.406% at the three horizons, respectively. In our experimental setting, ST-GT produced lower 60-minute errors than the classical baselines and the Vanilla LSTM, and it also outperformed the added DCRNN-lite, Graph WaveNet-lite, and GMAN-lite implementations in MAE across all three horizons. Ablation experiments showed that removing graph aggregation, temporal self-attention, time-of-day features, or multi-head attention increased prediction error. The results support the usefulness of jointly modeling spatial structure, temporal dependencies, and periodic context, while the single-seed design and fixed graph topology limit broader claims of statistical or state-of-the-art superiority.