1. Introduction
Traffic congestion prediction models that treat intersections independently fail to capture the propagation of congestion along connected corridors. Graph-based learning offers a natural way to encode road network topology alongside temporal traffic patterns.
2. Methodology
Loop-detector readings from 340 signalised intersections over six months were assembled into a graph where nodes represent intersections and edges represent road segments weighted by distance and historical correlation. A two-layer graph attention network extracted spatial features at each timestep, which were passed to a GRU to model temporal dependencies, trained with a rolling-window scheme.
3. Results
The model achieved MAPE of 9.8 percent, 11.4 percent and 13.9 percent at 15, 30 and 60-minute horizons respectively, consistently outperforming an LSTM-only baseline and a historical-average baseline across all horizons and across peak and off-peak periods.
4. Conclusion
Encoding road network topology through graph attention meaningfully improves short-horizon congestion forecasts over sequence-only models. Future work will integrate weather and event data as additional graph features.
References
[1] Yu B. et al., Spatio-temporal graph convolutional networks for traffic forecasting, IJCAI, 2018. [2] Velickovic P. et al., Graph attention networks, ICLR, 2018. [3] Li Y. et al., Diffusion convolutional recurrent neural network, ICLR, 2018.