Short-term Flow Forecasting with Spatio-temporal GCN

Flow is temporally auto-correlated and spatially propagated along the network. Classical time-series models ignore spatial structure; graph convolution explicitly encodes topology.

Leverage network topology and historical flow to forecast 15-min link flows, providing lead time for real-time management.

Built a graph with links as nodes and adjacency as edges; node features are the past 12 5-min flows; combined GCN + GRU to capture spatio-temporal dependence, trained on an 80/10/10 split with early stopping.

Spatial structure markedly improves short-term accuracy, but long-horizon forecasts remain sensitive to event-driven perturbation. Incorporating external features (weather, events) is the next step.