Online Demand Forecasting with Spatial-Temporal Graph Attention Networks: A Proof of Concept

Predicting future demand from the current state of the network is an unsolved challenge in dynamic traffic management systems which show the current state of traffic as well as demand and supply forecasts for simulations of response plans. In the context of the TANGENT H2020 project, data-driven methodologies are developed focusing on the real-time demand prediction problem.This paper presents a spatial-temporal graph attention network (ST-GAT) to model: 1) the geometry of the network including centroids as demand generation points; 2) the temporal resolution of traffic counts.Preliminary experiments show promising outcomes when quasi-perfect synthetic data is used for training. Yet more research is needed to fully cater for the challenging requirements of the demand prediction task in real-time settings.