Graph Neural Networks for Traffic Flow Prediction

Abstract

Network reconfigurations, including closures, new road segments, capacity changes, and speed changes, can redistribute traffic flows. Predicting this redistribution is important for rapid assessment of infrastructure changes and operational interventions. Conventional Stochastic User Equilibrium (SUE) methods can compute the resulting equilibrium state, but they require Origin-Destination (OD) demand and must be solved again for each reconfigured network. This thesis studies a more constrained setting: predicting post-edit equilibrium flows from the pre-edit network, old flows, and the reconfigured network, without OD demand as input.

A physics-informed graph learning framework, called ST-PINN GatedGCN, is developed for this task. First, edge alignment maps old-flow information onto the edge set of the reconfigured graph. The GatedGCN then learns flow propagation and topology changes through graph message passing. The training objective combines edge-flow regression with flow-conservation regularization. It fits SUE-generated labels while reducing node-level physical inconsistency. Inference is completely OD free as the demand is used only to generate labels.

Experiments use the Sioux Falls and EMA networks, with 10,000 network-pair samples for each network. The model obtains 12.92% WMAPE on Sioux Falls and 7.84% WMAPE on EMA, which is lower than the tested baselines. Retained edges are predicted more accurately than newly added edges, since new edges have no historical flow and may introduce new route choices. Ablations show that old-flow information, residual physical injection, the global message channel, and recurrent pressure update affect error and conservation behavior. Inference is about 197 times faster than SUE on Sioux Falls and about 1896 times faster on EMA.

These results suggest that old flows and graph structure contain useful information for fast scenario screening under network reconfiguration. The remaining errors reflect the uncertainty caused by unobserved OD demand, especially for newly added links.