Skip to content
Open access

Physics-Informed State Projection With Trajectory-Derived Delayed Propagation for Scalable Traffic Forecasting

2026 · IEEE Access · Vol 14, pp. 160076-160084 · 0 citations · 25 references

Abstract

Large-scale expressway forecasting requires a model to represent how traffic states propagate between road segments with heterogeneous travel delays. Existing temporal, graph-based, and attention-based predictors usually learn these dependencies implicitly from aggregated flow histories, which limits direct use of interpretable propagation strength and travel-delay evidence available in reconstructed trajectories. This paper proposes Physics-Informed State Projection (PISP), which extracts an edge-level propagation operator and an eight-bin discrete delay distribution from training trajectories and calibrates the fixed operator using training-only low-rank gradient corrections. Experiments cover 10,552 road-segment nodes and 3,984 samples from 48 complete dates, with eight 15-min observations used to predict the next six intervals under a chronological 70%/10%/20% split. Across three random seeds, PISP achieves an MAE of <inline-formula> <tex-math notation="LaTeX">$8.198\pm 0.040$ </tex-math></inline-formula>, an RMSE of <inline-formula> <tex-math notation="LaTeX">$19.771\pm 0.099$ </tex-math></inline-formula>, and a weighted absolute percentage error (WAPE) of <inline-formula> <tex-math notation="LaTeX">$19.518\%\pm 0.096\%$ </tex-math></inline-formula>. Its mean WAPE is the best among the evaluated AR, linear, recurrent, kernel-regression, graph-convolutional, and Transformer baselines, improving on the strongest comparison value by 0.10% while exhibiting lower cross-seed variation. The trajectory-derived propagation structure offers an interpretable and stable basis for short-term network forecasting and supports large-scale expressway monitoring and operational decision making.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.