Xue, Song (Monash University), Vu, Hai L. (Monash University)

Mixed-Granularity Physics-Informed Deep Learning for Traffic State Estimation

Scheduled for presentation during the Regular Session "S08b-Intelligent Modeling and Prediction of Traffic Dynamics" (WE-EA-T8), Wednesday, November 19, 2025, 14:10−14:30, Coolangata 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 19, 2025

Abstract

This study presents a mixed-granularity physics-informed deep learning (PIDL) framework for traffic state estimation (TSE), integrating microscopic car-following behavior and macroscopic density prediction. Unlike traditional PIDL models that rely on macroscopic flow assumptions or pre-defined fundamental diagrams, the proposed method uses a differentiable Intelligent Driver Model (IDM) to simulate vehicle-level interactions. These are softly aggregated via kernel-weighted assignment to produce grid-level traffic density. To ensure physical consistency, the IDM parameters are trained end-to-end using sparse dataset. The framework is validated using the NGSIM US-101 dataset under loop-detector scenarios. The estimated spatiotemporal traffic states are compared with ground-truth values using L^2 relative error. Results show that our model outperforms baseline approaches, particularly under sensor-sparse conditions. The proposed method effectively bridges microscopic dynamics and macro-scale estimation without requiring equilibrium traffic assumptions.