Dynamic origin-destination estimation using link counts : a physics-informed deep learning approach

Abstract

This paper proposes a novel Physics-Informed Neural Network (PINN) approach for solving the Dynamic Origin-Destination Estimation (DODE) problem. The method combines physical constraints and deep learning techniques, utilizing traffic flow observations and road network structure information to accurately estimate dynamic OD matrices using link counts. The main contributions of this study include: (1) proposing a novel spatiotemporal neural network structure (CNN & Line Graph Sample and AggreGatE (L-GSAGE)) that achieves high-accuracy OD estimation; (2) introducing the Physics-Informed Deep Learning (PIDL) + Dynamic Traffic Assignment Learner (DTAL) OD estimation method based on the proposed algorithm, which avoids directly encoding complex terms in the Dynamic Traffic Assignment (DTA) process by incorporating a machine learning surrogate into the model-driven component, thereby improving estimation accuracy; (3) conducting comprehensive experimental evaluations on two different networks (Sioux Falls and Grid networks), validating the advantages of the proposed method in terms of estimation accuracy and data efficiency compared to benchmark methods. The experimental results demonstrate that the PINN method performs exceptionally well under various demand levels and data missing scenarios, proving its robustness and effectiveness in practical applications. This study provides a novel and efficient approach for solving the dynamic OD estimation problem in transportation networks, which has significant implications for traffic planning and management.

Date of Award	2024
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Hong Kam LO (Supervisor)