Dynamic Origin-Destination Matrix Prediction with Line Graph Neural Networks and Kalman Filter

TL;DR

This work tackles dynamic O-D demand prediction by jointly modeling spatial-temporal dependencies on a road network. It introduces Fusion Line Graph Convolutional Networks (FL-GCN) that combine line-graph based link-to-node convolutions with historical O-D patterns, and couples them with a deviation-based Kalman filter to refine estimates in real time, using a mixing weight to balance contributions. The approach, validated on the New Jersey Turnpike dataset, achieves superior accuracy over baselines and demonstrates robustness across prediction horizons. The findings highlight the importance of integrating deep spatial-temporal learning with probabilistic filtering, and show that historical O-D information often carries more weight than instantaneous link flows in the FL-GCN architecture. The work suggests extensions to missing data, recurrent line-graph models, and richer traffic information for broader applicability.

Abstract

Modern intelligent transportation systems provide data that allow real-time dynamic demand prediction, which is essential for planning and operations. The main challenge of prediction of dynamic Origin-Destination (O-D) demand matrices is that demands cannot be directly measured by traffic sensors; instead, they have to be inferred from aggregate traffic flow data on traffic links. Specifically, spatial correlation, congestion and time dependent factors need to be considered in general transportation networks. In this paper we propose a novel O-D prediction framework combining heterogeneous prediction in graph neural networks and Kalman filter to recognize spatial and temporal patterns simultaneously. The underlying road network topology is converted into a corresponding line graph in the newly designed Fusion Line Graph Convolutional Networks (FL-GCNs), which provide a general framework of predicting spatial-temporal O-D flows from link information. Data from New Jersey Turnpike network are used to evaluate the proposed model. The results show that our proposed approach yields the best performance under various prediction scenarios. In addition, the advantage of combining deep neural networks and Kalman filter is demonstrated.

Figures (6)

• Figure 1: The structure of our model.
• Figure 2: Line graph transformation and graph convolution in links and nodes
• Figure 3: Section of NJ turnpike.
• Figure 4: Model comparison under different prediction steps.
• Figure 5: MAE, RMSE and RMSN with ratio of neural networks.