A Novel Graph Neural Network Method for Traffic State Estimation with Directional Wave Awareness.

Area of Science:

• Intelligent Transportation Systems
• Graph Neural Networks
• Traffic Flow Theory

Background:

• Traffic state estimation (TSE) is vital for managing and controlling intelligent transportation systems.
• Accurate TSE requires understanding complex, time-delayed correlations within road networks.

Purpose of the Study:

• To propose a novel physics-guided graph neural network for enhanced traffic state estimation.
• To integrate traffic flow theory into a graph neural network framework for more accurate predictions.

Main Methods:

• Constructed wave-informed anisotropic temporal graphs and merged them with spatial graphs for a unified spatiotemporal structure.
• Designed a four-layer diffusion graph convolutional network with squeeze-and-excitation attention for dynamic directional correlations.
• Incorporated the fundamental diagram equation into the loss function to ensure physically consistent estimations.

Main Results:

• The proposed physics-guided graph neural network demonstrated higher accuracy compared to benchmark methods on a real-world highway dataset.
• The model effectively captured complex traffic dynamics and time-delayed correlations across the road network.

Conclusions:

• The novel physics-guided graph neural network is effective for traffic state estimation.
• Integrating traffic flow theory improves the physical consistency and accuracy of TSE models.