A Physics-Informed Generative Car-Following Model for Connected Autonomous Vehicles.

Area of Science:

• Traffic flow dynamics
• Artificial intelligence in transportation

Background:

• Accurate car-following models are crucial for traffic safety and efficiency.
• Existing models often struggle with the complexity of mixed traffic flow (human-driven and autonomous vehicles).
• Hybrid approaches combining physics-based and data-driven methods show promise but often require explicit parameter tuning.

Purpose of the Study:

• To introduce a novel hybrid car-following model, the physics-informed conditional generative adversarial network (PICGAN).
• To enhance multi-step car-following modeling capabilities in mixed traffic flow scenarios.
• To develop a model that integrates physics-based principles with deep learning without explicit weighting parameters.

Main Methods:

• Developed the physics-informed conditional generative adversarial network (PICGAN) model.
• Utilized the inherent structure of Generative Adversarial Networks (GANs) to fuse physics-based and deep learning components.
• Validated the model using the NGSIM I-80 dataset for car-following behavior analysis.

Main Results:

• PICGAN demonstrated superior trajectory reproduction compared to conventional models.
• The model significantly improved stability and efficiency in mixed traffic flow simulations.
• Case studies confirmed the effectiveness and reliability of the proposed framework.

Conclusions:

• The PICGAN model offers a robust and effective solution for multi-step car-following prediction in mixed traffic.
• It eliminates the need for explicit weighting parameters, simplifying the integration of physics and data-driven approaches.
• PICGAN provides a strong foundation for developing advanced longitudinal control strategies for connected autonomous vehicles (CAVs).