Physics-Guided Graph Convolutional Network for Modeling Car-Following Behaviors Under Distributional Shift

Recent works on data-driven car-following modeling have shown that a Graph Convolutional Network (GCN)-based car-following model (CFM) can outperform other data-driven models such as Long Short-term Memory (LSTM) networks. Inspired by this result, a new physics-guided GCN-based CFM is proposed in this paper. The model has been extended from the previous GCN-based CFM by integrating a vehicle platoon graph along with other refinements. Furthermore, the first attempt to develop physics-guided graph-learning-based CFM is made in this paper by adopting physics-guided machine learning (PGML) in different modeling aspects. The effectiveness of the proposed model was demonstrated in both in-distribution and out-of-distribution scenarios, yielding safe and accurate car-following behaviors. The experiment results showed that under the most severe distributional shift, the proposed model outperformed the best pure data-driven model by approximately 26.25% and a well-calibrated physics-based model by about 11.45% in terms of collision-penalized gap root mean squared error. Future research can focus on improving the robustness of our proposed model under various distributional shifts, leveraging the insights and suggestions given in the paper.