GAT-Transformer-IDM: A Physics-Informed Spatio-Temporal Model for Platoon Dynamics Prediction

With the development of autonomous driving and vehicle-to-everything (V2X) technologies, vehicles are transitioning from manual driving to autonomous driving. Current research largely focuses on single-vehicle trajectory prediction, which struggles to effectively characterize platoon-level interactions, resulting in insufficient prediction accuracy and physical consistency in traffic fluctuation scenarios. To this end, this paper proposes a physics-informed GAT-Transformer-IDM model for predicting the longitudinal dynamics of a platoon operating on highways. GAT-Transformer is used to extract the spatio-temporal interaction features of the platoon, while differentiable IDM provides adaptive physical guidance, thus forming a unified physical information fusion learning framework. Accordingly, the extracted platoon trajectory data are first spatially encoded using the graph attention network (GAT) to extract vehicle interactions, learning simultaneous interactions such as following and avoiding. Transformer temporal encoding then captures the temporal evolution of the data and learns the historical behavior of each vehicle. Furthermore, by introducing a differentiable IDM, we achieve adaptive parameter learning and end-to-end gradient optimization. In end-to-end training, the output of GAT-Transformer is connected to the physical computation graph of IDM through differentiable parameters, so that the model can satisfy both data regularity and physical regularity during learning. Finally, the proposed GAT-Transformer-IDM model was trained on the highD dataset. Prediction experiments showed that the GAT-Transformer-IDM achieved significantly lower prediction errors for speed and relative distance than other baseline models. Compared to the baseline models, the GAT-Transformer-IDM reduced the mean squared error of speed prediction by up to 33% and the mean squared error of relative distance prediction by up to 29%.