A physics-based deep learning method for longitudinal trajectories prediction of platoon on highways

In recent years, with the popularization of autonomous vehicles and the development of vehicle networking technology, the demand for accurate prediction of vehicle behavior has become increasingly urgent. However, most current studies focus on single-vehicle trajectory prediction, lacking effective extraction and capture of vehicle behavior interaction features at the platoon scale. In terms of platoon-level trajectory prediction, especially in traffic fluctuation scenarios, the prediction accuracy still needs to be improved. More importantly, while ensuring high prediction accuracy, it is also necessary not to lose physical interpretability. To this end, this paper proposes a physics-based deep learning network CNN-Int-LSTM-IDM to predict the longitudinal trajectories of all vehicles in a following platoon on a highway. Specifically, the constraints of the physical model are used to prevent the pure data model from generating trajectories that violate physical laws, especially in traffic congestion scenarios, and the physical parameters are calibrated in advance on the trajectory dataset. In addition, this paper adopts a planned sampling mechanism in the model training process to prevent the error from accumulating and propagating over time, while considering the interaction behavior of all vehicles in the entire platoon to optimize the trajectory generation accuracy of the entire platoon. Besides, a convolutional neural network (CNN) is used in advance to extract the interaction features between vehicles in the platoon, and Int-LSTM then processes the temporal dependencies of these features. Finally, the proposed CNN-Int-LSTM-IDM is trained on the highD and NGSIM (Next Generation Simulation) dataset. The prediction experimental results show that the prediction error of CNN-Int-LSTM-IDM is significantly smaller than that of the three baseline models Int-LSTM, CNN-Int-LSTM, and Int-LSTM-IDM. Compared with these three baseline models, the mean square error of CNN-Int-LSTM-IDM on the highD dataset is reduced by 33 %, 21 %, and 18 %, respectively. In addition, in the simulation experiment, the performance of CNN-Int-LSTM-IDM is also more robust.