Car-following behavior modeling is of great importance for traffic simulation and analysis. Considering the multi-steps decision-making process in human driving, we propose a sequence to sequence (seq2seq) learning based car-following model incorporating not only memory effect but also reaction delay. Since the seq2seq architecture has the advantage of handling variable lengths of input and output sequences, in this paper, it is applied to car-following behavior modeling to memorize historical information and make multi-step predictions. We further compare the seq2seq model with a classical car-following model (IDM) and a deep learning car-following model (LSTM). The evaluation results indicate that the proposed model outperforms others for reproducing trajectory and capturing heterogeneous driving behaviors. Moreover, the platoon simulation demonstrates that the proposed model can well reproduce different levels of hysteresis phenomenon. The proposed model is further extended with spatial anticipation, which improves platoon simulation accuracy and traffic flow stability.

跟车行为建模对于交通仿真和分析非常重要。考虑到人类驾驶中的多步决策过程，我们提出了一种基于序列到序列（seq2seq）学习的汽车跟随模型，该模型不仅包含记忆效应而且还包含反应延迟。由于seq2seq体系结构具有处理可变长度的输入和输出序列的优点，因此在本文中，它被应用于汽车跟随行为建模中以存储历史信息并进行多步预测。我们进一步将seq2seq模型与经典的汽车跟随模型（IDM）和深度学习汽车跟随模型（LSTM）进行了比较。评估结果表明，所提出的模型在复制轨迹和捕获异构驾驶行为方面优于其他模型。此外，排仿真表明，所提出的模型可以很好地再现不同水平的磁滞现象。提出的模型通过空间预期得到进一步扩展，从而提高了排仿真的准确性和交通流的稳定性。