To improve the characteristics extraction ability and safety of the composite network model for car-following behavior, an improved composite network RD_BiLSTM_Attention (RDBLA), based on residual dense module (RDM), is proposed, and the hybrid-driven car-following model, based on RDBLA and intelligent driver model (IDM), is established through the optimal weighting theory. An actual driving dataset was obtained from OpenACC and preprocessed, using the dataset to train RDBLA and calibrate the IDM parameters. Finally, the performance of the RDBLA model and the hybrid-driven model is verified through simulation comparisons with existing models. The results show that the multi-step prediction and RDM in RDBLA enhance the ability of the network to extract driving characteristics and improve prediction accuracy. RDM increases inference efficiency by 48.39%. The hybrid-driven model achieves a balance between prediction accuracy and car-following safety, and the minimum driving space increased by 55.84% compared to RDBLA, with a stable car-following process. The total prediction error is reduced by 22.05% and 27.55% compared to the single IDM and RDBLA, respectively. Compared with the RBFNN-IDM model, the hybrid-driven model has higher traffic efficiency and can capture the asymmetric driving characteristics well.

为了提高跟驰行为复合网络模型的特征提取能力和安全性，提出了一种基于残差密集模块（RDM）的改进复合网络RD_BiLSTM_Attention（RDBLA），并建立了混合驱动的跟驰模型，基于RDBLA和智能驾驶员模型（IDM），通过最优权重理论建立。从OpenACC获取实际驾驶数据集并进行预处理，使用该数据集训练RDBLA并校准IDM参数。最后，通过与现有模型的仿真比较，验证了RDBLA模型和混合驱动模型的性能。结果表明，RDBLA中的多步预测和RDM增强了网络提取驾驶特征的能力，提高了预测精度。 RDM 推理效率提升 48.39%。混合驱动模型实现了预测精度和跟车安全性的平衡，最小行驶空间较RDBLA提升了55.84%，跟车过程稳定。与单一IDM和RDBLA相比，总预测误差分别降低了22.05%和27.55%。与RBFNN-IDM模型相比，混合驱动模型具有更高的交通效率，并且能够很好地捕捉非对称驾驶特性。