An Attention-Based Car-Following Model Based on Fused Data

Data-driven car-following models provide an alternative approach to explore the underlying and complex driving behavior, while they suffer from poor generalization ability that makes it difficult to be applied to traffic simulations. To address this issue, this paper proposes an attention-based data fusion car-following (ABF-CF) model. In the ABF-CF model, we propose an innovative approach to construct a multi-source dataset for training models by fusing field data with model simulation data. The major framework of the ABF-CF model is based on a sequence to sequence (seq2seq) learning with the attention mechanism. In addition, safety indicators are introduced in the inputs of the ABF-CF model to avoid possible unrealistic outputs and rear-end accidents. Case studies demonstrate that the ABF-CF model outperforms the existing data-driven and theory-driven models in reproducing vehicle trajectories with higher predictive accuracy and stability. Ablation experiments demonstrate the validity of the model structure. Also, the results indicate that the encoder-decoder structure and attention mechanism can improve the prediction accuracy, and data fusion can improve the safety performance. The platoon simulations demonstrate that the ABF-CF model can well reproduce oscillations in real traffic.