基于长短期记忆神经网络与注意力机制的智能汽车分车型跟驰模型

Car-Following Behavior Model for Intelligent Vehicles Based on LSTM and Attention Mechanisms with Identifiable Vehicle Types

考虑到跟驰车流中前车车型对智能汽车跟车行为的影响,采用长短期记忆(Long Short Term Memory,LSTM)神经网络,基于NGSIM数据集,通过One-Hot方法编码车型特征,并引入注意力机制(Attention Mechanism)生成输入特征的注意力权重,训练并建立了一种可根据前车车型产生不同跟驰行为的智能车辆跟驰模型(Identifiable Vehicle Type Car-Following Model,IVT-CF)。在不同前车车型的跟车场景中仿真发现,IVT-CF模型仿真车辆的速度和位移的均方误差(Mean Square Error,MSE)比不分车型的LSTM模型分别降低了23.8%、31.7%,比IDM模型分别降低了15.8%、18.7%,仿真精度更高。在混入大型车辆的车队跟驰场景仿真中发现,交通流速度和车头间距的收敛时间为92 s,该模型能较快收敛,具有较好的稳定性和抗干扰能力。

Considering the influence of preceding vehicle models in the following traffic flow on the intelligent vehicle following behavior,this paper adopted LSTM neural network structure to train and establish an Identifiable Vehicle Type Car-following(IVT-CF)model that can generate different following behaviors based on the preceding vehicle type.Based on the NGSIM dataset,vehicle features are encoded using the one-hot method,and the attention mechanism is introduced to generate the attention weights for the input features.The simulation results show that the IVT-CF model achieves higher simulation accuracy,as its speed and displacement MSE are 23.8%and 31.7%lower than those of the LSTM model without considering vehicle types,and 15.8%and 18.7%lower than those of the IDM model,respectively.In the simulation of the convoy following scenario with the presence of large vehicles,it is found that the convergence time for traffic flow speed and vehicle spacing is 92 s.The proposed model can converge faster,with better stability and anti-interference ability.