面向数据驱动的车辆跟驰行为建模

Modeling of Car-following Behavior Based on Data-driven Method

数据驱动类跟驰模型可充分挖掘车辆轨迹数据中的跟驰行为特性,从而实现对车辆跟驰行为的预测。为研究车辆跟驰行为特性,基于实测轨迹数据,运用数据驱动的方法并结合GM跟驰模型建模思路,基于注意力机制与驾驶员进行跟驰行为决策时着重关注重要信息机制相一致的特性,加入注意力机制,建立了基于CNN-BiLSTM-Attention的车辆跟驰模型。筛选NGSIM数据集中符合车辆跟驰特性的轨迹数据并进行降噪处理,通过试验选择最优模型网络结构,对模型进行训练,与LSTM,GRU,CNN-BiLSTM等数据驱动类跟驰模型进行对比分析。结果表明:基于CNNBiLSTM-Attention的车辆跟驰模型与LSTM模型相比,MAE减少约38.12%,R2提高约2.18%,MSE减少约23.45%;与GRU模型相比,MAE减少约19.05%,R2提高约1.13%,MSE减少约13.95%;与CNN-BiLSTM模型相比,MAE减少约1.06%,R2提高约0.15%,MSE减少约1.78%,即模型具有最优的加速度预测性能,可更好地描述车辆跟驰行为;基于CNN-BiLSTM-Attention的车辆跟驰模型与单一模型相比,可更好地提取交通数据特征并处理历史数据间的关系,捕捉历史信息并做出相应决策,与CNN-BiLSTM相比,Attention机制使得模型着重关注输入特征的重要部分,从而具有更高的预测精度。该模型可实现车辆加速度预测,可帮助理解车辆跟驰特性,未来可为车辆跟驰行为决策提供理论依据。

The data-driven car-following model can fully explore the characteristics of car-following behavior in vehicle trajectory data,so as to realize the prediction of car-following behavior.To study the characteristics of car-following behavior,based on the measured trajectory data,using the data-driven method and combining the GM car-following model modeling ideas,based on the attention mechanism and the characteristics of the driver’s attention to the important information mechanism when making car-following behavior decisions,the attention mechanism is added to establish a car-following model based on CNNBiLSTM-Attention.The trajectory data that conforms to the car-following characteristics in NGSIM is screened and processed for noise reduction.The optimal model network structureis selected for training through experiments,and compared with data-driven car-following models such as LSTM,GRU and CNN-BiLSTM.The result shows that(1)compared with LSTM model,the car-following model based on CNN-BiLSTM Attention has a decrease of MAE by 38.12%,an increase of R2 by 2.18%,and a decrease of MSE by 23.45%;(2)compared with GRU model,MAE decreases by 19.05%,R2 increases by 1.13%and MSE decreases by 13.95%;(3)compared with the CNN-BiLSTM model,MAE decreases by 1.06%,R2 increases by 0.15%and MSE decreases by 1.78%,indicating that the model has the best acceleration prediction performance and can better describe car-following behavior;(4)compared with the single model,the car-following model based on CNN-BiLSTM-Attention can better extract the features of traffic data and process the relationship among historical data,capture historical information and make corresponding decisions,the attention mechanism allows the model to focus on important parts of the input characteristics,thus providing higher prediction accuracy.The model can be used to predict vehicle acceleration,help to understand car-following characteristics,and provide theoretical basis for car-following behavior decisionmaking in the future.