智能网联车辆低渗透率下交叉口排队长度估计策略  

Queue Estimation Under Low Penetration of Connected and Automated Vehicle

作　　者：房山 杨澜[1] 赵祥模[1] 王威 魏诚 吴国垣 FANG Shan;YANG Lan;ZHAO Xiang-mo;WANG Wei;WEI Cheng;WU Guo-yuan(School of Information and Engineering,Chang'an University,Xi'an 710064,Shaanxi,China;Center for Environmental Research and Technology,University of California at Riverside,Riverside CA 92501,California,USA)

机构地区：[1]长安大学信息工程学院,陕西西安710064 [2]加州大学河滨分校环境研究与技术中心,加利福尼亚河滨CA 92501

出　　处：《中国公路学报》2024年第11期249-261,共13页China Journal of Highway and Transport

基　　金：国家重点研发计划项目(2021YFB2501205);长安大学博士研究生创新能力培养项目(300203211242);陕西省重点研发计划项目(2022GY-300)。

摘　　要：为了提高混合交通流条件下信号交叉口的车辆通行效率,提出一种智能网联车辆低渗透率下的信号交叉口车辆排队长度估计策略。首先,根据信号交叉口上游区域车辆的随机到达特性,构建考虑智能网联车辆与人类驾驶车辆组成的车辆排队场景。其次,以智能网联车辆的位移差,速度差以及加速度差为输入,以人类驾驶车辆位移差为输出,建立基于Seq2seq架构的车辆微观轨迹前/后向重构模型,采用时间注意力机制判断车辆行驶状态变化的关键时域,提高模型对车辆“走-停”波的重构能力。再次,以当前信号周期排队车辆数为输入,以车辆排队长度为输出,建立基于XGBoost的车辆排队长度估计模型,可在历史样本数据较少的条件下准确估计车辆排队长度。最后,试验基于NGSIM数据集进行模型训练,在不同智能网联车辆渗透率、单信号周期以及多信号周期等条件下验证所提方法性能。结果表明:在10%~30%的低渗透率条件下,与经典时间序列预测模型RNN、LSTM、Seq2seq以及CNN模型相比,所提出的车辆微观轨迹前/后向重构模型的损失函数收敛速度较快,稳定性更好,车辆轨迹均方根误差降低了8.9%~71.7%,且能够准确描述信号交叉口区域车辆的“走-停”波;相比于基于KNN、随机森林与多项式回归模型的排队长度估计方法,所提方法的均方根误差降低了13.56%~91.99%,排队长度估计的运行时间降低至约8 ms,有效证明了所提方法在交叉口车辆排队长度估计的精确性和实时性。To improve the efficiency of traffic at a signalized intersection under mixed traffic conditions,a vehicle queue estimation method with a low penetration of connected and automated vehicle(CAV)is proposed.Based on the random arrival characteristics of the mixed traffic flow in the upstream area of the signalized intersection,a queue estimation of mixed traffic flow was constructed considering the composition of CAV and human-driven vehicle(HV).With the displacement difference,speed difference,and acceleration difference of CAV as an input and displacement difference of HV as an output,a Seq2seq architecture-based vehicle microscopic trajectory forward/backward reconstruction under low penetration conditions was established.The model uses the temporal attention mechanism to determine the key time domain of the vehicle driving state change,and improves the ability to reconstruct the“stop-and-go”waves for the model.Additionally,with the number of vehicles queuing at the current signal cycle as an input and vehicle queuing length as an output,an XGBoost-based vehicle queue length estimation model was developed,which can accurately estimate the vehicle queue length under the condition of low historical sample data.The experiments were based on the NGSIM dataset for model training.The performance of the proposed method was verified under different conditions including different penetrations of CAV,single signal cycle,and multisignal cycle.Under the low penetration rate of 10%-30%,the loss function converges faster and has a better stability compared to the classical time series prediction models of RNN(recurrent neural network),LSTM(long short-term memory),Seq2seq(sequence to sequence),and CNN(convolutional neural network).The root mean square error(RMSE)of the vehicle trajectory is reduced by 8.9%-71.7%.The method could accurately describe the stop-and-go waves at the signalized intersection.Compared to the queue length estimation methods based on KNN(k-nearest neighbors),random forest,and polynomial regression mo

关 键 词：交通工程 车辆排队长度估计 车辆轨迹重构 数据驱动 

分 类 号：U491.264[交通运输工程—交通运输规划与管理]