This paper presents a novel method to estimate queue length at signalised intersections using connected vehicle (CV) data. The proposed queue length estimation method does not depend on any conventional information such as arrival flow rate and parameters pertaining to traffic signal controllers. The model is applicable for real-time applications when there are sufficient training data available to train the estimation model. To this end, we propose the idea of “k-leader CVs” to be able to predict the queue which is propagated after the communication range of dedicated short-range communication (the communication platform used in CV system). The idea of k-leader CVs could reduce the risk of communication failure which is a serious concern in CV ecosystems. Furthermore, a linear regression model is applied to weigh the importance of input variables to be used in a neural network model. Vissim traffic simulator is employed to train and evaluate the effectiveness and robustness of the model under different travel demand conditions, a varying number of CVs (i.e. CVs’ market penetration rate) as well as various traffic signal control scenarios. As it is expected, when the market penetration rate increases, the accuracy of the model enhances consequently. In a congested traffic condition (saturated flow), the proposed model is more accurate compared to the undersaturated condition with the same market penetration rates. Although the proposed method does not depend on information of the arrival pattern and traffic signal control parameters, the results of the queue length estimation are still comparable with the results of the methods that highly depend on such information. The proposed algorithm is also tested using large size data from a CV test bed (i.e. Australian Integrated Multimodal Ecosystem) currently underway in Melbourne, Australia. The simulation results show that the model can perform well irrespective of the intersection layouts, traffic signal plans and arrival patterns of vehicles. Based on the numerical results, 20% penetration rate of CVs is a critical threshold. For penetration rates below 20%, prediction algorithms fail to produce reliable outcomes.

中文翻译：

---

基于k-leader连接车辆概念的神经网络排队长度估计算法

本文提出了一种使用联网车辆（CV）数据估算信号交叉口队列长度的新方法。所提出的队列长度估计方法不依赖于任何常规信息，例如到达流量和与交通信号控制器有关的参数。当有足够的训练数据可用于训练估计模型时，该模型适用于实时应用。为此，我们提出了“ k -leader CVs”的思想，以便能够预测在专用短距离通信（CV系统中使用的通信平台）的通信范围之后传播的队列。k的想法-领导简历可能会降低通信失败的风险，这是简历生态系统中的一个严重问题。此外，线性回归模型用于权衡要在神经网络模型中使用的输入变量的重要性。Vissim交通模拟器用于在不同的旅行需求条件，不同数量的CV（即CV的市场渗透率）以及各种交通信号控制场景下训练和评估模型的有效性和鲁棒性。不出所料，当市场渗透率提高时，模型的准确性将因此提高。在拥挤的交通状况（饱和流量）下，与具有相同市场渗透率的不饱和状况相比，所提出的模型更为准确。尽管所提出的方法不依赖于到达模式和交通信号控制参数的信息，但是队列长度估计的结果仍可与高度依赖此类信息的方法的结果相媲美。还使用来自澳大利亚墨尔本当前正在进行的CV测试台（即，澳大利亚集成多模式生态系统）的大型数据对提出的算法进行了测试。仿真结果表明，该模型无论交叉口的布局，交通信号计划和车辆的到达方式如何都能表现良好。根据数值结果，CV的20％渗透率是一个关键阈值。对于低于20％的渗透率，预测算法无法产生可靠的结果。队列长度估计的结果仍可与高度依赖此类信息的方法的结果相媲美。还使用来自澳大利亚墨尔本当前正在进行的CV测试台（即，澳大利亚集成多模式生态系统）的大型数据对所提出的算法进行了测试。仿真结果表明，该模型无论交叉口的布局，交通信号计划和车辆的到达方式如何都能表现良好。根据数值结果，CV的20％渗透率是一个关键阈值。对于低于20％的渗透率，预测算法无法产生可靠的结果。队列长度估计的结果仍可与高度依赖此类信息的方法的结果相媲美。还使用来自澳大利亚墨尔本当前正在进行的CV测试台（即，澳大利亚集成多模式生态系统）的大型数据对所提出的算法进行了测试。仿真结果表明，该模型无论交叉口的布局，交通信号计划和车辆的到达方式如何都能表现良好。根据数值结果，CV的20％渗透率是一个关键阈值。对于低于20％的渗透率，预测算法无法产生可靠的结果。澳大利亚集成多模式生态系统）目前正在澳大利亚墨尔本进行。仿真结果表明，该模型无论交叉口的布局，交通信号计划和车辆的到达方式如何都能表现良好。根据数值结果，CV的20％渗透率是一个关键阈值。对于低于20％的渗透率，预测算法无法产生可靠的结果。澳大利亚集成多模式生态系统）目前正在澳大利亚墨尔本进行。仿真结果表明，该模型无论交叉口的布局，交通信号计划和车辆的到达方式如何都能表现良好。根据数值结果，CV的20％渗透率是一个关键阈值。对于低于20％的渗透率，预测算法无法产生可靠的结果。