基于卷积自适应平滑网络的高速公路交通状态估计方法  

Freeway Traffic State Estimation Based on Convolutional Adaptive Smoothing Network

作　　者：刘婧 陆良[1,2] 郑芳芳 刘晓波 LIU Jing;LU Liang;ZHENG Fang-fang;LIU Xiao-bo(School of Transportation and Logistics,Southwest Jiaotong University,Chengdu 611756,Sichuan,China;National Engineering Laboratory of Integrated Transportation Big Data Application Technology,Southwest Jiaotong University,Chengdu 611756,Sichuan,China;Department of Traffic Management Engineering,Shandong Police College,Jinan 250014,Shandong,China)

机构地区：[1]西南交通大学交通运输与物流学院,四川成都611756 [2]西南交通大学综合交通大数据应用技术国家工程实验室,四川成都611756 [3]山东警察学院交通管理工程学院,山东济南250014

出　　处：《中国公路学报》2024年第11期275-288,共14页China Journal of Highway and Transport

基　　金：国家重点研发计划项目(2021YFB1600100);国家自然科学基金项目(52072315)。

摘　　要：针对连续流场景的交通状态估计问题,提出了一种基于卷积神经网络的改进自适应平滑方法。该方法能够利用稀疏的固定断面观测数据,重构完整的交通状态。与传统的自适应平滑方法相比,所提出的方法不依赖于经验选择交通状态的传播特征,而是针对交通状态的传播特性,设计了3种各向异性卷积核,分别对应于交通波的前向传播、反向传播和双向传播特征,有效提取自由流和拥堵流状态的特征。基于交通流稳态基本关系和冲击波理论,进一步提出交通基本图(Fundamental Diagram, FD)权重算子,用于自适应加权不同的交通特征,优化交通状态的重构效果。不同于传统自适应平滑方法采用Sigmoid函数确定加权权重,所提出的算子在加权过程中明确包含了交通流波动的物理意义,并能直观反映交通流的一阶连续流特性。通过对高速公路真实轨迹数据集的验证,结果显示在所提出的深度学习估计框架下,各向异性卷积核能够在确保估计精度的同时,减少约1/3的待训练参数量,并使估计结果更符合交通流的物理特征。同时,所提方法能显著降低整体平均绝对误差(MAE)和平均相对百分比误差(MAPE),与传统自适应平滑方法相比,分别降低了22.3%和31.35%。特别是在拥堵状态下,所提方法能进一步减少估计误差,MAE和MAPE分别降低了31.1%和37.58%。此外,所提方法对微小交通扰动的敏感性较低,且随着估计点与观测断面距离的增加,误差增长速度更慢。在不同观测断面间距下的性能对比显示,所提方法在任意断面间距下均展现出最低的估计误差,并且随着断面间距的增加,误差增长速度最慢。这些结果从多个角度证明了所提方法的有效性和优越性。This study proposes a novel traffic state estimation method that leverages convolutional neural networks(CNNs)for the adaptive smoothing of cross-sectional traffic flow data to reconstruct a complete traffic state.Unlike traditional adaptive smoothing methods,which depend on empirical selection for traffic propagation characteristics,the proposed method utilizes three distinct anisotropic convolution kernels.These kernels are carefully designed to align with the specific propagation patterns of traffic waves(forward,backward,and bidirectional),thereby effectively distinguishing between free-flow and congested states.Furthermore,a novel fundamental diagram(FD)-based weight operator is introduced that adaptively weighs different traffic features to accurately depict traffic equilibrium conditions.Unlike traditional adaptive smoothing methods that use sigmoid functions to determine weighting,this operator explicitly incorporates the physical meaning of traffic flow fluctuations and clearly represents first-order characteristics of continuous traffic flow.Validation against field freeway trajectory data demonstrates that the proposed framework,with its anisotropic convolutional kernels,maintains high estimation accuracy and reduces the number of trainable parameters by approximately one-third,thus ensuring results more closely aligned with the physical properties of traffic flow.The proposed method significantly outperforms traditional adaptive smoothing techniques,reducing the overall mean absolute error(MAE)and mean absolute percentage error(MAPE)by 22.3 and 31.35%,respectively.Remarkably,under congested conditions,the proposed approach shows even greater precision,with MAE and MAPE reductions of 31.1 and 37.58%,respectively.Moreover,the proposed model demonstrates lower sensitivity to minor traffic disturbances,and a slower rate of error increase as the distance between the estimation locations and observation cross-sections increase.Across various observation section spacings,the proposed method consistently surp

关 键 词：交通工程 交通状态估计 卷积神经网络 自适应平滑方法 各向异性卷积核 交通波 

分 类 号：U491.2[交通运输工程—交通运输规划与管理]