基于生成对抗网络的交通流参数实时估计模型  

Real-time Traffic Flow Parameters Estimation Model Based on Generative Adversarial Network

作　　者：姚荣涵[1] 王荣贇 张文松 叶劲松[2] 孙锋[3] YAO Rong-han;WANG Rong-yun;ZHANGWen-song;YE Jin-song;SUN Feng(School of Transportation and Logistics,Dalian University of Technology,Dalian 116024,Liaoning,China;Key Laboratory of Transport Industry of Big Data Application Technologies for Comprehensive Transport,China Academy of Transportation Sciences,Beijing 100029,China;School of Transportation and Vehicle Engineering,Shandong University of Technology,Zibo 255049,Shandong,China)

机构地区：[1]大连理工大学,交通运输学院,辽宁大连116024 [2]交通运输部科学研究院,综合交通运输大数据应用技术交通运输行业重点实验室,北京100029 [3]山东理工大学,交通与车辆工程学院,山东淄博255049

出　　处：《交通运输系统工程与信息》2022年第3期158-167,共10页Journal of Transportation Systems Engineering and Information Technology

基　　金：国家自然科学基金(52172314);交通运输部科学研究院综合交通运输大数据应用技术交通运输行业重点实验室开放课题(2020B1203)。

摘　　要：为有效调控道路网时空资源,需实时估计交通流参数。若要准确估计交通流参数,应详细考虑道路网交通流时空特征。本文基于生成对抗网络,提出一种能捕捉交通流时空特征的实时估计模型,即TSTGAN模型。该模型包括生成器和判别器两部分,生成器利用门控卷积神经网络捕捉交通流的动态空间特征,使用基于注意力机制的长短期记忆神经网络分析交通流的动态时间特征;采用门控卷积神经网络与长短期记忆神经网络构建判别器;通过对抗方式训练生成对抗网络的生成器与判别器,实时获得交通流参数估计值。使用中国山东省淄博市12个卡口设备和美国加州洛杉矶市23个线圈检测器获得的交通流量数据,验证TSTGAN模型的可靠性。结果表明,TSTGAN模型引入的时空模块能有效提取交通流的时空特征,所得均方根误差和平均绝对误差比现有模型分别降低2.12%~42.41%和1.66%~40.49%,证明所提TSTGAN模型可以提高交通流参数的估计精度。To effectively allocate the spatio-temporal resources of a road network,it is necessary to estimate the traffic flow parameters in real time.The accurate estimation of traffic flow parameters requires the detailed consideration of the spatio-temporal characteristics of traffic flow in the road network.Based on the generative adversarial network,a real-time estimation model that can capture the spatio-temporal characteristics of traffic flow was formulated,that is,the TSTGAN model.This model included a generator and a discriminator.In the generator,the gated convolutional neural network was used to capture the dynamic spatial characteristics of traffic flow,and the long short-term memory neural network based on the attention mechanism was used to analyze the dynamic temporal characteristics of traffic flow.The discriminator consisted of the gated convolutional neural network and the long short-term memory neural network.The generator and discriminator in the generative adversarial network were trained by an adversarial mode,and the real-time estimated values of traffic flow parameters were obtained.The reliability of the TSTGAN model was validated using the traffic flow data obtained from 12 bayonet devices in Zibo City,Shandong Province,China,and 23loop detectors in Los Angeles,California,America.The results show that:the introduced spatio-temporal block in the TSTGAN model can effectively extract the spatio-temporal characteristics of traffic flow,and the obtained root mean square and mean absolute errors decrease by 2.12% ~42.41% and 1.66% ~40.49%,respectively,compared with those obtained from the existing models,which indicates that the formulated TSTGAN model can improve the estimation precision of traffic flow parameters.

关 键 词：智能交通 生成对抗网络 深度学习 交通流参数 时空特征 

分 类 号：U491[交通运输工程—交通运输规划与管理]