面向高速公路非检测点位的全域交通状态预测方法  

Global traffic state prediction method for non-sensing locations on freeways

作　　者：王亦兵 胡然 余宏鑫 李嘉恒 张玉杰 徐志刚[4] 何兆成 陆启荣 WANG Yi-bing;HU Ran;YU Hong-xin;LI Jia-heng;ZHANG Yu-jie;XU Zhi-gang;HE Zhao-cheng;LU Qi-rong(College of Civil Engineering and Architecture,Zhejiang University,Hangzhou 310058,Zhejiang,China;College of Civil Engineering,Tongji University,Shanghai 200092,China;Zhejiang Communications Investment Group Co.,Ltd.,Hangzhou 310020,Zhejiang,China;School of Information Engineering,Chang'an University,Xi'an 710064,Shaanxi,China;School of Intelligent Systems Engineering,Sun Yat-sen University,Shenzhen 518107,Guangdong,China)

机构地区：[1]浙江大学建筑工程学院,浙江杭州310058 [2]同济大学土木工程学院,上海200092 [3]浙江省交通投资集团有限公司,浙江杭州310020 [4]长安大学信息工程学院,陕西西安710064 [5]中山大学智能工程学院,广东深圳518107

出　　处：《交通运输工程学报》2025年第1期274-294,共21页Journal of Traffic and Transportation Engineering

基　　金：国家自然科学基金项目(52272315);浙江省重点研发计划(2024C01180,2022C01129);宁波市国际科技合作项目(2023H020)。

摘　　要：针对既有高速公路交通状态预测研究较少考虑非检测点位和道路拓扑变化的问题,分析了现有研究方法的局限性,提出了一种结合宏观交通流模型、扩展卡尔曼滤波、数据驱动长短时记忆网络(LSTM)的交通状态预测方法,充分发挥机器学习在时域特征表达与可信交通流模型在空间动态跟踪上的优势;基于有限检测点位的流量和速度数据,构建了高速公路网络交通流体模型(METANET)并完成全局模型参数和基本图参数标定,设计了基于METANET和扩展卡尔曼滤波的交通状态估计器,继而训练机器学习模型实现全部检测点位的交通状态预测,并驱动交通状态估计器实现全域交通状态预测。研究结果表明:本文提出的交通状态预测方法能够显著提升高速公路的流量和速度预测精度,其中5 min流量和速度预测平均绝对百分比误差为6.92%和5.29%,相比基线方法分别改善29.62%和24.28%;30 min流量和速度预测平均绝对百分比误差为10.02%和8.62%,相比基线方法分别改善24.84%和15.87%;本文方法充分考虑了出入口匝道流量对主线交通状态的影响,也明显改善了主线流量预测性能。In view of the problem that existing research on traffic state prediction of freeways rarely considers non-sensing locations or road topology changes,the limitations of existing research methods were analyzed.A traffic state prediction method combining macroscopic traffic flow model,extended Kalman filtering,and data-driven long short-term memory(LSTM) was proposed,aiming to fully leverage the advantages of machine learning in temporal feature expression and trustworthy traffic flow models in spatial dynamic tracking.Based on the flow and speed data of limited sensing locations,a model of ecoulement of traffic autoroute for networks(METANET) was constructed,and the global model parameters and fundamental diagram parameters were calibrated.A traffic state estimator based on METANET and extended Kalman filtering was designed.The machine learning model was trained to predict the traffic state of all sensing points,and the traffic state estimator was driven to predict the global traffic state.Research results show that the proposed traffic state prediction method can significantly improve the prediction accuracy of flow and speed of freeways.The mean absolute percentage errors of 5-minute flow and speed predictions are 6.92% and 5.29%,which perform 29.62% and 24.28% better than baseline method,respectively,and those of 30-minute flow and speed predictions are 10.02% and 8.62%,which perform 24.84% and 15.87% better than baseline method,respectively.In addition,the proposed method fully considers the impact of on/off ramp flow on the mainline traffic state,so the performance of mainline traffic flow prediction is significantly improved.

关 键 词：智能交通系统 交通状态预测 融合物理模型的机器学习 非检测点位 卡尔曼滤波 高速公路 

分 类 号：U491.14[交通运输工程—交通运输规划与管理]