可预知性特殊事件下的短时交通状态预测  

Short-term traffic state prediction under planned special events

作　　者：冯小原 陈咨霖 季楠 任毅龙[1,2,3] FENG Xiaoyuan;CHEN Zilin;JI Nan;REN Yilong(School of Transportation Science and Engineering,Beihang University,Beijing 100191,China;Hangzhou Innovation Institute,Beihang University,Hangzhou 310023,China;National Engineering Laboratory for Comprehensive Transportation Big Data Application Technology,Beihang University,Beijing 100191,China;Intelligent Transport System(ITS)R&D Center,Shanghai Urban Construction Design and Research Institute(Group)Co.,Ltd.,Shanghai 200125,China)

机构地区：[1]北京航空航天大学交通科学与工程学院,北京100191 [2]北京航空航天大学杭州创新研究院(余杭),杭州310023 [3]北京航空航天大学综合交通大数据应用技术国家工程实验室,北京100191 [4]上海市城市建设设计研究总院(集团)有限公司智慧城市设计研究院,上海200125

出　　处：《北京航空航天大学学报》2023年第10期2721-2730,共10页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家自然科学基金(U1964206)。

摘　　要：精准的短时交通状态预测是实施有效的交通管理与控制的重要依据。而可预知性特殊事件(PSEs)短时间内在其举办地点周边产生异常的交通出行需求,又因为事件发生数量少、数据样本收集困难等不利因素,往往造成预测精度难以保证。为此,通过实测数据分析了PSEs下短时交通演化特性,在此基础上,采用改进的K近邻(KNN)算法框架,提出一种短时交通状态的KNN(PSE-KNN)预测模型,并通过基于深度强化学习的实时超参数优化方法将其构建成自适应PSE-KNN(APSE-KNN)模型,最后以北京市演唱会场景为例对所提模型的效果进行了验证。结果表明:所提模型在多步预测实验中,相对于其他7种对比预测模型,平均减少残差值12.43%、降低绝对值百分比误差29.90%。证明所提模型有优异的快速调整能力,其更适应于PSEs场景下短时交通状态预测任务。Accurate short-term traffic state prediction is an important basis for effective traffic management and control.The planned special events(PSEs)generate abnormal traffic demand around the venue in a short time.However,due to the limited number of the special events and the difficulty in data sample collection,the prediction accuracy is hard to guarantee.Therefore,the short-term traffic evolution characteristics under PSEs are analyzed by measured data.On this basis,a short-term traffic state prediction model is established by using the framework of improved K-nearest neighbor(KNN)algorithm.Therefore,the evolution characteristics of short-term traffic under PSEs are analyzed through real event data,and a short-term traffic state KNN(PSE-KNN)prediction model was proposed.Moreover,through real-time super parameter optimization method based on Deep reinforcement learning,we constructed into an adaptive PSE-KNN(APSE-KNN)model.Finally,the effect of the model is verified by taking the concert scene in Beijing as an example.The results show that in the multi-step prediction experiment,compared with the other seven comparative prediction models,the proposed prediction model reduces the mean residual error by 12.43%and the mean absolute percentage error by 29.90%on average.These results prove that this model has excellent rapid adjustment ability and is more suitable for short-term traffic state prediction task under PSEs.

关 键 词：短时交通状态预测 可预知性特殊事件 K近邻 深度确定性策略梯度 强化学习 

分 类 号：U491.14[交通运输工程—交通运输规划与管理]