基于谱域超图卷积网络的交通流预测模型  

Traffic Flow Prediction Model Based on Spectral Hypergraph Convolutional Network

作　　者：尹宝才 王竟成[1,2] 张勇 胡永利 孙艳丰[1,2] YIN Baocai;WANG Jingcheng;ZHANG Yong;HU Yongli;SUN Yanfeng(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Beijing Key Laboratory of Multimedia and Intelligent Software Technology,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]北京工业大学多媒体与智能软件技术北京市重点实验室,北京100124

出　　处：《北京工业大学学报》2024年第2期152-164,共13页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(62072015,U1811463,61632006);国家重点研发计划资助项目(2018YFB1600903)。

摘　　要：针对传统图结构难以对节点间的隐含复杂关联关系建模的问题,利用超图对交通流数据进行高阶表示,提出基于谱域超图卷积网络的交通流预测方法。首先,通过动态超边刻画数据特征层面的关系,利用谱域超图卷积,包括基于傅里叶和图小波的超图卷积及门控时序卷积,在多尺度上提取交通流的时空特征,实现端到端的节点级交通流预测。然后,采用北京市以及美国加利福尼亚州真实历史数据集进行预测实验。消融实验通过孤立和重构网络模型验证了所提方法的有效性。全时段和早高峰交通流预测的实验结果表明,该方法预测准确率高于目前主流交通流预测模型。The traditional graph structure ignores the implicit complex relationship between nodes to a certain extent.Aiming at the problem,hypergraph was used to represent traffic flow data at a high level,and a traffic flow prediction method was proposed based on hypergraph convolutional network in spectral domain.First,spectral domain hypergraph convolution and gated temporal convolution were used to extract the spatiotemporal characteristics of traffic flow at multiple scales by describing the relationship at the data feature level through dynamic hyperedges,and end-to-end node-level traffic prediction was realized.Afterward,the real historical data sets of Beijing and California were used to conduct prediction experiments.The ablation experiments verify the effectiveness of the proposed method by isolating and reconstructing the network model;the full-time and morning peak traffic flow prediction experiments show that the prediction accuracy of the proposed method is higher than that of the current mainstream traffic forecasting models.

关 键 词：图神经网络 超图理论 多元时序预测 深度学习 大数据分析 智慧交通 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]