基于混合图卷积的多通道时空交通流预测模型  

Multi-channel spatial-temporal traffic flow prediction based on hybrid static-dynamic graph convolution

作　　者：张雄涛 郑景玉 申情[1,2] 孙丹枫 蒋云良[1,2,4] ZHANG Xiongtao;ZHENG Jingyu;SHEN Qing;SUN Danfeng;JIANG Yunliang(School of Information Engineering,Huzhou University,Huzhou 313000,China;Zhejiang Province Key Laboratory of Smart Management and Application of Modern Agricultural Resources,Huzhou 313000,China;School of Computer,Hangzhou Dianzi University,Hangzhou 310018,China;School of Computer Science and Technology,Zhejiang Normal University,Jinhua 321004,China)

机构地区：[1]湖州师范学院信息工程学院,浙江湖州313000 [2]浙江省现代农业资源智慧管理与应用研究重点实验室,浙江湖州313000 [3]杭州电子科技大学计算机学院,浙江杭州310018 [4]浙江师范大学计算机科学与技术学院,浙江金华321004

出　　处：《电信科学》2023年第9期97-110,共14页Telecommunications Science

基　　金：国家自然科学基金资助项目(No.U22A20102);浙江省“尖兵”“领雁”研发攻关计划项目(No.2023C01150)。

摘　　要：针对交通流预测模型没有考虑道路上下文相关性和空间依赖关系动态性的问题,提出一种基于混合图卷积的多通道时空交通流预测模型(MHGCN)。该模型采用三明治结构(即中间多通道空间模块,两边时间模块)提取时空特征,多通道空间模块又分为静态图卷积模块和动态图卷积模块。静态图卷积模块同时从拓扑空间结构、语义空间结构及其组合中提取特定和公共的特征;动态图卷积模块对不同的特征分配不同的权重,从未知的图结构中提取动态的空间特征。时间模块中采用多头注意力机制提取全局时间特征,采用时间门控机制提取局部时间特征。该模型从不同的空间结构中提取空间信息,从不同时间间隔提取时间信息,建立全局、全面的时空关系。实验结果表明,MHGCN模型在4个公开数据集上的性能优于现有的交通流预测模型。Aiming at the problem that the traffic flow prediction model did not consider the correlation of road context and the dynamics of spatial dependency,a multi-channel spatial-temporal traffic flow prediction based on hybrid static-dynamic graph convolution(MHGCN)was proposed.A sandwich structure(i.e.multi-channel spatial module in the middle and temporal module on both sides)was used in the model to extract spatial-temporal features,and the multi-channel spatial module was divided into static graph convolution module and dynamic graph convolution module.The static graph convolution module simultaneously extracted specific and common features from topological spatial structures,semantic spatial structures,and their combinations.The dynamic graph convolution module assigned different weights to different features and extracts dynamic spatial features from unknown graph structures.In the temporal module,the multi-head attention mechanism was used to extract the global temporal features,and the temporal gating mechanism extracted the local temporal features.The model extracted spatial information from different spatial structures and temporal information from different time intervals to establish a global and comprehensive spatial-temporal relationship.The experimental results show that the MHGCN performs better than the existing traffic flow prediction models on four real world traffic flow datasets.

关 键 词：智能交通 动态图卷积 多头注意力 时空相关性 多通道 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]