面向短时交通流预测的动态时空图卷积网络  

Dynamic Spatiotemporal Graph Convolutional Networks for Short-Term Traffic Flow Prediction

作　　者：符琦[1] 巴冰 黄崇景 蒋云霞[1] FU Qi;BA Bing;HUANG Chongjing;JIANG Yunxia(School of Computer Science and Engineering,Hunan University of Science and Technology,Xiangtan 411201,China)

机构地区：[1]湖南科技大学计算机科学与工程学院,湖南湘潭411201

出　　处：《湖南科技大学学报（自然科学版）》2024年第1期70-79,共10页Journal of Hunan University of Science And Technology：Natural Science Edition

基　　金：湖南省自然科学基金资助项目(2021JJ30274);湖南省自然科学基金资助项目(2021JJ30278)。

摘　　要：针对现有的交通流量预测模型缺乏对交通数据动态时空相关性建模能力的问题,提出一种新的基于深度学习的动态时空图卷积网络(DSTGCN)模型,该模型无需给出道路网络信息,从交通数据中即可建模时空相关性.动态时空图卷积层包含2个主要部分:(1)动态邻接矩阵生成模块:使用时间自相关机制与空间注意力机制捕捉交通数据中的动态时空相关性;(2)时空图卷积:使用图卷积和标准二维卷积对信息进行高效聚合.DSTGCN通过堆叠动态时空图卷积层,能够捕捉不同时间级别的时空依赖关系.文章提出的方法在美国加利福尼亚州高速公路流量公开数据集上进行了实验,结果表明:所提出的DSTGCN模型在各项评价指标中均优于现有的基准方法.在PeMSD04数据集上,与当前较新的GeoMAN和ASTGCN模型相比,MAE分别降低了4.00和2.16,验证了所提模型在交通流预测中的有效性.Aiming at the problem that the existing traffic flow prediction models lack the ability to model the dynamic spatiotemporal correlation of traffic data,a new deep learning-based dynamic spatiotemporal graph convolutional network(DSTGCN)model is proposed.The model can model spatiotemporal correlations from traffic data without giving road network information.The dynamic spatiotemporal graph convolutional layer consists of two main parts,one is the dynamic adjacency matrix generation module,which uses temporal auto-correlation mechanism and spatial attention mechanism to capture dynamic spatiotemporal correlations in traffic data;the other is the spatiotemporal graph convolution,which uses graph convolution and standard 2D convolution to efficiently aggregate information.By stacking dynamic spatiotemporal graph convolutional layers,DSTCCN is able to capture spatiotemporal dependencies at different temporal levels.The method proposed in this paper is tested on a public California freeway datase.The results show that the proposed DSTGCN model outperforms the existing benchmark methods in every evaluation index.On the PeMSDO4 dataset,compared with the current newer GeoMAN and ASTGCN models,the MAE is reduced by 4.00 and 2.16 respectively,verifying the effectiveness of the proposed model in traffic flow prediction.

关 键 词：交通流预测 时空图卷积 自相关机制 时空图建模 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]