深度时空混合图卷积的城市交通预测模型  

Urban Traffic Prediction Based on Deep Spatio-temporal Hybrid Graph Convolution

作　　者：郭海锋[1] 许宏伟 周子盛 GUO Haifeng;XU Hongwei;ZHOU Zisheng(Institute of Cybersecurity,Zhejiang University of Technology,Hangzhou 310000,China;College of Information Engineering,Zhejiang University of Technology,Hangzhou 310000,China)

机构地区：[1]浙江工业大学网络空间安全研究院,杭州310000 [2]浙江工业大学信息工程学院,杭州310000

出　　处：《小型微型计算机系统》2025年第1期97-103,共7页Journal of Chinese Computer Systems

基　　金：国家重点研发计划项目(2019YFE0126100)资助;国家自然科学基金项目(52072343)资助;浙江省自然科学基金项目(LY20E080023)资助。

摘　　要：由于交通网络复杂的时空相关性和交通数据的非线性,给交通预测带来了很大的挑战.现有的方法主要关注路网的时空特征,分别对时间相关性和空间相关性进行建模来模拟时空依赖关系.随着城市道路网络的进一步扩大,导致模型对路网空间特征的挖掘能力不足.此外,交通运行状态受到外部环境因素的干扰,交通流在路段传递效应的影响下会出现较大波动.为解决上述问题,提出深度时空混合图卷积模型,利用图卷积网络和图注意力网络的残差连接分别汇聚路网全局和局部信息,扩展图卷积的感受野范围,从而增强路网空间特征的提取能力.受Transformer在长序列预测上的启发,同时为减少计算复杂度,通过引入Informer模型来处理路网数据潜在的时间依赖性,实现对交通流参数的长期预测能力,并对城市天气和POI(医院,学校,商场)等外部因素进行编码来增强路网信息的属性.为验证所提出模型的性能,在真实数据集上开展实验,对模型进行准确性和可行性分析.实验结果表明,深度时空混合图卷积模型预测精度最高达到75.1%,较Transformer和Informer分别提升了2.5%和2.3%,在不同预测范围下都超过了其他基线模型,具有长期的交通预测能力.Due to the complex spatio-temporal correlations and nonlinear patterns in the traffic network,traffic prediction presents substantial challenges.Existing methods primarily focus on the spatio-temporal features of road networks,separately modelingtemporal and spatial correlations to simulate spatio-temporal dependencies.As urban road networks continue to expand,existing models may lack the ability to fully exploit the spatial characteristics of road networks.Furthermore,the traffic operational state is influenced by external environmental factors,leading to significant fluctuations in traffic flow due to segment transmission effects.To address these issues,a deep spatio-temporal hybrid graph convolution model is proposed.The residual connected graph neural network and graph attention network are used to aggregate the global and local information of the road network,respectively,thereby extending the receptive field of graph convolutions to enhance the extraction capability of spatial features.Inspired by the Transformer′s success in long sequence prediction and to reduce computational complexity,the Informer model is introduced to handle the potential temporal dependencies in road network data.This achieves long-term prediction capabilities for traffic flow parameters and encodes external factors such as city weather and points of interest to enhance road network information attributes.To validate the performance of the proposed model,accuracy and feasibility analyses are conducted on real-world datasets.Experimental results demonstrate that the deep spatio-temporal hybrid graph convolution model achieves a peak accuracy of up to 75.1%,outperforming Transformer and Informer models by 2.5%and 2.3%respectively.It surpasses other baseline models across different prediction ranges,showcasing its long-term traffic prediction capabilities.

关 键 词：交通预测 时空依赖 道路网络 图神经网络 长期预测 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]