结合自相关注意力动态卷积的交通预测网络  

Auto-correlation Attention and Dynamics Convolution Based Spatial-temporal Networks for Traffic Forecasting

作　　者：王思超 汤颖[1] WANG Sichao;TANG Ying(College of Computer Science and Technology,College of Software,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]浙江工业大学计算机科学与技术学院、软件学院,杭州310023

出　　处：《小型微型计算机系统》2024年第6期1418-1425,共8页Journal of Chinese Computer Systems

基　　金：浙江省自然科学基金重点项目(LZ23F020010)资助;国家自然科学基金面上项目(61972355)资助;国家自然科学基金重大项目(72192820)资助。

摘　　要：交通流量预测作为智能交通系统(ITS)的重要任务之一,受到极大关注,其常被建模为时空序列预测问题,准确把握交通数据的时间-空间相关性成为了解决此问题的关键,现有的工作往往采用循环神经网络以捕获时间依赖性,采用图卷积网络以捕获空间依赖性,两者尚未有机的结合且捕获时空依赖的能力有限,导致预测精度不佳.本文提出了用于交通预测的基于自相关注意力和动态卷积的时空网络(AADCSN),设计采用类Transformer架构,结合自相关注意力与动态学习图卷积有效捕获交通数据的时间特征与空间特征,并引入数据蒸馏技术和多种嵌入表示有效提升预测性能.论文选用4个真实数据集和9个先进的基线方法进行比较,实验结果表明,本文提出的模型在几乎所有对比指标上都优于基线模型.Traffic flow prediction,as one of the important tasks of intelligent transportation systems(ITS),has received great attention.The problem of traffic flow prediction is usually modeled as a spatio-temporal sequence prediction problem and accurately capturing the spatio-temporal dependence of traffic data has become the key to solving this problem.Existing works often use recurrent neural networks to capture the time-dependence and graph convolutional networks to capture spatial-dependence.However,the above two networks have not been effectively combined and have limited ability to capture spatio-temporal dependence.In this paper,we propose an Auto-Correlation Attention and Dynamics Convolution based Spatial-Temporal Networks(AADCSN)for traffic flow prediction,which is designed with a Transformer-like architecture.The combination of auto-correlation attention and dynamic learning graph convolutional networks effectively captures the temporal and spatial features of traffic data.Besides,the data distilling and multiple embedding representations of input data have been introduced to effectively improve the prediction accuracy.We perform experiments on four real datasets and compare our results with nine advanced baselines.Experimental results show that the model proposed in this paper outperforms the baseline model in almost all comparison indicators.

关 键 词：交通流量预测 TRANSFORMER 自相关注意力机制 动态图卷积网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]