预训练长短时空交错Transformer在交通流预测中的应用  

Pre-trained long-short spatiotemporal interleaved Transformer for traffic flow prediction applications

作　　者：马莉 王永顺[1] 胡瑶[1] 范磊 MA Li;WANG Yongshun;HU Yao;FAN Lei(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;School of Electronic and Electrical Engineering,Lanzhou Petrochemical University of Vocational Technology,Lanzhou 730060,China;School of Telecommunication Engineering,Lanzhou Bowen College of Science and Technology,Lanzhou 730101,China)

机构地区：[1]兰州交通大学电子与信息工程学院,甘肃兰州730070 [2]兰州石化职业技术大学电子电气工程学院,甘肃兰州730060 [3]兰州博文科技学院电信工程学院,甘肃兰州730101

出　　处：《浙江大学学报(工学版)》2025年第4期669-678,共10页Journal of Zhejiang University:Engineering Science

基　　金：国家自然科学基金资助项目(6136606);甘肃省教育厅高校教师创新基金资助项目(2023B-294).

摘　　要：为了削弱和消除短期交通流预测普遍存在的时空幻影现象,基于Transformer网络和自监督预训练-全监督训练框架,提出新型预训练长短时空交错Transformer模型.采用自监督预训练的方式获得长期时空异质性,设计时空交错模块进行交互获得长期时空异质交互性.设计短时空循环Transformer,将短期时空序列循环压缩提取至能够表现整个短期时空序列独特时空特征的空间片上.在长期时空交错的时空异质交互性指导下,将未来时间与近似特征匹配,重建未来短期时空序列.比较不同交通流预测模型在4个交通流标准数据集和2个交通速度数据集上的预测精度和多步长.实验结果表明,相比当前先进模型,所提模型提升了交通数据预测的精确性.To mitigate and eliminate the common spatiotemporal illusions in short-term traffic flow prediction,a novel pre-training long-short spatiotemporal interleaved Transformer model was proposed,based on the Transformer network and a self-supervised pre-training to fully supervised training framework.Long-term spatiotemporal heterogeneity was acquired by the self-supervised pre-training,and a spatiotemporal interleaving module was designed to interact and obtain the long-term spatiotemporal heterogeneous interactivity.A short spatiotemporal recurrent Transformer was designed to compress and extract the short-term spatiotemporal sequences onto a spatial slice,which represented the unique spatiotemporal features of the entire short-term sequence.Guided by the long-term spatiotemporal interleaved heterogeneous interactivity,similar features were matched on the future timeline to reconstruct the future short-term spatiotemporal sequence.Different traffic flow prediction models were compared in terms of accuracy and multi-step predictions in four traffic flow benchmark datasets and two traffic speed datasets.Experimental results show that the proposed model improves the accuracy of traffic data prediction compared to current state-of-the-art models.

关 键 词：智能交通 交通流预测 TRANSFORMER 深度学习 自监督 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]