交通运输系统工程与信息

交通运输系统工程与信息 ›› 2026, Vol. 26 ›› Issue (5): 227-237.DOI: 10.16097/j.cnki.1009-6744.2026.05.019

• 智能交通系统与信息技术 • 上一篇    下一篇

考虑轨迹语义与时空关联的高速公路行程时间预测方法

邹复民*1 ，江林文1 ，蔡祈钦1,2 ，廖律超1 ，赖树坤3   

  1. 1. 福建理工大学，福建省汽车电子与电驱动技术重点实验室，福州350118；2.漳州职业技术学院， 装备制造与车辆工程学院，福建漳州363000；3.福建省高速公路信息科技有限公司，福州350011
  • 收稿日期:2026-03-07 修回日期:2026-07-14 接受日期:2026-07-25 出版日期:2026-10-25 发布日期:2026-10-21
  • 作者简介:邹复民（1976— ），男，湖南隆回人，教授，博士。
  • 基金资助:
    国家自然科学基金(41971340)；福建省青年科技人员育成项目(2025350443)。

Expressway Travel Time Estimation Method Considering Trajectory Semantics and Spatiotemporal Correlations

ZOU Fumin*1, JIANG Linwen1, CAI Qiqin1,2, LIAO Lvchao1, LAI Shukun3   

  1. 1. Fujian Provincial Key Laboratory of Automotive Electronics and Electric Drive Technology, Fujian University of Technology, Fuzhou 350118, China; 2. College of Equipment Manufacturing and Vehicle Engineering, Zhangzhou Institute of Technology, Zhangzhou 363000, Fujian, China; 3. Fujian Provincial Expressway Information Technology Co Ltd, Fuzhou 350011, China
  • Received:2026-03-07 Revised:2026-07-14 Accepted:2026-07-25 Online:2026-10-25 Published:2026-10-21
  • Supported by:
    National Natural Science Foundation of China (41971340)；Cultivated Young Talents of Fujian Province, China (2025350443)。

PDF

摘要： 针对现有行程时间预测研究中存在未深入挖掘轨迹数据中路段的语义关联，并且无法充分处理行程所蕴含的丰富时空信息问题，本文提出一种考虑轨迹语义与时空关联的行程时间预测方法。首先，设计语义感知轨迹表示学习模块，将行程路段序列视为轨迹语义单元，通过多层Transformer编码器学习路段上下文关联，并引入掩码预测自监督任务优化路段表示质量，获得融合路段语义和拓扑关联的轨迹表示；其次，构建属性嵌入模块，将车辆类型、出发时间和路段长度等多维特征映射到统一特征空间，以增强模型对行程外部影响因素的表达能力；最后，设计自适应门控注意力网络，协同建模全局路径依赖与相邻路段局部关联，并通过注意力输出模块动态分配不同路段对行程时间的贡献。基于福建省3条高速公路电子不停车收费（ETC）和人工收费（MTC）数据的实验结果表明，所提方法在平均绝对误差（MAE）、平均绝对百分比误差（MAPE）和均方根误差（RMSE）指标上均优于对比方法，误差分别降低1.44%、1.46%和1.22%，预测效率较最优基线方法提升19.30%。消融实验进一步验证了轨迹表示学习、自适应门控注意力网络和注意力输出模块可显著提升预测性能。综上，本文通过准确建模轨迹表示，并且学习潜在的行程时空规律，有效地提升了高速公路行程时间预测精度和效率。

关键词: 智能交通, 行程时间预测, 表示学习, 高速公路, 注意力机制

Abstract: To address the insufficient mining of semantic correlations among road segments in trajectory data and the inadequate modeling of rich spatiotemporal information in existing travel time estimation studies, this paper proposes a travel time estimation method that considers trajectory semantics and spatiotemporal correlations. First, a semantic-aware trajectory representation learning module is designed, in which the road segment sequence of a trip is regarded as a trajectory semantic unit. A multi-layer Transformer encoder is used to learn the contextual correlations among the road segments, and a masked prediction self-supervised task is introduced to optimize the quality of road segment representations, thereby obtaining the trajectory representations that integrate the road segment semantics and topological correlations. Second, an attribute embedding module is constructed to map multidimensional features, such as vehicle type, departure time, and road segment length, into a unified feature space, which enhances the ability of model to represent external influencing factors. Third, an adaptive gated attention network is designed to jointly model the global path dependencies and local correlations among the adjacent road segments, while an attention output module dynamically assigns the contribution of each road segment to travel time. Experimental results on the Electronic Toll Collection (ETC) and Manual Toll Collection (MTC) data from three expressways in Fujian Province show that the proposed method outperforms comparison methods in terms of Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE). The errors are reduced by 1.44%, 1.46%, and 1.22%, respectively, and the prediction efficiency is improved by 19.30% compared with the best baseline method. Ablation experiments further verify the effectiveness of trajectory representation learning, adaptive gated attention network, and attention output module in improving prediction performance. In conclusion, the proposed method enhances the accuracy and efficiency of expressway travel time estimation through precise trajectory representation modeling and the extraction of underlying spatiotemporal patterns.

Key words: intelligent transportation, travel time estimation, representation learning, expressway, attention mechanism

中图分类号: