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 (1): 318-328.DOI: 10.16097/j.cnki.1009-6744.2026.01.029

• 工程应用与案例分析 • 上一篇    下一篇

基于两级自适应空间建模的多视距高速公路交通流预测

邹复民*1，陈培烨1，蔡祈钦1，廖律超1，罗永煜2   

  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. 福建省高速公路信息科技有限公司，福州350011
  • 收稿日期:2025-11-24 修回日期:2025-12-23 接受日期:2025-12-29 出版日期:2026-02-25 发布日期:2026-02-17
  • 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)。

Traffic Flow Prediction of Multi-horizon Expressway Based on Two-level Adaptive Spatial Modeling

ZOU Fumin*1, CHEN Peiye1, CAI Qiqin1, LIAO Lvchao1, LUO Yongyu2   

  1. 1. Fujian Key Laboratory for Automotive Electronics and Electric Drive, Fujian University of Technology, Fuzhou 350118, China; 2. Fujian Provincial Expressway Information Technology Co Ltd, Fuzhou 350011, China
  • Received:2025-11-24 Revised:2025-12-23 Accepted:2025-12-29 Online:2026-02-25 Published:2026-02-17
  • Supported by:
    National Natural Science Foundation of China(41971340)；Foundation for Cultivated Young Talents of Fujian Province, China (2025350443)。

PDF

摘要： 交通流预测是智能交通系统的核心能力，可支撑交通管理部门的实时调度与路网优化。然而，高速公路交通流在时间上变化剧烈，在空间上呈现动态关联，使得静态拓扑难以准确刻画路段间的依赖关系，同时预测视距增加时误差易快速累积，进一步加大建模难度。为此，本文提出两级自适应时空网络模型（TLASTN）。模型首先通过多尺度卷积与双向门控循环单元（BiGRU）提取时间序列特征，捕捉多尺度时序模式；随后在逐帧动态图上进行两级空间建模：第1级以路段特征相似为主，结合路网拓扑距离进行约束，从而生成稀疏化动态邻接并形成掩码，用于筛选物理合理的候选邻居；第2级在该掩码约束下采用图注意力机制，对候选邻居分配动态权重，以刻画不同交通状态下的精细化空间依赖。预测框架采用共享编码器与独立卷积预测头的分层结构，可同时支持5、15、30、60min多视距预测。在福建省G15沈海高速和G25长深高速电子收费系统（ETC）门架数据上的实验表明，TLASTN在所有视距上均取得最优性能，其中，在G15上，相比先进基线模型的平均绝对百分比误差（MAPE）指标降低3.0%~5.9%。研究表明，在逐帧动态图上采用两级空间建模能够有效提升动态场景下的交通流预测精度，为高速公路运行管理与决策提供可行技术方案。

关键词: 智能交通, 交通流预测, 两级自适应时空网络, 高速公路, 动态图学习

Abstract: Traffic flow prediction is a core capability of intelligent transportation systems, which supports the real-time traffic dispatching and road network optimization for traffic management authorities. However, the traffic flow of highway exhibits rapid variations and dynamic spatial correlations, which makes static topologies inadequate for accurately characterizing inter-segment dependencies. Meanwhile, the errors of prediction tend to accumulate when the forecasting horizon increases, which further increase the difficulty of modeling. To address these challenges, this paper proposes a Two-Level Adaptive Spatio-Temporal Network (TLASTN). The model first extracts temporal features and captures multi-scale sequential patterns through multi-scale convolutions and a bidirectional GRU. Subsequently, a two-level spatial modeling is conducted on frame-wise dynamic graphs. At the first level of this model, the similarity of segment feature is taken as the primary criterion. This model is constrained by the distance of road network topological to generate a sparse dynamic adjacency matrix and mask, which is used to screen physically reasonable candidate neighbors. At the second level, a graph attention mechanism is applied under the mask constraint to assign dynamic weights to the candidate neighbors, which enables a fine-grained modeling of the spatial dependencies under different traffic states. The prediction framework adopts a hierarchical structure with a shared encoder and independent convolutional prediction heads, which allows simultaneous multi-horizon forecasting at 5, 15, 30, and 60 minutes. Experiments on the ETC gantry data from the G15 Shenhai Expressway and the G25 Changshen Expressway in Fujian Province demonstrate that TLASTN achieves the best performance across all forecasting horizons, among which, on the G15 dataset, TLASTN reduces MAPE by 3.0%~5.9% compared with baseline models. The results indicate that adopting a two-level spatial modeling on frame-wise dynamic graphs can effectively improve the accuracy of traffic flow prediction in dynamic scenarios, which provides a feasible technical solution for expressway operation management and decision-making.

Key words: intelligent transportation, traffic flow prediction, two-level adaptive spatio-temporal network, expressway, dynamic graph learning

中图分类号: