Abstract: Traffic flow prediction is one of the core tasks of intelligent transportation systems, with the primary challenge lying in the high coupling of non-Euclidean spatial dependencies and dynamic temporal characteristics within complex road networks. To address the limitations of existing methods in modeling spatial-temporal dependencies and fusing multi-source information, a novel model, STFformer, was proposed. The proposed model designed a global spatiotemporal interaction module, enabling joint modeling of spatial dependencies and temporal dynamic evolution of global nodes. Meanwhile, a weight-sharing dynamic graph convolution module was incorporated to refine node features through a time-varying adjacency matrix, capturing latent non-Euclidean structures and enhancing the model’s perception ability towards dynamic changes in road network topology. Furthermore, a gated enhancement fusion module was designed to realize adaptive integration of multi-source spatial-temporal information and stable representation, thereby improving the model’s generalization and robustness. The validity verification was conducted on a real traffic flow benchmark dataset. The research results show that the proposed model outperforms baseline models across three prediction metrics, confirming its predictive accuracy and stability in complex spatial-temporal scenarios.

Key words: traffic engineering; traffic flow forecasting; spatiotemporal interaction; dynamic graph convolution; attention mechanism; gating mechanism

摘要： 交通流预测是智能交通系统的核心任务之一，其关键挑战在于复杂路网中非欧几里得空间依赖与动态时序特征的高度耦合。针对现有方法在时空依赖建模及多源信息融合方面的局限性，提出了一种新的STFformer模型。该模型设计了全局时空交互模块，实现了对全局节点空间依赖与时间动态演化的联合建模；同时引入权重共享的动态图卷积模块，通过时变邻接矩阵提炼节点特征，以捕捉潜在的非欧几里得结构，增强了模型对路网拓扑动态变化的感知能力；此外，构建了门控增强融合模块，实现多源时空信息的自适应融合与稳定表达，提升了模型的泛化性与鲁棒性；并在真实交通流基准数据集上进行有效性验证。研究结果表明：该模型在3个预测指标上均优于基线模型，验证了其在复杂时空场景下的预测精度与稳定性。

关键词: 交通工程；交通流预测；时空交互；动态图卷积；注意力机制；门控机制