Scholarlink

<p>Accurate traffic flow prediction is essential for advancing intelligent transportation systems (ITS), enabling efficient urban mobility, congestion mitigation, and real-time decision-making. However, achieving reliable forecasts remains challenging due to the intricate spatiotemporal dependencies and inherent time delay effects in traffic networks, where disturbances propagate unevenly across nodes. Existing methods often overlook these dynamic interactions and delays, resulting in suboptimal performance, especially in congested or volatile scenarios. To address these issues, we propose a Delay-Aware Dynamic Graph Network (DADGN), which integrates three key innovations: (1) a hierarchical spatiotemporal clustering module leveraging Deep Embedding Clustering (DEC) with contrastive learning to adaptively group nodes exhibiting similar traffic patterns; (2) a Delayed Spatiotemporal Graph (DSTG) propagation module that captures time-varying topologies and employs cascading update gates to model delay propagation effectively; and (3) the substitution of conventional Multi-Layer Perceptrons (MLPs) with Kolmogorov-Arnold Networks (KANs), utilizing learnable spline functions to boost interpretability and nonlinear feature extraction. Extensive experiments on four benchmark datasets (PEMS03, PEMS04, PEMS08, and PEMS-BAY) demonstrate that DADGN consistently outperforms state-of-the-art baselines across short-term and mid-term horizons, while offering improved interpretability and robustness. These results highlight DADGN’s potential for reliable deployment in practical traffic management and urban mobility optimization.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

KAN-based delay-aware dynamic graph networks with spatiotemporal clustering for traffic flow forecasting

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Lujuan Ma,
  • Huan Xu,
  • Jianlin Zhou,
  • Xiaoping Deng

摘要

Accurate traffic flow prediction is essential for advancing intelligent transportation systems (ITS), enabling efficient urban mobility, congestion mitigation, and real-time decision-making. However, achieving reliable forecasts remains challenging due to the intricate spatiotemporal dependencies and inherent time delay effects in traffic networks, where disturbances propagate unevenly across nodes. Existing methods often overlook these dynamic interactions and delays, resulting in suboptimal performance, especially in congested or volatile scenarios. To address these issues, we propose a Delay-Aware Dynamic Graph Network (DADGN), which integrates three key innovations: (1) a hierarchical spatiotemporal clustering module leveraging Deep Embedding Clustering (DEC) with contrastive learning to adaptively group nodes exhibiting similar traffic patterns; (2) a Delayed Spatiotemporal Graph (DSTG) propagation module that captures time-varying topologies and employs cascading update gates to model delay propagation effectively; and (3) the substitution of conventional Multi-Layer Perceptrons (MLPs) with Kolmogorov-Arnold Networks (KANs), utilizing learnable spline functions to boost interpretability and nonlinear feature extraction. Extensive experiments on four benchmark datasets (PEMS03, PEMS04, PEMS08, and PEMS-BAY) demonstrate that DADGN consistently outperforms state-of-the-art baselines across short-term and mid-term horizons, while offering improved interpretability and robustness. These results highlight DADGN’s potential for reliable deployment in practical traffic management and urban mobility optimization.