Scholarlink

<p>Accurate traffic flow prediction serves as a linchpin of intelligent transportation systems (ITS), underpinning optimized traffic control strategies and sustainable urban development initiatives. Traditional approaches are constrained by their failure to dynamically adapt to evolving and complex road network topologies while maintaining computational efficiency for large-scale deployment scenarios. To address this limitation, an innovative architecture is proposed. It synergistically integrates the Dynamic Graph Neural Network (DGNN), Transformer, and Long Short-Term Memory (LSTM), collectively abbreviated as DGNNTL. The framework comprises two core components: the Dynamic Graph Attention Network (DyGAT) for autonomous topology inference, and a multiscale hybrid temporal modeling module that fuses Transformer and LSTM paradigms. DyGAT-driven adaptive adjacency generation is employed, where a trainable relational scoring matrix can dynamically detect traffic interaction patterns and congestion propagation paths without relying on predefined topological constraints. To enhance predictive accuracy while improving computational efficiency, a dual-path temporal processor incorporates windowed Transformer layers for city-level cyclical trend modeling and residual LSTM units with memory gates to capture fine-grained intersection-scale fluctuations. DyGAT-optimized spatiotemporal fusion is realized via a grouped attention mechanism, yielding real-time predictive capabilities with interpretable feature interactions. This framework establishes a novel interpretable traffic prediction paradigm through dynamic graph attention mechanisms, thereby enabling adaptive signal control systems and providing computational foundations for dynamic routing applications.</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

Dynamic Graph Neural Network-Transformer-LSTM Based Multi-Scale Spatio-Temporal Traffic Forecasting

  • 期刊论文

  • 发表日期：

    2026年2月10日

查看全文
  • Chengyue Hu,
  • Yu Wang,
  • Xiangyue Zhang,
  • Min Zheng,
  • Guifang Feng,
  • Jia liu,
  • Yuyun Kang

摘要

Accurate traffic flow prediction serves as a linchpin of intelligent transportation systems (ITS), underpinning optimized traffic control strategies and sustainable urban development initiatives. Traditional approaches are constrained by their failure to dynamically adapt to evolving and complex road network topologies while maintaining computational efficiency for large-scale deployment scenarios. To address this limitation, an innovative architecture is proposed. It synergistically integrates the Dynamic Graph Neural Network (DGNN), Transformer, and Long Short-Term Memory (LSTM), collectively abbreviated as DGNNTL. The framework comprises two core components: the Dynamic Graph Attention Network (DyGAT) for autonomous topology inference, and a multiscale hybrid temporal modeling module that fuses Transformer and LSTM paradigms. DyGAT-driven adaptive adjacency generation is employed, where a trainable relational scoring matrix can dynamically detect traffic interaction patterns and congestion propagation paths without relying on predefined topological constraints. To enhance predictive accuracy while improving computational efficiency, a dual-path temporal processor incorporates windowed Transformer layers for city-level cyclical trend modeling and residual LSTM units with memory gates to capture fine-grained intersection-scale fluctuations. DyGAT-optimized spatiotemporal fusion is realized via a grouped attention mechanism, yielding real-time predictive capabilities with interpretable feature interactions. This framework establishes a novel interpretable traffic prediction paradigm through dynamic graph attention mechanisms, thereby enabling adaptive signal control systems and providing computational foundations for dynamic routing applications.