Scholarlink

<p>Integrated transport hubs require reliable, fine-grained forecasts of crowd distribution to safeguard operations and sustainable urban mobility. We present Group Evolution Mechanism Embedded Network (GEME-Net), a passenger flow distribution forecasting architecture that fuses multimodal data, including video-derived counts, digital twin-based mobility chains, and railway/metro operations information, via multi-graph spatial representations and event-aware temporal modules, with a distilled lightweight student model for deployment. In a real-world case at Shanghai Hongqiao, GEME-Net consistently outperforms statistical, convolutional, recurrent, graph-based and Transformer baselines across MAE, RMSE and WMAPE, while retaining inference latency compatible with near-real-time use. Ablations indicate that schedule encoding and event-driven frequency enhancement, together with learned long-range and community graphs, are principal contributors to accuracy. By coupling operational signals with spatial semantics, our approach improves hub-scale situation awareness and short-horizon decision support, offering a practical route to resilient crowd management without asserting broader societal or policy impacts.</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

Passenger flow distribution forecasting at integrated transport hub via group evolution mechanism and multimodal data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月27日

查看全文
  • Zhicheng Dai,
  • Dewei Li,
  • Soora Rasouli,
  • Yan Feng,
  • Hua Li,
  • Linhan Zou,
  • Ruonan Zhang

摘要

Integrated transport hubs require reliable, fine-grained forecasts of crowd distribution to safeguard operations and sustainable urban mobility. We present Group Evolution Mechanism Embedded Network (GEME-Net), a passenger flow distribution forecasting architecture that fuses multimodal data, including video-derived counts, digital twin-based mobility chains, and railway/metro operations information, via multi-graph spatial representations and event-aware temporal modules, with a distilled lightweight student model for deployment. In a real-world case at Shanghai Hongqiao, GEME-Net consistently outperforms statistical, convolutional, recurrent, graph-based and Transformer baselines across MAE, RMSE and WMAPE, while retaining inference latency compatible with near-real-time use. Ablations indicate that schedule encoding and event-driven frequency enhancement, together with learned long-range and community graphs, are principal contributors to accuracy. By coupling operational signals with spatial semantics, our approach improves hub-scale situation awareness and short-horizon decision support, offering a practical route to resilient crowd management without asserting broader societal or policy impacts.