Scholarlink

<p>Rapid flood forecasting and early warning represent key non-structural measures currently employed to mitigate flood disasters induced by short-term intense rainfall. Enhancing the accuracy, lead time, and generalization capability of flood forecasting in small watersheds remains a critical challenge. In this study, we propose a novel rapid flood forecasting model that integrates Graph Neural Networks (GNN) with Transformer architectures, leveraging deep learning techniques. The model incorporates static physical attributes of surface grids and utilizes hydrodynamic model simulation results to guide learning. Specifically, the GNN component captures spatial relationship weights among grids and their neighboring cells, while the Transformer component models rainfall time series to dynamically track flood evolution processes. Experimental results demonstrate that the proposed GNN-Transformer-based watershed flood forecasting model outperforms traditional numerical models. The predicted results exhibit a mean relative error of no more than 15% compared to hydrodynamic simulation outputs (including cross-sectional discharge, flow velocity, and water depth), with a NSE exceeding 0.990, and a MAE not exceeding 1.220. Moreover, computational efficiency is improved by a factor of 100–200 relative to conventional hydrodynamic models, while fully meeting the accuracy requirements for flood forecasting. Compared to traditional forecasting approaches, this method offers robust technical support for advancing the intelligence of flood management systems across forecasting, warning, rehearsal, and contingency planning.</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

A Study on Rapid Dynamic Flood Forecasting in Small Watersheds Using a GNN-Transformer Approach Integrated with Spatial Physical Information

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Xinxin Pan,
  • Jingming Hou,
  • Donglai Li,
  • Yanhong Wang,
  • Xingyi Li,
  • Jiantao Sun,
  • Chenchen Fan,
  • Yongping Yang

摘要

Rapid flood forecasting and early warning represent key non-structural measures currently employed to mitigate flood disasters induced by short-term intense rainfall. Enhancing the accuracy, lead time, and generalization capability of flood forecasting in small watersheds remains a critical challenge. In this study, we propose a novel rapid flood forecasting model that integrates Graph Neural Networks (GNN) with Transformer architectures, leveraging deep learning techniques. The model incorporates static physical attributes of surface grids and utilizes hydrodynamic model simulation results to guide learning. Specifically, the GNN component captures spatial relationship weights among grids and their neighboring cells, while the Transformer component models rainfall time series to dynamically track flood evolution processes. Experimental results demonstrate that the proposed GNN-Transformer-based watershed flood forecasting model outperforms traditional numerical models. The predicted results exhibit a mean relative error of no more than 15% compared to hydrodynamic simulation outputs (including cross-sectional discharge, flow velocity, and water depth), with a NSE exceeding 0.990, and a MAE not exceeding 1.220. Moreover, computational efficiency is improved by a factor of 100–200 relative to conventional hydrodynamic models, while fully meeting the accuracy requirements for flood forecasting. Compared to traditional forecasting approaches, this method offers robust technical support for advancing the intelligence of flood management systems across forecasting, warning, rehearsal, and contingency planning.