Scholarlink

<p>Capturing nearshore wave propagation is hindered by coarse global data and computationally expensive high-resolution models. Deep learning provides rapid, efficient downscaling alternatives, though it relies heavily on extensive reliable datasets. This paper proposes a statistical downscaling model developed using Attentional Long Short-Term Memory (ATT-LSTM) to predict the significant wave height in nearshore area from global wave data, replacing physical downscaling based on numerical modeling. The significant wave height and wind of ERA5 are taken as low-resolution inputs, and the high-resolution significant wave height is derived from the unstructured grid continuous wave simulation of the wave-current coupling model (SWAN-ADCIRC) as a reference. ATT-LSTM outperforms traditional and LSTM baselines (overall RMSE is 0.106, correlation coefficient R is 0.83). It robustly predicts both typhoon extremes and low-wave events (R&gt;0.99) while achieving a 100-fold speedup over physical models. This approach offers a valuable solution for downscaling environmental fields, especially when computational power and observational data are limited.</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

Downscaling prediction of significant wave height in nearshore waters

  • 期刊论文

  • 发表日期：

    2026年4月14日

查看全文
  • Shenwen Xu,
  • Qiushi Hao

摘要

Capturing nearshore wave propagation is hindered by coarse global data and computationally expensive high-resolution models. Deep learning provides rapid, efficient downscaling alternatives, though it relies heavily on extensive reliable datasets. This paper proposes a statistical downscaling model developed using Attentional Long Short-Term Memory (ATT-LSTM) to predict the significant wave height in nearshore area from global wave data, replacing physical downscaling based on numerical modeling. The significant wave height and wind of ERA5 are taken as low-resolution inputs, and the high-resolution significant wave height is derived from the unstructured grid continuous wave simulation of the wave-current coupling model (SWAN-ADCIRC) as a reference. ATT-LSTM outperforms traditional and LSTM baselines (overall RMSE is 0.106, correlation coefficient R is 0.83). It robustly predicts both typhoon extremes and low-wave events (R>0.99) while achieving a 100-fold speedup over physical models. This approach offers a valuable solution for downscaling environmental fields, especially when computational power and observational data are limited.