Scholarlink

<p>Ocean forecasting is essential to modern ocean science and management, but traditional models face challenges when processing complex ocean data. In this paper, a multivariate ocean forecasting model combining Kolmogorov-Arnold Networks (KAN) and Transformer is proposed based on Pangu-Weather. This approach aims to accurately capture and predict complex associations among ocean multi-variables to provide accurate and stable ocean forecasts. In order to enhance the accuracy and interpretability of the self-attention computation, KAN is used to replace the MLP layer in multi-head self-attention, which enhances the model’s ability to model nonlinear dynamic relationships. In addition, Swin-Transformer is improved for the multivariate ocean forecasting task by using pooled window instead of shifted window, which enriches the multiscale information extraction capability and optimizes the global feature capture. In this paper, experiments are conducted on the South China Sea adopting the analyzed data provided by CMEMS and compared with Pangu-Weather. The experiment results show that Ocean-Kansformer achieves satisfactory prediction accuracy and stability.</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

Ocean-Kansformer: A model combining KAN and transformer for ocean forecasting

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Yongmei Zhang,
  • Lei Zhang,
  • Linghua Tian

摘要

Ocean forecasting is essential to modern ocean science and management, but traditional models face challenges when processing complex ocean data. In this paper, a multivariate ocean forecasting model combining Kolmogorov-Arnold Networks (KAN) and Transformer is proposed based on Pangu-Weather. This approach aims to accurately capture and predict complex associations among ocean multi-variables to provide accurate and stable ocean forecasts. In order to enhance the accuracy and interpretability of the self-attention computation, KAN is used to replace the MLP layer in multi-head self-attention, which enhances the model’s ability to model nonlinear dynamic relationships. In addition, Swin-Transformer is improved for the multivariate ocean forecasting task by using pooled window instead of shifted window, which enriches the multiscale information extraction capability and optimizes the global feature capture. In this paper, experiments are conducted on the South China Sea adopting the analyzed data provided by CMEMS and compared with Pangu-Weather. The experiment results show that Ocean-Kansformer achieves satisfactory prediction accuracy and stability.