Scholarlink

<p>El Niño-Southern Oscillation (ENSO) prediction is one of the most important climate forecasting tasks. Traditional ENSO prediction methods are limited by complex physical models and empirical rules, leading to insufficient accuracy of their predictions. In particular, new climate impacts have exacerbated the complexity of ENSO changes. In this study, we proposed a novel spatio-temporal neural network model, LSTA-Swin, that aggregates long- and short-term information, to improve the accuracy of ENSO prediction. By CMIP6, SODA, and GODAS data, LSTA-Swin outperforms the traditional CanCM4, CCSM3, and GFDLaer04 dynamic forecasting systems by 7.57, 23, and 13.63% respectively, in terms of the average correlation skill of the Niño3.4 index with 11 months in advance. Besides, it significantly surpasses mainstream deep learning models, such as CNN and 3D-Geoformer, in the long-term correlation skill of the Niño3.4 index and can predict effectively up to 20 months. Under the optimal conditions, LSTA-Swin demonstrates its ability to alleviate the Spring Predictability Barrier hence simulate the spatio-temporal changes of super-strong ENSO. This study shows the great potential of utilizing innovative deep learning methods with existing multi-sourced data to efficiently learn the complex dynamic changes of ENSO.</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

LSTA-Swin: a long- and short-term aggregated spatio-temporal neural network for enhanced ENSO forecasting

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Wei Fang,
  • Xiao-Zhi Zhang,
  • Yijing Li,
  • Zhang-Jie Fu,
  • Hai-Yan Fu

摘要

El Niño-Southern Oscillation (ENSO) prediction is one of the most important climate forecasting tasks. Traditional ENSO prediction methods are limited by complex physical models and empirical rules, leading to insufficient accuracy of their predictions. In particular, new climate impacts have exacerbated the complexity of ENSO changes. In this study, we proposed a novel spatio-temporal neural network model, LSTA-Swin, that aggregates long- and short-term information, to improve the accuracy of ENSO prediction. By CMIP6, SODA, and GODAS data, LSTA-Swin outperforms the traditional CanCM4, CCSM3, and GFDLaer04 dynamic forecasting systems by 7.57, 23, and 13.63% respectively, in terms of the average correlation skill of the Niño3.4 index with 11 months in advance. Besides, it significantly surpasses mainstream deep learning models, such as CNN and 3D-Geoformer, in the long-term correlation skill of the Niño3.4 index and can predict effectively up to 20 months. Under the optimal conditions, LSTA-Swin demonstrates its ability to alleviate the Spring Predictability Barrier hence simulate the spatio-temporal changes of super-strong ENSO. This study shows the great potential of utilizing innovative deep learning methods with existing multi-sourced data to efficiently learn the complex dynamic changes of ENSO.