Scholarlink

<p>The El Niño-Southern Oscillation (ENSO) is an important climate phenomenon, a large-scale coupled ocean-atmosphere interannual variability. It mainly occurs in the equatorial eastern Pacific region and has profound impacts on global climate and ecosystems. Conventional graph convolutional networks (GCNs) perform well in ENSO prediction, but they rely on static adjacency matrices and are difficult to adapt to dynamic changes in node characteristics and relationships. To address this problem, we design an adaptive dynamic graph learning structure that enables the model to adapt to these dynamic changes by explicitly modeling the information flow of edge connections. This approach not only improves the interpretability of the forecasting process, but also characterizes the feedback mechanisms and the impact of the North Pacific Oscillation (NPO) on ENSO events through connections between nodes. To capture the seasonal variations of ENSO, we introduce the Transformer to capture the long-term dependencies in the series. In addition, the fused spatio-temporal feature representation is enhanced and redundant information is suppressed by introducing the node feature attention module. Experiments on CMIP5 and SODA datasets show that the proposed DGL-GT method achieves correlation coefficients of more than 79% over six forecast lead times and is able to forecast ENSO up to 15 months ahead of most state-of-the-art methods.</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 dynamic graph-based multiscale spatio-temporal feature enhancement network applied to ENSO prediction

  • 期刊论文

  • 发表日期：

    2025年6月21日

查看全文
  • Wei Shao,
  • Guoxiang Tong

摘要

The El Niño-Southern Oscillation (ENSO) is an important climate phenomenon, a large-scale coupled ocean-atmosphere interannual variability. It mainly occurs in the equatorial eastern Pacific region and has profound impacts on global climate and ecosystems. Conventional graph convolutional networks (GCNs) perform well in ENSO prediction, but they rely on static adjacency matrices and are difficult to adapt to dynamic changes in node characteristics and relationships. To address this problem, we design an adaptive dynamic graph learning structure that enables the model to adapt to these dynamic changes by explicitly modeling the information flow of edge connections. This approach not only improves the interpretability of the forecasting process, but also characterizes the feedback mechanisms and the impact of the North Pacific Oscillation (NPO) on ENSO events through connections between nodes. To capture the seasonal variations of ENSO, we introduce the Transformer to capture the long-term dependencies in the series. In addition, the fused spatio-temporal feature representation is enhanced and redundant information is suppressed by introducing the node feature attention module. Experiments on CMIP5 and SODA datasets show that the proposed DGL-GT method achieves correlation coefficients of more than 79% over six forecast lead times and is able to forecast ENSO up to 15 months ahead of most state-of-the-art methods.