| 摘要: |
| 厄尔尼诺-南方涛动（ENSO）是全球热带太平洋地区最显著的气候事件之一，能对全球气候系统产生重要影响，引发干旱、洪水和热浪等极端气候事件.准确预测ENSO的发生对于农业生产、水资源管理、灾害防范和经济规划具有重要意义.然而，ENSO具有非线性和复杂特性，准确预测其强度、持续时间和发生时机具有较大挑战性.针对这一问题，提出了一种基于线性注意力机制的时空Transformer（Linformer-ST）模型，引入了线性注意力机制，取代传统的Softmax注意力机制，将时空特征的建模复杂度从O(n2)降至O(nlog (n))，显著提升了计算效率.在CMIP6数据集上进行了预训练，在SODA数据集上进行了迁移学习，并在GODAS数据集上进行了验证.实验结果表明，该模型在Nino 3.4海表温度异常预测中表现优异，在20个月的预测范围内保持较高的相关性和精度. |
| 关键词: 厄尔尼诺-南方涛动（ENSO） Nino 3.4指数 Linformer-ST模型 时空预测 |
| DOI：10.20192/j.cnki.JSHNU（NS）.2025.02.010 |
| 分类号:TP391.4 |
| 基金项目: |
|
| Application of multi-head spatiotemporal attention mechanism in long-term ENSO prediction |
|
WU Xinjiao, LIAN Jie
|
|
College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 201418, China
|
| Abstract: |
| The El Niño-Southern Oscillation（ENSO） was one of the most significant climate phenomena in the tropical Pacific region， exerting a significant impact on the global climate system and capable of triggering extreme climate events such as drought， floods， and heatwaves. Accurate prediction of ENSO was crucial for agricultural production， water resource management， disaster prevention and economic planning. However， due to its nonlinear and complex characteristics， it was challenging to accurately predict the intensity， duration， and timing. To address this issue， a spatiotemporal transformer model based on linear attention mechanism （Linformer-ST） was proposed in this paper. In the constructed model the traditional Softmax attention mechanism was replaced with a linear attention mechanism， which reduced the computational complexity of spatiotemporal feature modeling from O(n2) to O(nlog (n))， and improved computational efficiency significantly. The model was pre-trained on the CMIP6 dataset， fine-tuned on the SODA dataset， and validated on the GODAS dataset. Experimental results demonstrated that the model performed exceptionally well in predicting Nino 3.4 sea surface temperature anomalies， maintaining high correlation and accuracy over a 20-month prediction horizon. |
| Key words: El Ni?o-Southern Oscillation （ENSO） Nino 3.4 index Linformer-ST model spatiotemporal prediction |
