Acta Scientiarum Naturalium Universitatis Pekinensis ›› 2021, Vol. 57 ›› Issue (6): 1071-1078.DOI: 10.13209/j.0479-8023.2021.114

Previous Articles     Next Articles

Spring Predictability Barrier Phenomenon in ENSO Prediction Model Based on LSTM Deep Learning Algorithm

ZHOU Pei1, HUANG Yingjie1, HU Bingyi1,2, WEI Jun1,3,†   

  1. 1. Key Laboratory of Tropical Atmosphere-Ocean System (MOE), School of Atmospheric Sciences, Sun Yat-Sen University, Guangzhou 510275 2. HSBC Business School, Peking University, Shenzhen 518055 3. School of Marine Sciences, Guangxi University, Nanning 530004
  • Received:2020-12-25 Revised:2021-04-26 Online:2021-11-20 Published:2021-11-20
  • Contact: WEI Jun, E-mail: weijun5(at)mail.sysu.edu.cn

基于LSTM深度学习的ENSO预测及其春季预报障碍研究

周佩1, 黄颖婕1, 胡冰逸1,2, 韦骏1,3,†   

  1. 1. 热带大气海洋系统科学教育部重点实验室, 中山大学大气科学学院, 广州 510275 2. 北京大学汇丰商学院, 深圳 518055 3. 广西大学海洋学院, 南宁 530004
  • 通讯作者: 韦骏, E-mail: weijun5(at)mail.sysu.edu.cn
  • 基金资助:
    广东省重点领域研发计划(2020B1111020003)、国家重点研发计划(2016YFA0202704)、广西壮族自治区特聘专家经费(2018B08)和国家自然科学基金(41976007, 91958101)资助

Abstract:

A LSTM (long-short term memory) model is applied to the prediction of the Nino3.4 index, and the spring prediction barrier (SPB) issue has been further investigated in the LSTM model. The results show that the model can predict the trend of the Nino3.4 index well, yet revealing different performance in different El Nino events. For the 1997/1998 El Nino and 2015/2016 El Nino, which are strong EP El Nino events, the model performes well on the prediction of Nino3.4 index trend and peaks, and anomaly correlation coefficient (ACC) reaches more than 0.93. But for the weak CP El Nino events, e.g. the 1991/1992 El Nino and 2002/2003 El Nino, it shows relatively poor performance on the prediction of the peak. In the growing period, the maximum season growth rate of prediction error are in AMJ quarter, which indicates obvious SPB phenomenon. However, in the decaying period, the maximum have similar distribution in the same type of events: for the weak CP El Nino events, the maximum are in AMJ quarter, indicating obvious SPB phenomenon; for strong EP El Nino events, the maximum are in other quarter, indicating that there is no SPB phenomenon. The differences in the performance among individuals may be related to the development characteristics of the event itself (such as event type and intensity).

Key words: LSTM, ENSO, prediction error, SPB, Nino3.4 index

摘要:

利用长短期记忆网络(LSTM)深度学习算法构建一个热带太平洋Nino3.4指数预测模型, 并分析模型的季节预报误差。结果表明, LSTM模型能够较好地预测厄尔尼诺事件的变化趋势, 但针对不同类型的厄尔尼诺事件有不同的表现。对于1997/1998和 2015/2016强东部型厄尔尼诺事件, 该模型能较准确地预测事件的趋势和峰值, 距平相关系数(ACC)达到0.93以上。但是, 对于 1991/1992和2002/2003弱中部型厄尔尼诺事件, 在峰值预测方面表现不好。在厄尔尼诺增长期, 预报误差的季节增长率最大值皆处于4—6月, 存在明显的春季预报障碍(SPB)现象。在衰减期, 同类型事件的季节增长率最大值分布相似: 弱中部型厄尔尼诺事件的最大值皆处于春季, 存在明显的SPB现象; 强东部型厄尔尼诺事件的最大值分散在其他季度, 不存在SPB现象。个体事件间存在一定的差异, 可能与事件的特征(如事件类型和强度)有关。

关键词: 长短期记忆人工神经网络(LSTM), ENSO, 预报误差, 春季预报障碍(SPB), Nino3.4指数