IOCAS-IR  > 海洋环流与波动重点实验室
基于深度学习的ENSO建模及其预测和可预报性研究
周路
Subtype博士
Thesis Advisor张荣华
2024-05-16
Degree Grantor中国科学院大学
Place of Conferral中国科学院海洋研究所
Degree Name理学博士
Keyword深度学习 ENSO预测 混合建模方法 多变量三维场预测 物理可解释性分析
Abstract

厄尔尼诺-南方涛动(El Niño-Southern Oscillation, ENSO)作为气候系统中最强的年际变率信号,可通过大气桥和海洋过程影响全球各地的天气和气候。因此,准确开展ENSO相关的海气耦合过程模拟和预测具有重要的科学和社会价值。当前,数据驱动的深度学习模型已成为ENSO预测研究的重要工具之一,其相比于动力模式显著改善了ENSO的春季预报障碍和跨年度准确预测问题,但也存在着计算复杂度高、预测任务单一和可解释性差等问题。围绕以上问题,本文基于深度学习方法系统地开展了“建模-应用-解释”等一系列创新研究:首先基于主振荡型分析(Principal Oscillation Pattern analysis, POP)和深度学习算法提出了一种计算高效、性能优越的混合型ENSO预测模型,以探索低成本、高性能预测模型建模方法;进一步,基于Transformer架构和时空自注意力机制发展了首个热带海气系统多变量三维(3D)场年际异常预测模型,以拓展ENSO智能预测模型的应用范围;最后基于所开发的多变量三维场智能预测模型开展了一系列初始场及变量敏感性试验以探究模型的物理可解释性和ENSO可预报性等。具体工作和结论如下:

1)发展了POP分析与深度学习模型相融合的混合型ENSO预测模型:POP-Net,为物理过程分析方法与深度学习的混合建模研究提供了新思路。该混合型模型利用POP方法从多变量高维时空场中提取具有特定周期和传播特性的ENSO相关模态信息,并将其编码融合进CNN-LSTM模型,从而改善ENSO预测技巧。结果显示,POP-NetNiño3.4区海表温度(Sea Surface Temperature, SST)异常的有效预测时长可达17个月,且大幅改善了春季预报障碍问题,其ENSO预测性能达到了先进的深度学习模型水平。此外,POP-Net相比于单独的POP统计模式和CNN-LSTM模型拥有更快的模型收敛速度和更高的ENSO预测精度,表明了在深度学习模型中显式地融入ENSO相关的大尺度年际异常模态信息可显著降低初始场噪音对模型的干扰并提高预测准确性。

2)基于Transformer架构和时空自注意力机制率先成功构建了数据驱动的热带海气系统多变量三维场年际异常预测模型:3D-Geoformer,突破了基于人工智能方法对ENSO相关多变量耦合过程的表征和预测瓶颈,推动ENSO智能预测实现了从单变量、单点时间序列预测到多变量、3D立体场预测的重要跨越。特别是,该研究针对海气变量场的强时间依赖性和空间非局地相关性等耦合特征,开发了适用于高维气候数据建模预测多头时空自注意力算法算法克服了卷积循环神经网络在构建空间遥相关关系和时序依赖关系方面不足,通过分别执行时间和空间注意力计算来一次性获取任意时空位置变量间的相关关系。3D-Geoformer对热带太平洋海表风应力场及上层海洋3D温度场的预测试验结果显示,该模型对赤道中西太平洋的风应力异常和热带太平洋大部分区域的3D海温异常的有效预测时长可达912个月以上,对Niño3.4SST异常的有效预测时长超过18个月此外,得益于3D-Geoformer出色的多变量3D场预测能力,模型可以至少提前6个月准确预测并区分两类El Niño事件中SST的异常时空演变模型的多变量和ENSO预测性能均明显优于当前先进的动力模式。以上结果表明,在数据驱动的3D-Geoformer中,通过以海气耦合的方式合理表征ENSO相关的Bjerknes反馈过程,可有效提升模型对多变量场协同演变过程的预测能力,这为进一步开展ENSO机制和可预报性研究提供了创新平台。

3)基于3D-Geoformer开展了初始场敏感性分析、海气过程表征合理性验证及ENSO海温前兆信号识别等创新研究,从技术和物理角度揭示了海气耦合建模对提高模型模拟和预测ENSO多变量时空演变过程的必要性,并探讨了ENSO跨年度可预报性来源。具体结论包括:1)模型的ENSO预测性能很大程度上取决于初始场输入数据的信噪比。一方面,适当延长初始场序列长度(Time Interval, TI)可为模型提供更多可预测信息,有助于模型捕捉与ENSO相关的长期海洋记忆信号,从而提升模型预测能力。另一方面,冗长的初始场信息会加剧输入场噪音对预测的干扰,同时受ENSO可预报性上限的限制,反而造成模型预测性能的降低。因此,寻求最适TI使初始场可预测信号的信息增益效应与噪音干扰的衰减效应之间取得平衡,是实现模型ENSO预测性能最优化的关键。2)模型能够合理表征多变量耦合关系并量化其对ENSO预测的影响:在海气变量对ENSO预测的贡献方面,除赤道海温异常的主导作用外,热带中太平洋海表风应力和赤道外次表层海温场也分别对ENSO的短期(<6个月)和跨年度>10个月)预测有重要影响。其中热带中太平洋海表风应力对北半球春夏季的ENSO预测有重要影响,反映了在模型中考虑风场的高频可预测信息及其与海温场的相互作用有助于改善ENSO预测的春季预报障碍问题;赤道外海温场对提前一年左右的ENSO预测贡献超过12%,凸显了赤道外海洋过程对ENSO长期预测的重要性。在海气耦合建模对复杂ENSO事件的预测影响方面,以2015-2016年超强El Niño事件和2020-2022年持续性La Niña事件的回报分析为例,验证了3D-Geoformer可以准确表征海气Bjerknes反馈过程,揭示了合理表征海气耦合过程对准确预测复杂ENSO事件强度和时空演变的必要性。3)沿赤道-赤道外北太平洋传播的海温年际异常信号可作为ENSO预测的海温前兆因子,为模型准确开展ENSO跨年度预测提供了物理基础利用3D-Geoformer和变量敏感度分析方法成功识别了上层海温年际异常信号沿赤道和赤道外北太平洋的逆时针环形演变路径路径由赤道和10°N太平洋区域的纬向路径及太平洋东、西边界的经向路径首尾相接而成。本文详细描述了观测中热带太平洋上层海温异常信号沿该路径的演变特征,以3D-Geoformer3D演变过程的表征能力,阐述了其与ENSO起源的关系,探究了其作为海温前兆因子ENSO跨年度预测中的作用,为3D-Geoformer准确开展ENSO跨年度预测提供了合理的物理解释

综上,本文针对ENSO智能预测模型在计算复杂度、预测泛化能力和可解释性等方面存在的问题,开展了混合型模型开发、多变量三维场气候预测模型构建以及ENSO可预报性分析等创新研究,系统地探讨了深度学习技术在海气系统建模、ENSO过程表征及预测中的前沿进展及应用分析。本论文对于推动物理与深度学习融合建模,突破数据驱动的高维时空数据预测难题,促进深度学习模型的可解释性研究都具有重要的科学和应用价值。

Other Abstract

El Niño-Southern Oscillation (ENSO) is the most prominent variation on interannual climate variability in the climate system, which influences the global weather and climate system by atmospheric bridge and ocean pathways. Therefore, accurately simulating and predicting ENSO-related ocean-atmosphere conditions has substantial scientific and societal value. Currently, data-driven deep learning (DL) models have been one of the crucial tools for process understanding and ENSO predictions, with skillful ENSO prediction at lead times of more than one year has been achieved and the spring predictability barrier (SPB) being substantially alleviated. However, challenges still exist in ENSO simulations and predictions using DL-based methods, such as high computational complexity, weak prediction generalization ability, and lack of interpretability. This dissertation systematically conducts a series of innovative research using DL-based methods, ranging from modeling to application and further to interpretative analyses: firstly, a computationally efficient hybrid approach is proposed by combining process-based principal oscillation pattern (POP) analysis with DL model for long-lead ENSO predictions. Additionally, to construct a multitask prediction model, a specific self-attention-based DL model is developed for three-dimensional (3D) multivariate predictions of ENSO based on the much sought-after Transformer architecture. Finally, a series of sensitivity experiments are conducted based on the multivariate prediction model to investigate its physical interpretability and the sources of ENSO predictability. The main work and conclusions are as follows:

(1) A novel hybrid model, named POP-Net, is developed for improved ENSO predictions by combining the POP analysis procedure with the CNN-LSTM algorithm. In this hybrid configuration, the POP-based pre-processing acts to extract the periodically varying ENSO-related patterns and corresponding temporal coefficients of multivariate fields while filtering unrelated noise. Next, the POP-related temporal coefficients are embedded and fed into the CNN-LSTM model for further optimization procedure to obtain the final output as predictions. Consequently, an improved prediction is achieved in the POP-Net relative to POP and CNN-LSTM models. The results indicate that the POP-Net shows a high-correlation skill for Niño3.4 sea surface temperature (SST) anomaly prediction made 17 months in advance and alleviates the SPB problem. The ENSO prediction skills are comparable to the state-of-the-art DL-based models. Otherwise, the results indicate that this new hybrid model achieved an enhanced prediction capability without using an overly complex and advanced DL algorithm, nor did it use too many training manipulations. It indicates that value-added artificial neural networks for improved ENSO predictions are possible by including process-oriented analyses to enhance signal representations.

(2) An innovative self-attention-based DL model, named 3D-Geoformer, is developed for spatiotemporal multivariate predictions of ENSO by incorporating multiheaded spatiotemporal attention modules into the Transformer architecture. The 3D-Geoformer signifies a major advancement in DL-based ENSO predictions, leaping from previous predictions that focused on specific regional ENSO indices or a single variable to predicting 3D upper-ocean temperature anomalies and wind stress anomalies. In particular, the 3D-Geoformer eliminates convolution and recurrence operations and replaces them with self-attention mechanisms to establish multivariable relationships in parallel regardless of their spatial and temporal distances. The unique properties of the self-attention mechanism reinforce the nonlocal long-range modeling ability to establish multivariate interconnections by performing attention analyses on the temporal and spatial axes separately. Consequently, it is particularly suitable for capturing the multivariate coupling dynamics among atmospheric and oceanic anomalies during ENSO evolution. To quantitatively assess the prediction skills of 3D-Geoformer, multivariable anomaly correlation skills are calculated for the spatial distribution and Niño indices. The correlation skill of upper-ocean temperature fields remains consistently above 0.5 over most of the tropical Pacific for 12-month lead time predictions, and the wind stress anomalies over the central-western equatorial Pacific can be reasonably predicted at lead times exceeding 9 months. Furthermore, the model achieves high correlation skills for predicting Niño3.4 SST anomalies with an 18-month lead time and exhibits the capability to distinguish among different types of central Pacific and eastern Pacific El Niño at lead times exceeding 6 months, outperforming most statistical or dynamical models. These effective and accurate predictions using the 3D-Geoformer can be partly attributed to adequate representations of the ocean-atmosphere system in a coupled manner, which is in accordance with the Bjerknes feedback mechanism. Overall, the 3D-Geoformer overcomes current limitations in characterizing and predicting ENSO-related ocean-atmosphere coupled processes using data-driven methods and provides a crucial tool for process understanding and ENSO predictions.

(3) A suite of sensitivity experiments is conducted using the 3D-Geoformer to make robust connections between model interpretability and dynamical processes, including the sensitivity of model initial conditions, the importance of ocean-atmosphere coupled modeling, and the predictability sources for long-lead ENSO predictions. The main conclusions include: 1) The ENSO prediction performance of the 3D-Geoformer largely depends on the signal-to-noise ratio of the input predictors. Generally, including more information on 3D temperature and surface wind fields during long times intervals (TI) in the initial conditions tends to precondition future evolution more effectively, thus enhancing the predictability for SST in the tropical Pacific. However, the lengthy information in the initial condition also exacerbates noise interference, which in turn results in the degradation of model prediction performance. Consequently, determining the optimal TI value that strikes a balance between the skill-increase-effect gained from predictable signals in the initial fields and the skill-decrease-effect owing the noise interference is critical for achieving optimal ENSO prediction performance using the 3D-Geoformer. 2) The 3D-Geoformer is able to correctly characterize the coupled ocean-atmosphere dynamics and reasonably quantify their effects on ENSO predictions. The intrinsic properties and fundamental coupling in the ocean-atmosphere system with differently time scale-dependent effects provide predictable information for ENSO predictions at different lead times. Regarding the contributions of multivariable on ENSO predictions, the wind stress anomalies over the central Pacific and off-equatorial temperature anomalies in predictors makes significant contributions on short-term (<6 months) and long-term (>10 months) ENSO predictions, respectively. Specifically, the effects of central Pacific wind stress greatly enhances the ENSO predictions for the boreal spring and summer seasons, which are the target seasons with low prediction skills in the 3D-Geoformer. To some extent, the alleviation of the SPB problem in the model can be partly attributed to the effect of wind stress in the modeling. In contrast, the low-frequency ocean memory contained in the off-equatorial temperature anomalies has larger effects than of the wind stress on long-term ENSO predictions. Its contribution to the ENSO prediction skills is more than 12% for the prediction made one-year in advance, implying that the long-term ENSO predictions using the 3D-Geoformer rely more on the temperature anomalies than wind stress predictors. In terms of the effects of ocean-atmosphere coupling process on ENSO predictions, the case analyses of the 2015-2016 super El Niño and 2020-2022 multi-year La Niña hindcasts demonstrate that the prediction skills for complex ENSO events can be enhanced with the multivariate synergic ocean-atmosphere dynamics represented following the Bjerknes feedback mechanism during ENSO cycles. 3) ENSO-related thermal precursors serving as initial conditions during multi-month time intervals are identified in the equatorial-northern Pacific, preconditioning the input predictors to provide for long-lead ENSO predictability. Results reveal the existence of upper-ocean temperature anomaly pathways and consistent phase propagations of thermal precursors around the tropical Pacific: eastward along the equator, westward across the off-equatorial tropical North Pacific, and apparent meridional phase connections both in the western and eastern boundaries. From the prediction perspective, it is illustrated that 3D thermal fields and their basinwide evolution during long TIs act to enhance long-lead ENSO prediction skills. These physically explainable results indicate that neural networks can represent predictable precursors in the input predictors for successful ENSO predictions.

In summary, this dissertation examines the challenges posed by DL-based ENSO prediction models regarding computational complexity, generalization capability, and physical interpretability by conducting innovative research on DL algorithm development and ENSO predictability analyses. It systematically explores the frontier advancements and applications of data-driven methods in high-dimensional multivariate modeling within geoscience and ENSO-related process predictions. These results are of great scientific and practical implications for advancing hybrid modeling, facilitating predictions of high-dimensional multivariate fields, and enhancing the interpretability of DL models.

MOST Discipline Catalogue理学::海洋科学
Language中文
Table of Contents

第1章 绪论    1
1.1 研究背景和意义    1
1.1.1 ENSO预测的物理基础    3
1.1.2 用于ENSO预测的数值与统计模式    4
1.2 基于人工智能的海气耦合系统建模及ENSO预测研究    7
1.2.1 人工智能在海气系统建模及ENSO预测中的应用    7
1.2.2 ENSO智能预测的可解释性研究    10
1.3 主要研究内容    12
1.4 本文结构    13
第2章 资料与方法    15
2.1 数据介绍    15
2.1.1 再分析数据    16
2.1.2 CMIP6模拟数据    17
2.2 用于ENSO分析的统计方法介绍    19
2.2.1 主振荡型分析方法    19
2.2.2 预测技巧评估方法    20
第3章 基于深度学习的混合型模型及对ENSO的预测研究    23
3.1 引言    23
3.2 基于POP的线性预测模型    23
3.3 POP-Net:融合物理分析方法和神经网络的混合型ENSO预测模型    27
3.3.1 POP-Net模型架构与形式化表达    29
3.3.2 损失函数与模型训练    30
3.3.3 模型性能评估    31
3.4 本章小结    36
第4章 热带海气系统多变量三维场智能表征和ENSO预测研究    39
4.1 引言    39
4.2 3D-Geoformer:基于时空注意力机制的多变量三维场气候预测模型    40
4.2.1 3D-Geoformer模型架构及形式化表达    40
4.2.2 训练数据集与损失函数    46
4.2.3 变量敏感度分析方法    47
4.2.4 模型性能评估    48
4.3 本章小结    67
第5章 基于3D-Geoformer对ENSO可解释性及可预报性的进一步应用    69
5.1 引言    69
5.2 初始场对模型ENSO预测能力的影响    70
5.3 热带太平洋风场及次表层海温场对ENSO预测的影响    72
5.3.1 海表风场在模型预测中的作用    72
5.3.2 次表层海温场在模型预测中的作用    80
5.4 热带太平洋海温前兆信号识别及其对ENSO跨年度预测的影响    84
5.4.1 赤道-赤道外海温年际异常演变路径的观测特征    84
5.4.2 3D-Geoformer对上层海温异常演变路径的表征及其对ENSO可预报性的影响    89
5.4.3 ENSO相关的海温前兆信号传播路径的表征    90
5.4.4 海温前兆因子对ENSO跨年度可预报性的影响    93
5.5 本章小结    96
第6章 总结与展望    97
6.1 研究成果总结    97
6.2 本文创新点    100
6.3 不足与展望    100
参考文献    103
致谢    115
作者简历及攻读学位期间发表的学术论文与其他相关学术成果    117

Document Type学位论文
Identifierhttp://ir.qdio.ac.cn/handle/337002/185222
Collection海洋环流与波动重点实验室
Recommended Citation
GB/T 7714
周路. 基于深度学习的ENSO建模及其预测和可预报性研究[D]. 中国科学院海洋研究所. 中国科学院大学,2024.
Files in This Item:
File Name/Size DocType Version Access License
周路-202218006812032-基(8710KB)学位论文 延迟开放CC BY-NC-SAView 2025-7-1后可获取
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[周路]'s Articles
Baidu academic
Similar articles in Baidu academic
[周路]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[周路]'s Articles
Terms of Use
No data!
Social Bookmark/Share
File name: 周路-202218006812032-基于深度学习的ENSO建模及其预测和可预报性研究-最终版.pdf
Format: Adobe PDF
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.