Prediction and correction of ENSO using an intelligent Air-Sea coupling model based on the Transformer architecture
-
摘要: 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，可对全球的天气和气候产生重要的影响。在全球变暖下，ENSO的演变愈发呈现出复杂、多样的特征，其模拟与预测已成为气候领域极具挑战性的课题。本研究引入基于 Transformer 架构开发的热带海气系统多变量智能预测模型—3D-Geoformer，开展 ENSO 预测的误差分析及订正研究。3D-Geoformer 模型不同于多数智能模型的仅对 ENSO 相关的单变量场或时间序列进行预测，实现了对热带太平洋海气系统多变量三维场的准确表征和预测，保证了 ENSO 预测所需物理过程的完整性。同时，本文针对3D-Geoformer 模型在ENSO中存在的春季预测技巧低、赤道西太平洋海温预测能力较弱和极端 ENSO 事件预测强度偏低等问题，提出了基于经验正交分解（EOF）的季节预测误差订正技术，并应用于对3D-Geoformer预测结果的订正检验。在订正关系构建阶段，通过对 1983～2009 年的多变量预测场和预测误差场进行EOF分析，构建二者主成分序列间的线性关系，并用于后续误差订正。在测试阶段，利用预测场的EOF主成分系数以及与误差场主成分的线性关系，便可算出对应的预测误差场主成分，进而得到预测误差场和校正的预测场。结果显示，使用3D-Geoformer模型对赤道西太平洋海表温度（SST）预测时，预测误差在0.15℃ 以下；赤道中东太平洋SST预测误差缩减46.7%。通过比较EOF订正前后的3D-Geoformer模型对赤道太平洋SST预测结果的异常相关系数（ACC）的差值，结果发现，ACC的差值均有正值区，表明经过EOF订正后的模型预测准确度提高，且优化了3D-Geoformer模型在训练过程中使用第六次耦合模式比较计划（CMIP6）的气候模式数据引起的“冷舌偏差”问题。模型对提前12个月对2015～2016年El Niño的预测订正结果显示，赤道西太平洋地区SST误差控制在0.5℃以内，赤道东太平洋SST预测误差减小约75%，误差范围缩至±0.5℃以内。本研究揭示了基于EOF分解的季节预测误差订正方法在改善模式预测中的应用价值，为进一步提高智能模型预测ENSO的精度提供了新方法，也为地球科学领域相关的模拟预测、误差分析研究提供了新思路。
-
关键词:
- ENSO预测 /
- EOF统计订正 /
- Transformer /
- 3D-Geoformer
Abstract: El Niño-Southern Oscillation (ENSO), as the most prominent interannual variability signal in the climate system, exerts significant impacts on global weather and climate. Under global warming, ENSO evolution has increasingly exhibited the characteristics of complex and diverse rendering its simulation and prediction a particularly challenging subject within climatology. This study introduces 3D-Geoformer, an advanced multi-variable intelligent prediction model for the tropical sea-air system based on Transformer architecture, to conduct error analysis and correction research for ENSO predictions. Unlike many existing models that focus solely on univariate fields or time series related to ENSO, the 3D-Geoformer model achieves accurate characterization and prediction of the multi-variable three-dimensional field of the tropical Pacific sea-air system while preserving the integrity of the physical processes essential for ENSO prediction. To address specific issues in ENSO predictions by the 3D-Geoformer model, such as low spring forecasting skills, weak SST forecasting ability in the western equatorial Pacific, and inadequate forecasting intensity for extreme ENSO events, this study proposes a seasonal forecasting error correction technique based on empirical orthogonal function (EOF) decomposition. This method is applied to correct the prediction results of the 3D-Geoformer model. During the construction phase of the correction relationship, EOF analysis was used to establish the linear relationship between the principal component sequences of the multivariable prediction field and the prediction error field from 1983 to 2009. Subsequently, this relationship was utilized for subsequent error corrections. In the testing phase, the EOF principal component coefficients of the prediction field and their linear relationships with the main components of the error were employed to calculate the corresponding principal components of the prediction error, thereby obtaining the prediction error field and the corrected prediction field. The experimental results indicate that when the 3D-Geoformer model is employed for forecasting the sea - surface temperature (SST) in the western equatorial Pacific, the prediction error remains below 0.15 °C. Notably, the prediction bias of the 3D-Geoformer model regarding the sea temperature in the western equatorial Pacific, induced by the “cold tongue bias” inherent in climate models, is substantially mitigated. Concurrently, there is a remarkable 46.7% reduction in the prediction error of the sea-surface temperature (SST) in the central and eastern equatorial Pacific. Through a meticulous comparison of the disparities in the anomaly correlation coefficients (ACC) between the SST prediction outcomes of the 3D-Geoformer model with and without Empirical Orthogonal Function (EOF) correction in the equatorial Pacific, it is discerned that positive-value regions are consistently present in the ACC differences. This finding strongly suggests that the EOF-corrected model exhibits enhanced prediction accuracy, effectively alleviating the “cold tongue bias” issue arising from the utilization of climate model data from the Sixth Coupled Model Inter-comparison Project (CMIP6) during the training phase of the 3D-Geoformer model. For the 2015−2016 El Niño event, forecast corrections made 12 months in advance show that the SST error in the western equatorial Pacific is controlled within 0.5°C, and the SST error in the eastern equatorial Pacific is reduced by approximately 75%, with the error range narrowed to within ±0.5°C. This study underscores the application value of the seasonal forecast error correction method based on EOF decomposition in enhancing model prediction accuracy, providing a novel approach to improving the precision of ENSO intelligent predictions, and offering new insights into simulation prediction and error analysis in earth science.-
Key words:
- ENSO forecast /
- statistical revision of EOF /
- Transformer /
- 3D-Geoformer
-
图 1 用于ENSO预测的3D-Geoformer模型结构。模型整体为编码器-解码器结构，编码器和解码器均基于多层叠加的时空自注意力算法，输入连续多个月的多变量异常场通过编码器提取特征矩阵，解码器利用该特征矩阵和已生成的预测场滚动预测下个月的变量场
Fig. 1 The architecture of the 3D-Geoformer model for ENSO prediction. The overall structure of the model follows an encoder-decoder framework. Both the encoder and the decoder are grounded on the multi-layer stacked spatio-temporal self-attention algorithm. Multiple consecutive months of multi-variable anomaly fields are inputted, and the encoder extracts the feature matrix therefrom. Subsequently, the decoder utilizes this feature matrix along with the previously generated prediction fields to perform a rolling prediction of the variable fields for the next month
图 2 (a)图为3D-Geoformer模型的滚动预测流程图。输入连续12个月的多变量异常场，每一步预测未来1个月的预测场。每一步预测中都会考虑前12个月的所有信息，重复滚动预测直至得到未来20个月的预测场；(b)图为 Niño3.4指数的预测相关系数随起报月份和提前预报时间的分布
Fig. 2 (a) Flowchart of the rolling prediction process of the 3D-Geoformer model. Multivariable anomaly fields for 12 consecutive months are inputted, and the model predicts the fields for the upcoming 1 month. The input fields extract spatio-temporal information via the spatio-temporal self-attention module of the encoder and compress it into a feature matrix. Subsequently, the decoder predicts the multivariable fields for the next month using the feature matrix and the previously generated prediction fields. In each prediction step, all information from the previous 12 months is taken into account, and the rolling prediction is repeated until the prediction fields for the next 20 months are obtained. (b) Distribution of the prediction correlation coefficients of the Niño3.4 index with respect to the starting prediction month and the lead time
图 3 3D-Geoformer模型预测的赤道SST误差场随提前预测时间的分布；a−d分别为以1月、4月、7月和10月为目标月份的结果；时间为1983−2009年
Fig. 3 Distribution of the Equatorial SST Error Field Predicted by the 3D-Geoformer Model with Lead Prediction Time. Panels a−d show the results with January, April, July, and October as the target months, respectively. The time period is from 1983 to 2009
图 4 同图3，但为3D-Geoformer模型预测结果EOF订正后的赤道SST误差场
Fig. 4 Similar to Figure 3, but showing the equatorial sea surface temperature error fields after Empirical Orthogonal Function correction of the prediction results from the 3D-Geoformer model
图 5 提前6−12月的预报，EOF订正后的 (a−c) 纬向风 (
$ {\tau }_{x} $ )，(d−f) 经向风（$ {\tau }_{y} $ ）和 (g−i) SST的预测异常相关系数与未订正的预测异常相关系数（ACC）的差值。时间为1983−2009年Fig. 5 For the forecasts with a lead time of 6−12 months, the differences between the predicted anomaly correlation coefficients of (a−c) zonal wind (taux), (d−f) meridional wind (tauy), and (g−i) sea surface temperature after Empirical Orthogonal Function correction and those of the uncorrected predicted anomaly correlation coefficients. The time span is from 1983 to 2009
图 6 在提前6−12月的预报时间中，经过EOF订正后次表层海温与对应的未订正的次表层海温的ACC差值
Fig. 6 For the forecasts with a lead time of 6−12 months, the anomaly correlation coefficients differences between the subsurface sea temperature after Empirical Orthogonal Function correction and the corresponding uncorrected subsurface sea temperature
图 7 (a)热带太平洋2012−2016年SST异常的时空演变。海温异常来自于GODAS再分析数据（单位：℃）；(b) 2014−2016年再分析数据（黑线）、IRI模式预测（蓝线和绿线分别为IRI动力模式和统计模式集合平均的预测值）和3D-Geoformer模型（红线）的Niño3.4区SST异常演变。每条彩色线为模型从该月起报得到的未来9个月的预测值
Fig. 7 (a) Spatio-temporal evolution of the sea surface temperature (SST) anomaly map in the tropical Pacific from 2012 to 2016. The SST anomalies are derived from the GODAS reanalysis data (in units of °C). (b) Evolution of the SST anomalies in the Niño3.4 region from 2014 to 2016, including the reanalysis data (black line), the predictions of the IRI model (the blue and green lines represent the ensemble-averaged predicted values of the IRI dynamical and statistical models, respectively), and the 3D-Geoformer model (red line). Each colored line represents the predicted values for the next nine months starting from the corresponding month by the respective model
图 8 (a−f) 2015年4月−2016年2月期间热带太平洋海表风应力异常（箭头）和SST异常（填色）的时空演变。数据来自于GODAS再分析数据；(g−l) 3D-Geoformer模型从2015年4月起报得到的同期的海表风应力异常和SST异常（填色）的时空演变
Fig. 8 (a−f) Spatio-temporal evolution of sea surface wind stress anomalies (vectors) and sea surface temperature anomalies (shading) in the tropical Pacific from April 2015 to February 2016. The data are sourced from the GODAS reanalysis data. (g−l) Spatio-temporal evolution of sea surface wind stress anomalies and sea surface temperature anomalies (shading) during the same period, which are predicted by the 3D-Geoformer model starting from April 2015
图 9 3D-Geoformer模型提前12个月对2015年6月−2016年1月热带太平洋SST预测的误差分布（单位：°C）
Fig. 9 Error distribution of the 3D-Geoformer model's 12-month-ahead predictions for the sea surface temperature in the tropical Pacific from June 2015 to January 2016. Unit is °C for temperature
图 10 同图9， 但为经过EOF订正后的3D-Geoformer模型提前12个月对2015年6月−2016年1月热带太平洋SST预测的误差分布（单位：°C）
Fig. 10 Similar to Figure 9, but showing the error distribution of the 3D-Geoformer model's 12-month-ahead predictions for the sea surface temperature in the tropical Pacific from June 2015 to January 2016 after Empirical Orthogonal Function correction. Unit is °C for temperature
表 1 本文所用到的CMIP6模式信息
Tab. 1 The information of the CMIP6 models employed in this paper
序号 模式名称 Member_id table_id 时间 1 BCC-ESM1 rlilplf1
Monthly ocean
data(Omon)
1850−20142 CanESM5 rlilplf1 3 CESM2-FV2 rlilplf1 4 CNRM-CM6-1 rlilplf2 5 GISS-E2-1-G rlilplf1 6 INM-CM4-8 rlilplf1 7 MCM-UA-1-0 rlilplf1 8 MPI-ESM1-2-HR rlilplf1 9 NorCPM1 rlilplf1 10 NorESM2-LM rlilplf1 11 UKESM1-0-LL rlilplf2 12 AWI-CM-1-1-MR rlilplf1 13 CAMS-CSM1-0 rlilplf1 14 E3SM-1-1 rlilplf1 15 FGOALS-F3-L rlilplf1 16 GFDL-CM4 rlilplf1 17 HadGEM3-GC31-LL rlilplf3 18 IPSL-CM6A-LR rlilplf1 19 MIROC-ES2L rlilplf2 20 MIROC6 rlilplf1 21 MRI-ESM2-0 rlilplf1 22 NESM3 rlilplf1 23 SAM0-UNICON rlilplf1 
下载: 导出CSV
-
[1] Zhang Ronghua, Rothstein L M, Busalacchi A J. Origin of upper-ocean warming and El Niño change on decadal scales in the tropical Pacific Ocean[J]. Nature, 1998, 391(6670): 879−883. doi: 10.1038/36081 [2] Capotondi A, Wittenberg A T, Newman M, et al. Understanding ENSO diversity[J]. Bulletin of the American Meteorological Society, 2015, 96(6): 921−938. doi: 10.1175/BAMS-D-13-00117.1 [3] Zhang Ronghua, Yu Yongqiang, Song Zhenya, et al. A review of progress in coupled ocean-atmosphere model developments for ENSO studies in China[J]. Journal of Oceanology and Limnology, 2020, 38(4): 930−961. doi: 10.1007/s00343-020-0157-8 [4] 陈明诚. 厄尔尼诺和拉尼娜生命史不对称机理研究[J]. 气象学报, 2017, 25(2): 67−70. (查阅网上资料, 未找到本条文献信息, 请确认)Chen Mingcheng. Research on the asymmetric mechanisms of the life cycles of El Niño and La Niña[J]. Journal of Meteorological Research, 2017, 25(2): 67−70. [5] 高川, 陈茂楠, 周路, 等. 2020—2021年热带太平洋持续性双拉尼娜事件的演变[J]. 中国科学: 地球科学, 2022, 52(12): 2353-2372.Gao Chuan, Chen Maonan, Zhou Lu, et al. The 2020–2021 prolonged La Niña evolution in the tropical Pacific[J]. Science China Earth Sciences, 2022, 65(12): 2248-2266. [6] Yeh S W, Kug J S, Dewitte B, et al. El Niño in a changing climate[J]. Nature, 2009, 461(7263): 511−514. doi: 10.1038/nature08316 [7] Timmermann A, An S I, Kug J S, et al. El Niño-Southern Oscillation complexity[J]. Nature, 2018, 559(7715): 535−545. doi: 10.1038/s41586-018-0252-6 [8] Zhang Ronghua, Gao Chuan, Feng Licheng. Recent ENSO evolution and its real-time prediction challenges[J]. National Science Review, 2022, 9(4): nwac052. doi: 10.1093/nsr/nwac052 [9] 张荣华. 用于厄尔尼诺-南方涛动(ENSO)研究的海气耦合模式综述: 中间型和混合型模式[J]. 海洋与湖沼, 2024, 55(1): 1−23.Zhang Ronghua. A review of progress in coupled ocean-atmosphere model developments for Enso studies: intermediate coupled models and hybrid coupled models[J]. Oceanologia et Limnologia Sinica, 2024, 55(1): 1−23. [10] Chen Dake, Cane M A, Kaplan A, et al. Predictability of El Niño over the past 148 years[J]. Nature, 2004, 428(6984): 733−736. doi: 10.1038/nature02439 [11] Tang Youmin, Zhang Ronghua, Liu Ting, et al. Progress in ENSO prediction and predictability study[J]. National Science Review, 2018, 5(6): 826−839. doi: 10.1093/nsr/nwy105 [12] Eyring V, Bony S, Meehl G A, et al. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization[J]. Geoscientific Model Development, 2016, 9(5): 1937−1958. doi: 10.5194/gmd-9-1937-2016 [13] Luo Jingjia, Masson S, Behera S K, et al. Extended ENSO predictions using a fully coupled ocean–atmosphere model[J]. Journal of Climate, 2008, 21(1): 84−93. doi: 10.1175/2007JCLI1412.1 [14] Ding Ruiqiang, Tseng Y H, Di Lorenzo E, et al. Multi-year El Niño events tied to the North Pacific Oscillation[J]. Nature Communications, 2022, 13(1): 3871. doi: 10.1038/s41467-022-31516-9 [15] Zhu Jieshun, Kumar A, Huang Bohua, et al. The role of off -equatorial surface temperature anomalies in the 2014 El Niño prediction[J]. Scientific Reports, 2016, 6(1): 19677. doi: 10.1038/srep19677 [16] Dong Changming, Xu Guangjun, Han Guoqing, et al. Recent developments in artificial intelligence in oceanography[J]. Ocean-Land-Atmosphere Research, 2022, 2022: 9870950. [17] Zheng Gang, Li Xiaofeng, Zhang Ronghua, et al. Purely satellite data-driven deep learning forecast of complicated tropical instability waves[J]. Science Advance, 2020, 6(29): eaba1482. [18] Zhu Yuchao, Zhang Ronghua, Moum J N, et al. Physics-informed deep-learning parameterization of ocean vertical mixing improves climate simulations[J]. National Science Review, 2022, 9(8): nwac044. doi: 10.1093/nsr/nwac044 [19] Zhou Lu, Zhang Ronghua. A hybrid neural network model for ENSO prediction in combination with principal oscillation pattern analyses[J]. Advances in Atmospheric Sciences, 2022, 39(6): 889−902. doi: 10.1007/s00376-021-1368-4 [20] Taylor J, Feng Ming. A deep learning model for forecasting global monthly mean sea surface temperature anomalies[J]. Frontiers in Climate, 2022, 4: 932932. doi: 10.3389/fclim.2022.932932 [21] Gong Bing, Langguth M, Ji Yan, et al. Temperature forecasting by deep learning methods[J]. Geoscientific Model Development, 2022, 15(23): 8931−8956. doi: 10.5194/gmd-15-8931-2022 [22] Mu Bin, Cui Yuehan, Yuan Shijin, et al. Simulation, precursor analysis and targeted observation sensitive area identification for two types of ENSO using ENSO-MC v1.0[J]. Geoscientific Model Development, 2022, 15(10): 4105−4127. doi: 10.5194/gmd-15-4105-2022 [23] Devlin J, Chang Mingwei, Lee K, et al. BERT: pre-training of deep bidirectional transformers for language understanding[C]//Proceedings of 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Minneapolis: ACL, 2019. [24] Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16×16 words: transformers for image recognition at scale[C]//Proceedings of the 9th International Conference on Learning Representations. OpenReview. net, 2021. (查阅网上资料, 未找到对应的出版地信息, 请确认) [25] Zhou Lu, Zhang Ronghua. A self-attention–based neural network for three-dimensional multivariate modeling and its skillful ENSO predictions[J]. Science Advances, 2023, 9(10): eadf2827. doi: 10.1126/sciadv.adf2827 [26] Ham Y G, Kim J H, Kim E S, et al. Unified deep learning model for El Niño/Southern oscillation forecasts by incorporating seasonality in climate data[J]. Science Bulletin, 2021, 66(13): 1358−1366. doi: 10.1016/j.scib.2021.03.009 [27] 秦正坤, 林朝晖, 陈红, 等. 基于EOF/SVD的短期气候预测误差订正方法及其应用[J]. 气象学报, 2011, 69(2): 289−296. doi: 10.11676/qxxb2011.024Qin Zhengkun, Lin Zhaohui, Chen Hong, et al. The bias correction methods based on the EOF/SVD for short-term climate prediction and their applications[J]. Acta Meteorologica Sinica, 2011, 69(2): 289−296. doi: 10.11676/qxxb2011.024 [28] Yu Yue, Lin Zhaohui, Qin Zhengkun. Improved EOF-based bias correction method for seasonal forecasts and its application in IAP AGCM4.1[J]. Atmospheric and Oceanic Science Letters, 2018, 11(6): 499−508. doi: 10.1080/16742834.2018.1529532 [29] 姚方玲, 秦正坤, 林朝晖, 等. 基于旋转经验正交分解的流域降水气候预测误差订正方法[J]. 气候与环境研究, 2023, 28(3): 327−342. doi: 10.3878/j.issn.1006-9585.2022.22071Yao Fangling, Qin Zhengkun, Lin Zhaohui, et al. Bias correction method based on rotated empirical orthogonal function for seasonal precipitation prediction on basin scale[J]. Climatic and Environmental Research, 2023, 28(3): 327−342. doi: 10.3878/j.issn.1006-9585.2022.22071 [30] Zhang Ronghua, Zhou Lu, Gao Chuan, et al. A transformer-based coupled ocean-atmosphere model for ENSO studies[J]. Science Bulletin, 2024, 69(15): 2323−2327. doi: 10.1016/j.scib.2024.04.048 [31] Newman M, Sardeshmukh P D. Are we near the predictability limit of tropical Indo-Pacific sea surface temperatures?[J]. Geophysical Research Letters, 2017, 44(16): 8520−8529. doi: 10.1002/2017GL074088 [32] 郑依玲, 陈泽生, 王海, 等. 2015/2016年超强厄尔尼诺事件基本特征及生成和消亡机制[J]. 热带海洋学报, 2019, 38(4): 10−19.Zheng Yiling, Chen Zesheng, Wang Hai, et al. Features of 2015/2016 extreme El Niño event and its evolution mechanisms[J]. Journal of Tropical Oceanography, 2019, 38(4): 10−19. [33] Rivera Tello G A, Takahashi K, Karamperidou C. Explained predictions of strong eastern Pacific El Niño events using deep learning[J]. Scientific Reports, 2023, 13(1): 21150. doi: 10.1038/s41598-023-45739-3 [34] Xu Zichun, Zhao Yang, Deng Zhongwen. The possibilities and limits of AI in Chinese judicial judgment[J]. AI & Society, 2022, 37(4): 1601−1611. -
计量
- 文章访问数: 19
- HTML全文浏览量: 6
- PDF下载量: 3
- 被引次数: 0
