A cuboid attention-based IOD prediction model
-
摘要: 印度洋偶极子（Indian Ocean Dipole, IOD）是热带印度洋海表温度呈现东西反位相变化的主要气候模态，对区域和全球气候演变具有重要影响。现有的IOD预测方法多依赖于多变量耦合或传统统计模型，存在计算复杂度高、多变量噪声干扰等挑战。针对这些挑战，本文提出了一种基于立方体注意力机制的IOD预测模型（Cuboid Attention-based IOD Prediction Model, CAIPM）。该模型仅以海表温度异常（Sea Surface Temperature Anomaly, SSTA）作为输入变量，加入融合了滑动窗口、时序差分与空间卷积的时空梯度增强模块（Spatio-Temporal Gradient Enhancement Module, STGEM），以强化时空特征提取，通过立方体注意力结构有效捕捉SSTA场中的时空依赖关系，直接输出未来时空范围的SSTA预测结果，进而计算出IOD指数。实验表明，CAIPM在IOD指数预测中显著优于传统统计方法及当前主流深度学习模型，CAIPM在提前12个月的IOD指数预测中，其皮尔逊相关系数相较于CNN、CNN-LSTM、TCN和ConvLSTM模型分别取得了约32%、22%、16%和6%的性能提升。Abstract: The Indian Ocean Dipole (IOD) is the dominant climate mode in the tropical Indian Ocean, characterized by an east–west dipole pattern in sea-surface temperature anomalies that exerts substantial influence on regional and global climate variability. Current IOD forecasting methods predominantly rely on multivariate coupled or traditional statistical models, which pose significant challenges such as high computational complexity and multivariate noise interference. To address these challenges, this study proposes a cuboid attention-based IOD prediction model(CAIPM). The model takes sea surface temperature anomalies as the sole input variable, incorporates a Spatio-Temporal Gradient Enhancement Module (STGEM), which integrates sliding windows, temporal difference, and spatial convolution operations to enhance spatiotemporal feature extraction. By effectively capturing spatiotemporal dependencies within the Sea Surface Temperature Anomaly (SSTA) field through a cuboid attention mechanism, it directly outputs future spatiotemporal SSTA predictions, from which the IOD index is subsequently calculated. Experimental results demonstrate that the CAIPM significantly outperforms traditional statistical methods and current mainstream deep learning models in predicting the IOD index. Specifically, for a 12-month lead prediction, the Pearson correlation coefficient (PCC) of CAIPM is 32%, 22%, 16%, and 6% higher than that of the CNN, CNN-LSTM, TCN, and ConvLSTM models, respectively.
-
图 5 超参数(cuboid_size + shift)对模型误差指标的影响
Fig. 5 Effect of hyperparameters (cuboid_size + shift) on model error metrics
图 6 CAIPM预测2023年1–12月的DMI指数与DMI指数真实值比较
Fig. 6 Comparison of predicted DMI Index by CAIPM model and actual DMI Index from January 2023 to December 2023
表 1 CMIP6数据信息
Tab. 1 CMIP6 data information
序号 模式名称 所属国 原始分辨率 研发机构 1 ACCESS-ESM1-5 澳大利亚 1° × 1° CSIRO 2 FGOALS-g3 中国 1° × 1° CAS 3 IPSL-CM6A-LR 法国 1.5° × 1.25° IPSL 4 MIROC6 日本 1° × 1° MIROC 5 MRI-ESM2-0 日本 0.5° × 0.5° MRI 6 NorESM2-MM 挪威 1° × 1° NCC 
下载: 导出CSV
表 2 SODA和GODAS数据信息
Tab. 2 Data information of SODA and GODAS
序号 再分析数据 版本号 所属国 原始分辨率 研发机构 1 SODA 3.4.2 美国 0.5° × 0.5° UMD 2 GODAS v8 美国 1° × 1° NCEP
下载: 导出CSV
表 3 各模型评价指标对比结果
Tab. 3 Comparison results of evaluation metrics for each model
模型 MSE MAE RMSE No STGEM 0.6627 0.6445 0.8261 No Attention-I 0.6543 0.6112 0.8137 No Attention-O 0.7057 0.6782 0.7943 All no 0.7164 0.6957 0.8451 CAIPM 0.6352 0.5838 0.7936 注：表中MSE、MAE和RMSE数据为测试集中所有样本在相应模型预测下的平均值。
下载: 导出CSV
-
[1] 郑梦轲, 方巍, 张霄智. 深度学习在印度洋偶极子预测中的应用研究综述[J]. 海洋学研究, 2024, 42(3): 51−63 doi: 10.3969/j.issn.1001-909X.2024.03.004Zheng Mengke, Fang Wei, Zhang Xiaozhi. Review of application of deep learning in Indian Ocean Dipole prediction[J]. Journal of Marine Sciences, 2024, 42(3): 51−63. doi: 10.3969/j.issn.1001-909X.2024.03.004 [2] 滕宇威, 张文君, 刘超, 等. ENSO空间形态变异对ENSO-IOD关系年代际减弱的可能作用[J]. 气象学报, 2020, 78(2): 210−220.Teng Yuwei, Zhang Wenjun, Liu Chao, et al. Possible role of ENSO spatial pattern variation in interdecadal weakening of ENSO-IOD relationship[J]. Acta Meteorologica Sinica, 2020, 78(2): 210−220. [3] An S I, Park H J, Kim S K, et al. Main drivers of Indian Ocean Dipole asymmetry revealed by a simple IOD model[J]. npj Climate and Atmospheric Science, 2023, 6(1): 93. doi: 10.1038/s41612-023-00422-2 [4] Bahiyah A, Wirasatriya A, Mardiansyah W, et al. Massive SST-front anomaly in the tip of Sumatra waters triggered by extreme positive IOD 2019 event[J]. International Journal of Remote Sensing, 2024, 45(23): 8919−8936. doi: 10.1080/01431161.2023.2268821 [5] Sun Shuangwen, Fang Yue, Zu Yongcan, et al. Increased occurrences of early Indian Ocean Dipole under global warming[J]. Science Advances, 2022, 8(47): eadd6025. doi: 10.1126/sciadv.add6025 [6] Magee A D, Kiem A S. Using indicators of ENSO, IOD, and SAM to improve lead time and accuracy of tropical cyclone outlooks for Australia[J]. Journal of Applied Meteorology and Climatology, 2020, 59(11): 1901−1917. doi: 10.1175/JAMC-D-20-0131.1 [7] Roy I, Mliwa M, Troccoli A. Important drivers of East African monsoon variability and improving rainy season onset prediction[J]. Natural Hazards, 2024, 120(1): 429−445. doi: 10.1007/s11069-023-06223-3 [8] Cherchi A, Terray P, Ratna S B, et al. Indian Ocean Dipole influence on Indian summer monsoon and ENSO: a review[M]//Chowdary J, Parekh A, Gnanaseelan C. Indian Summer Monsoon Variability. Amsterdam: Elsevier, 2021: 157−182. [9] Cai Wenju, Yang Kai, Wu Lixin, et al. Opposite response of strong and moderate positive Indian Ocean Dipole to global warming[J]. Nature Climate Change, 2021, 11(1): 27−32. doi: 10.1038/s41558-020-00943-1 [10] Iskandar I, Lestari D O, Saputra A D, et al. Extreme positive Indian Ocean Dipole in 2019 and its impact on Indonesia[J]. Sustainability, 2022, 14(22): 15155. doi: 10.3390/su142215155 [11] Saji N H, Goswami B N, Vinayachandran P N, et al. A dipole mode in the tropical Indian Ocean[J]. Nature, 1999, 401(6751): 360−363. [12] Cai Wenju, Santoso A, Wang Guojian, et al. Increased frequency of extreme Indian Ocean Dipole events due to greenhouse warming[J]. Nature, 2014, 510(7504): 254−258. doi: 10.1038/nature13327 [13] Sarkar P P, Janardhan P, Roy P. A novel deep neural network model approach to predict Indian Ocean dipole and Equatorial Indian Ocean oscillation indices[J]. Dynamics of Atmospheres and Oceans, 2021, 96: 101266. doi: 10.1016/j.dynatmoce.2021.101266 [14] Feng Yuan, Li Chen, Sun Tianying. The study based on the deep learning for Indian Ocean Dipole (IOD) index predication[C]//Proceedings of the ACM Turing Award Celebration Conference – China. Hefei: Association for Computing Machinery, 2021: 23−27. [15] Feng Ming, Boschetti F, Ling Fenghua, et al. Predictability of sea surface temperature anomalies at the eastern pole of the Indian Ocean Dipole—using a convolutional neural network model[J]. Frontiers in Climate, 2022, 4: 925068. doi: 10.3389/fclim.2022.925068 [16] Liu Jun, Tang Youmin, Wu Yanling, et al. Forecasting the Indian Ocean Dipole with deep learning techniques[J]. Geophysical Research Letters, 2021, 48(20): e2021GL094407. doi: 10.1029/2021GL094407 [17] Ling Fenghua, Luo Jingjia, Li Yue, et al. Multi-task machine learning improves multi-seasonal prediction of the Indian Ocean Dipole[J]. Nature Communications, 2022, 13(1): 7681. doi: 10.1038/s41467-022-35412-0 [18] 方巍, 张霄智, 齐媚涵. MEPM模型: 基于深度学习的多变量厄尔尼诺−南方涛动预测模型[J]. 地球科学与环境学报, 2024, 46(3): 285−297. doi: 10.19814/j.jese.2023.08029Fang Wei, Zhang Xiaozhi, Qi Meihan. MEPM: multivariate ENSO prediction model based on deep learning[J]. Journal of Earth Sciences and Environment, 2024, 46(3): 285−297. doi: 10.19814/j.jese.2023.08029 [19] 张霄智, 方巍, 王淏西. 基于Swin-Transformer和时空融合注意力机制的ENSO预测[J]. 海洋学报, 2024, 46(12): 111−121. doi: 10.12284/hyxb2024127Zhang Xiaozhi, Fang Wei, Wang Haoxi. ENSO prediction based on Swin-Transformer and spatio-temporal fusion attention mechanism[J]. Haiyang Xuebao, 2024, 46(12): 111−121. doi: 10.12284/hyxb2024127 [20] 周天军, 邹立维, 陈晓龙. 第六次国际耦合模式比较计划(CMIP6)评述[J]. 气候变化研究进展, 2019, 15(5): 445−456.Zhou Tianjun, Zou Liwei, Chen Xiaolong. Commentary on the coupled model intercomparison project phase 6 (CMIP6)[J]. Climate Change Research, 2019, 15(5): 445−456. [21] Carton J A, Giese B S. A reanalysis of ocean climate using simple ocean data assimilation (SODA)[J]. Monthly Weather Review, 2008, 136(8): 2999−3017. doi: 10.1175/2007MWR1978.1 [22] Saha S, Nadiga S, Thiaw C, et al. The NCEP climate forecast system[J]. Journal of Climate, 2006, 19(15): 3483−3517. doi: 10.1175/JCLI3812.1 [23] Gao Zhihan, Shi Xingjian, Wang Hao, et al. Earthformer: exploring space-time transformers for earth system forecasting[C]//Proceedings of the 36th International Conference on Neural Information Processing Systems. New Orleans: Curran Associates Inc. , 2022: 1841. [24] Ham Y G, Kim J H, Luo Jingjia. Deep learning for multi-year ENSO forecasts[J]. Nature, 2019, 573(7775): 568−572. doi: 10.1038/s41586-019-1559-7 [25] Sun Ming, Chen Lin, Li T, et al. CNN-based ENSO forecasts with a focus on SSTA zonal pattern and physical interpretation[J]. Geophysical Research Letters, 2023, 50(20): e2023GL105175. doi: 10.1029/2023GL105175 [26] Fan Jin, Zhang Ke, Huang Yipan, et al. Parallel spatio-temporal attention-based TCN for multivariate time series prediction[J]. Neural Computing and Applications, 2023, 35(18): 13109−13118. doi: 10.1007/s00521-021-05958-z [27] Zha Wenshu, Liu Yuping, Wan Yujin, et al. Forecasting monthly gas field production based on the CNN-LSTM model[J]. Energy, 2022, 260: 124889. doi: 10.1016/j.energy.2022.124889 [28] Elmaz F, Eyckerman R, Casteels W, et al. CNN-LSTM architecture for predictive indoor temperature modeling[J]. Building and Environment, 2021, 206: 108327. doi: 10.1016/j.buildenv.2021.108327 [29] Geng Huantong, Hu Zhongyan, Wang Tianlei. ConvLSTM based temperature forecast modification model for North China[J]. Journal of Tropical Meteorology, 2022, 28(4): 405−412. doi: 10.46267/j.1006-8775.2022.030 [30] Moishin M, Deo R C, Prasad R, et al. Designing deep-based learning flood forecast model with ConvLSTM Hybrid Algorithm[J]. IEEE Access, 2021, 9: 50982−50993. doi: 10.1109/ACCESS.2021.3065939 -
图(6) / 表(3)
计量
- 文章访问数: 287
- HTML全文浏览量: 114
- PDF下载量: 90
- 被引次数: 0
