Prediction of Inter-annual Signal of Global Mean Surface Temperature Based on Deep Learning Approach
-
摘要: 利用集合经验模态分解（Ensemble Empirical Mode Decomposition, EEMD）有效地分解了全球平均表面温度（Global Mean Surface Temperature, GMST）时间序列，得到其不同尺度的、不同特征的子序列（Intrinsic Mode Function, IMF）。在此基础上，利用在预测长期、复杂、非线性变化的时间序列上具有显著优势的滑动自回归机器学习（Autoregressive Integrated Moving Average, ARIMA）模型和长短期记忆网络（Long Short-Term Memory, LSTM）模型开展GMST年际信号预测研究。结果表明：深度学习模型LSTM能很好地拟合并预测了长程相关性强的子序列（第2～6个IMF），而代表GMST年际尺度变化的IMF1则在一定程度上受到太平洋大西洋多重气候信号的影响和调制，因此进一步将3个气候指数作为预报前兆因子加入预测模型来更准确地预测IMF1的时间演变。通过利用多套GMST数据的对比，最终选定了考虑实时ENSO信息的LSTM（ENSO）模型来提前预测年际GMST信号，并预测2020年将有较大概率会成为史上最热的年份之一。
-
关键词:
- 全球平均表面温度 /
- 年际信号时间序列预测 /
- 集合经验模态分解 /
- 长短期记忆神经网络 /
- 深度学习预测模型
Abstract: Global Mean Surface Temperature (GMST) research and prediction is still an essential theoretical basis for climate change and disaster prevention. Because GMST series contains multi-scale variation characteristics that are highly complex and nonlinear, the Ensemble Empirical Mode Decomposition (EEMD) was adopted to effectively decompose the GMST time series to Intrinsic Mode Functions (IMFs), which obtains different scales and different characteristics. The machine learning model Autoregressive Integrated Moving Average (ARIMA) and the deep learning model Long Short-Term Memory (LSTM) present substantial advantages in predicting long-term, complex, and nonlinear time series to carry out GMST inter-annual signal prediction research. The results reveal that the deep learning model fits and predicts the sub-sequences with strong long-term correlation (IMF2-6). IMF1, which represents the inter-annual scale change of GMST is affected by the Pacific Ocean and the Atlantic Ocean multi-climate signal. Three climate indexes should be added as forecast precursor factors into the prediction model to predict IMF1 more accurately. This paper finally selected the LSTM(ENSO) model that considers real-time ENSO to predict the inter-annual GMST signal in advance by comparing multiple sets of GMST data and found that 2020 will have a greater probability of becoming one of the hottest years in history. -
图 1 长短期记忆网络（Long Short-Term Memory, LSTM）模型内部结构
Figure 1. Structure of the Long Short-Term Memory (LSTM)
图 2 集合经验模态分解（EEMD）方法对测试集全球平均表面温度（Global Mean Surface Temperature, GMST）数据进行分解：（a）原始序列；（b−f）各个本征模态分量；（h）残差项
Figure 2. Global Mean Surface Temperature (GMST) test set data decomposed using the Ensemble Empirical Mode Decomposition (EEMD) method: (a) The original sequence; (b−f) each Intrinsic Mode Function; (c) residual term
图 3 （a−f）ARIMA模型和（g−l）LSTM模型对于EEMD分解后的各个子序列（IMF1−6）预测结果（红色线为实际值，蓝色线为预测值，虚线为95%的信度区间的上、下限）
Figure 3. Predicted values of each IMFs (IMF1−6) decomposed by EEMD from (a–f) ARIMA model and (g–l) LSTM model (red line means the observed value, the blue line means the predicted value, and the dashed lines mean the 95% confidence interval of the prediction)
图 4 （a）1948～1997年、（b）1998～2007年和（c）2008～2017年所对应的年GMST时间序列回归至SST（ERSST V5数据）（黑点部分通过95%信度检验）
Figure 4. Time series of GMST corresponding to (a) 1948–1997, (b) 1998–2007, and (c) 2008–2017 are regressed to ERSST V5 data. The black dots mean exceeding the 95% confidence level
图 5 ARIMA、LSTM、LSTM-ENSO模型对于4套数据（a）HadCRUT4、（b）GISS、（c）BEST、（d）NOAA中1960～2019年GMST的预测结果（黑色为实际观测值，蓝色为ARIMA预测值，绿色为普通LSTM预测值，红色为更新ENSO的多变量LSTM预测值）
Figure 5. Prediction results of the three models (ARIMA, LSTM, and LSTM–ENSO) from 1960 to 2019 in the four GMST data sets (a) HadCRUT4, (b) GISS, (c) BEST, and (d) NOAA. The black line is the observed value, the blue line is the ARIMA predicted value, the green line is the ordinary LSTM predicted value, and the red line is the multivariate LSTM predicted value of the Sync updated real-time ENSO
图 6 ARIMA、LSTM、LSTM（ENSO）模型下的四套数据（a）HadCRUT4 、（b）GISS、（c）NOAA、（d）BEST误差对比（均方根误差RMSE、平均绝对误差MAE、平均绝对百分比误差MAPE）
Figure 6. Different error metrics (RMSE: Root-mean-square error, MAE: Mean absolute error, MAPE: Mean absolute percentage error) from (a) HadCRUT4, (b) GISS, (c) NOAA, and (d) BEST comparison under ARIMA, LSTM, and LSTM (ENSO) models
图 7 LSTM（ENSO）模型下HadCRUT4、GISS、NOAA、BEST 四套数据GMST的预测不确定性箱线图（红色为观测不确定性范围）
Figure 7. Range of forecast uncertainty of the four GMST datasets HadCRUT4, GISS, NOAA, and BEST under the LSTM (ENSO) model (box plot) (red is observed uncertainty)
表 1 资料信息及用途
Table 1. Information and purpose of the data sets
数据名称 时长 气候态 分辨率 用途 HadCRUT4 1850年以来 1961～1990年 5°（纬度）×5°（经度） 拟合试验和预测 GISS 1880年以来 1951～1980年 2°（纬度）×2°（经度） 拟合试验和预测 NOAA 1880年以来 1971～2000年 5°（纬度）× 5°（经度） 拟合试验和预测 BEST 1750年以来 1951～1980年 1°（纬度）× 1°（经度） 拟合试验和预测 ERSST V5 1854年以来 无 2°（纬度）× 2°（经度） 归因与气候指数 
下载: 导出CSV
表 2 LSTM的超参数设置
Table 2. Hyper-parameter settings of LSTM
设置参数 描述 设置值 输入节点数 输入训练集长度 100～160 输出节点数 预报时长 1 激活函数 控制“门”的输入输出 Relu 优化算子 改变学习率 AdaDelta 损失函数 减少梯度 mae 正则优化 防止过度拟合 0.1 反向传播步数 误差反向传播的步长 3 迭代次数 整个训练集迭代次数 2000
下载: 导出CSV
-
[1] Blaes S, Burwick T. 2017. Few-shot learning in deep networks through global prototyping [J]. Neural Networks, 94: 159−172. doi: 10.1016/j.neunet.2017.07.001 [2] 陈卓, 孙龙祥. 2018. 基于深度学习LSTM网络的短期电力负荷预测方法 [J]. 电子技术, 47(1): 39−41. doi: 10.3969/j.issn.1000-0755.2018.01.011Chen Zhuo, Sun Longxiang. 2018. Short-term electrical load forecasting based on deep learning LSTM networks [J]. Electronic Technology (in Chinese), 47(1): 39−41. doi: 10.3969/j.issn.1000-0755.2018.01.011 [3] Dimri T, Ahmad S, Sharif M. 2020. Time series analysis of climate variables using seasonal ARIMA approach [J]. Journal of Earth System Science, 129(1): 149. doi: 10.1007/s12040-020-01408-x [4] 封国林, 王启光, 侯威, 等. 2009. 气象领域极端事件的长程相关性 [J]. 物理学报, 58(4): 2853−2861. doi: 10.3321/j.issn:1000-3290.2009.04.115Feng Guolin, Wang Qiguang, Hou Wei, et al. 2009. Long-range correlation of extreme events in meterorological field [J]. Acta Physica Sinica (in Chinese), 58(4): 2853−2861. doi: 10.3321/j.issn:1000-3290.2009.04.115 [5] Foster G, Rahmstorf S. 2011. Global temperature evolution 1979–2010 [J]. Environmental Research Letters, 6(4): 044022. doi: 10.1088/1748-9326/6/4/044022 [6] Ghimire B, Williams C A, Masek J, et al. 2014. Global albedo change and radiative cooling from anthropogenic land cover change, 1700 to 2005 based on MODIS, land use harmonization, radiative kernels, and reanalysis [J]. Geophys. Res. Lett., 41(24): 9087−9096. doi: 10.1002/2014GL061671 [7] Guo Y N, Cao X Q, Liu B N, et al. 2020. El Niño index prediction using deep learning with ensemble empirical mode decomposition [J]. Symmetry, 12(6): 893. doi: 10.3390/sym12060893 [8] Hansen J, Ruedy R, Sato M, et al. 2010. Global surface temperature change [J]. Rev. Geophys., 48(4): RG4004. doi: 10.1029/2010RG000345 [9] Hawkins E, Sutton R. 2009. The potential to narrow uncertainty in regional climate predictions [J]. Bull. Amer. Meteor. Soc., 90(8): 1095−1108. doi: 10.1175/2009BAMS2607.1 [10] Ji Z, Chai X L, Yu Y L, et al. 2020. Improved prototypical networks for few-shot learning [J]. Pattern Recognition Letters, 140: 81−87. doi: 10.1016/j.patrec.2020.07.015 [11] 蒋国兴. 2007. 偏最小二乘回归方法（PLS）在短期气候预测中的应用研究 [D]. 南京信息工程大学硕士学位论文, 67pp. Jang Guoxing. 2007. The application for the Partial Least-Squares Regression (PLS) in the short-term climate forecast [D]. M. S. thesis (in Chinese), Nanjing University of Information Science and Technology, 67pp. [12] Kajtar J B, Collins M, Frankcombe L M, et al. 2019. Global mean surface temperature response to large-scale patterns of variability in observations and CMIP5 [J]. Geophys. Res. Lett., 46(4): 2232−2241. doi: 10.1029/2018GL081462 [13] Kosaka Y, Xie S P. 2013. Recent global-warming hiatus tied to equatorial Pacific surface cooling [J]. Nature, 501(7467): 403−407. doi: 10.1038/nature12534 [14] Kosaka Y, Xie S P. 2016. The tropical Pacific as a key pacemaker of the variable rates of global warming [J]. Nature Geoscience, 9(9): 669−673. doi: 10.1038/ngeo2770 [15] 李佳佳, 贺新光, 卢希安. 2019. 长江流域月降水的EEMD多时间尺度遥相关分析 [J]. 长江流域资源与环境, 28(8): 1898−1908. doi: 10.11870/cjlyzyyhj201908013Li Jiajia, He Xinguang, Lu Xi’an. 2019. Multi-time scale teleconnection analysis of monthly precipitation in the Yangtze River Basin based on the EEMD [J]. Resources and Environment in the Yangtze Basin (in Chinese), 28(8): 1898−1908. doi: 10.11870/cjlyzyyhj201908013 [16] 李立娟, 王斌, 周天军. 2007. 外强迫因子对20世纪全球变暖的综合影响 [J]. 科学通报, 52(15):1820–1825. Li Lijuan, Wang Bin, Zhou Tianjun. 2007. The comprehensive impact of external forcing factors on global warming in the 20th century [J]. Chinese Science Bulletin, 52(22): 3148−3154. doi: 10.1360/csb2007-52-15-1820 [17] 陆静文, 周天军, 黄昕, 等. 2020. 表面气温内部变率估算方法的比较研究 [J]. 大气科学, 44(1): 105−121. doi: 10.3878/j.issn.1006-9895.1901.18235Lu Jingwen, Zhou Tianjun, Huang Xin, et al. 2020. A comparison of three methods for estimating internal variability of near-surface air temperature [J]. Chinese Journal of Atmospheric Sciences (in Chinese), 44(1): 105−121. doi: 10.3878/j.issn.1006-9895.1901.18235 [18] 马尚谦, 张勃, 杨梅, 等. 2019. 基于EEMD的华北平原1901～2015年旱涝灾害分析[J]. 干旱区资源与环境, 33(3): 62–68.Ma Shangqian, Zhang Bo, Yang Mei, et al. Analysis of drought and flood disasters in the North China Plain from 1901 to 2015 based on EEMD [J]. Journal of Arid Land Resources and Environment (in Chinese), 33(3): 62–68. doi: 10.13448/j.cnki.jalre.2019.075 [19] Meehl G A, Teng H Y, Arblaster J M. 2014. Climate model simulations of the observed early-2000s hiatus of global warming [J]. Nature Climate Change, 4(10): 898−902. doi: 10.1038/nclimate2357 [20] Schneider T. 2007. Climate modelling: Uncertainty in climate-sensitivity estimates [J]. Nature, 446(7131): E1. doi: 10.1038/nature05707 [21] Sheffield J, Wood E F. 2008. Projected changes in drought occurrence under future global warming from multi-model, multi-scenario, IPCC AR4 simulations [J]. Climate Dyn., 31(1): 79−105. doi: 10.1007/s00382-007-0340-z [22] Shindell D T, Lamarque J F, Schulz M, et al. 2013. Radiative forcing in the ACCMIP historical and future climate simulations [J]. Atmos. Chem. Phys., 13(6): 2939−2974. doi: 10.5194/acp-13-2939-2013 [23] Strobach E, Bel G. 2020. Learning algorithms allow for improved reliability and accuracy of global mean surface temperature projections [J]. Nature Communications, 11(1): 451. doi: 10.1038/s41467-020-14342-9 [24] Vose R S, Arndt D, Banzon V F, et al. 2012. NOAA’s merged land–ocean surface temperature analysis [J]. Bull. Amer. Meteor. Soc., 93(11): 1677−1685. doi: 10.1175/BAMS-D-11-00241.1 [25] 王鑫, 吴际, 刘超, 等. 2018. 基于LSTM循环神经网络的故障时间序列预测 [J]. 北京航空航天大学学报, 44(4): 772−784. doi: 10.13700/j.bh.1001-5965.2017.0285Wang Xin, Wu Ji, Liu Chao, et al. 2018. Exploring LSTM based recurrent neural network for failure time series prediction [J]. Journal of Beijing University of Aeronautics and Astronautics (in Chinese), 44(4): 772−784. doi: 10.13700/j.bh.1001-5965.2017.0285 [26] 杨函. 2017. 基于深度学习的气象预测研究 [D]. 哈尔滨工业大学硕士学位论文, 73pp. Yang Han. 2017. Research on weather forecasting based on Deep Learning [D]. M. S. thesis (in Chinese), Harbin Institute of Technology, 73pp. [27] 云翔. 2019. 全球地表温度数据的改进以及增暖分析 [D]. 中国气象科学研究院硕士学位论文, 88pp. Yun Xiang. 2019. Improvement of surface temperature and warming analysis [D]. M. S. thesis (in Chinese), Chinese Academy of Meteorological Sciences, 88pp. [28] 张莹, 谭艳春, 彭发定, 等. 2019. 基于EEMD和ARIMA的海温预测模型研究 [J]. 海洋学研究, 37(1): 9−14. doi: 10.3969/j.issn.1001-909X.2019.01.002Zhang Yin, Tan Yanchun, Peng Dingfa, et al. 2019. Study on time series prediction model of sea surface temperature based on ensemble empirical mode decomposition and autoregressive integrated moving average [J]. Journal of Marine Sciences (in Chinese), 37(1): 9−14. doi: 10.3969/j.issn.1001-909X.2019.01.002 [29] 周天军, 吴波. 2017. 年代际气候预测问题: 科学前沿与挑战 [J]. 地球科学进展, 32(4): 331−341. doi: 10.11867/j.issn.1001-8166.2017.04.0331Zhou Tianjun, Wu Bo. 2017. Decadal climate prediction: Scientific frontier and challenge [J]. Advances in Earth Science (in Chinese), 32(4): 331−341. doi: 10.11867/j.issn.1001-8166.2017.04.0331 [30] Zwiers F, Hegerl G. 2008. Climate change: Attributing cause and effect [J]. Nature, 453(7193): 296−297. doi: 10.1038/453296a -
点击查看大图
计量
- 文章访问数: 201
- HTML全文浏览量: 36
- PDF下载量: 29
- 被引次数: 0
