Prediction of Sea Surface Wind Speed in China's Offshore Areas Based on Weighted Loss Function
-
摘要: 数值模式的海面风速预报存在偏差，基于数值模式的统计模型虽然能在一定程度上减小偏差，但由于大风样本严重不足，其对大风的预报效果仍然差强人意。在系统检验ECMWF的24 h海面风速预报的基础上，利用XGBoost模型构建了适用于中国近海的海面风速预报订正模型。该模型不仅从整体上具有不错的预报准确度，还通过采用加权损失函数的训练方式，明显改善了大风的预报效果。利用2022年1月—2023年1月我国近海14个浮标观测数据对模型进行独立检验，模型的平均误差和均方根误差分别为0.11 m·s-1和1.75 m·s-1，其中，对于7~9级风，模型的预报准确度较ECMWF显著提高，均方根误差分别降低了15%、25%、24%。此外，当模型应用在未参与训练的网格点上时，模型仍然能获得较ECMWF更准确的预报。该模型操作方便容易使用，可为中国近海海面风速预报尤其是大风预报提供参考。Abstract: Sea surface wind speed forecasts based on numerical models often exhibit deviations. Although statistical models based on numerical models can reduce the deviation to a certain extent, their performance in predicting strong winds remains suboptimal due to the scarcity of strong wind samples. This study systematically evaluated the 24-hour sea surface wind speed forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF) and employed the XGBoost model to develop a correction model tailored for China's offshore areas. This model not only demonstrated good overall forecasting accuracy, but also significantly enhanced the prediction of strong winds through the use of a weighted loss function during training. The model was independently tested using observational data from 14 buoys in China's offshore waters from January 2022 to January 2023. The average error and root mean squared error of the model were 0.11 m s-1 and 1.75 m s-1, respectively. The forecast accuracy of the model for wind speeds classified as levels 7 to 9 was significantly improved compared to that of the ECMWF, with root mean squared errors reduced by 15%, 25%, and 24%, respectively. Furthermore, when applied to grid points not included in the training, the model continued to provide more accurate forecasts than the ECMWF. This model is easy to operate and use and can provide reference information for sea surface wind speed forecasting, especially strong wind forecasting, in China's coastal waters.
-
Key words:
- sea surface wind speed /
- forecast correction /
- weighted loss function
-
图 3 观测的(横坐标)与ECMWF预报的(纵坐标)10 m风速的散点分布
填色代表样本数，黑色实线为对角线，红色虚线为线性拟合回归线，左上角为拟合回归线方程，右上角为R2。
图 10 FCM对不同站点在风速小于4 m·s-1(a)和大于14 m·s-1(b)条件下预报的RMSE
填色代表站点RMSE，三角形(正方形)代表该站点RMSE低于(高于)ECMWF，其大小代表FCM的RMSE较ECMWF偏低(偏高)量。
表 1 不同风速样本下各站点ME和RMSE的标准差、最小值和最大值
样本 ME RMSE 标准差 最小值 最大值 标准差 最小值 最大值 总样本 0.31 -0.58 0.47 0.27 1.59 2.59 < 4 m·s-1样本 0.49 0.64 2.70 0.60 1.56 4.02 > 14 m·s-1s样本 0.61 -3.34 -1.27 0.53 1.85 3.65 4~14 m·s-1样本 0.25 -0.64 0.20 0.14 1.50 2.02 
下载: 导出CSV
表 2 备选预报因子
要素名称 层次 经向风 10 m、100 m、950 hPa、850 hPa、700 hPa、500 hPa 纬向风 10 m、100 m、950 hPa、850 hPa、700 hPa、500 hPa 气温 2 m、950 hPa、850 hPa、700 hPa、500 hPa 位势高度 950 hPa、850 hPa、700 hPa、500 hPa 相对湿度 950 hPa、850 hPa、700 hPa、500 hPa 海平面气压 --
下载: 导出CSV
表 3 最佳预报因子组合
要素名称 层次(预报时效) 风速 10 m(00 h/24 h)、100 m(00 h/24 h)、950 hPa(00 h/24 h)、700 hPa(24 h)、500 hPa(00 h/24 h) 风向 10 m(00 h/24 h)、100 m(24 h)、950 hPa(24 h)、850 hPa(24 h)、500 hPa(00 h/24 h) 气温 2 m(00 h/24 h)、950 hPa(00 h/24 h)、850 hPa(00 h/24 h)、700 hPa(00 h)、500 hPa(24 h) 位势高度 950 hPa(00 h/24 h)、850 hPa(00 h/24 h)、700 hPa(00 h)、500 hPa(24 h) 相对湿度 950 hPa(24 h)、850 hPa(24 h)、500 hPa(24 h) 海平面气压 海平面(00 h/24 h)
下载: 导出CSV
-
[1] 许小峰, 顾建峰, 李永平. 海洋气象灾害[M]. 北京: 气象出版社, 2009. [2] 尹尽勇, 徐晶, 曹越男, 等. 我国海洋气象预报业务现状与发展[J]. 气象科技进展, 2012, 2(6): 17-26. [3] ZHENG C W, LI C Y, PAN J, et al. An overview of global ocean wind energy resource evaluations[J]. Renewable and Sustainable Energy Reviews, 2016, 53(1. ): 1 240-1 251. [4] MU M, DUAN W S, WANG J C. The predictability problems in numerical weather and climate prediction[J]. Adv Atmos Sci, 2002, 19(2): 191-204. [5] 季晓阳, 吴辉碇, 杨学联. 海面风场数值预报的历史和现状[J]. 海洋预报, 2005, 22(S1): 167-171. [6] 李敏, 王辉, 金啟华. 中国近海海面风场预报方法综述[J]. 海洋预报, 2009, 26(3): 114-120. [7] 高山红, 张新玲, 吴增茂. 渤海海面风场的一种动力诊断方法[J]. 海洋学报(中文版), 2001, 23(6): 51-58. [8] 毛绍荣, 林镇国, 梁健, 等. 广东沿海强东北季风的概率预报方法[J]. 热带气象学报, 2003, 17(1): 94-100. [9] 刘京雄, 唐文伟, 朱持则, 等. 浙闽沿海和台湾海峡海域冬季大风风速计算方法探讨[J]. 台湾海峡, 2004, 23(1): 8-13. [10] 林良勋, 程正泉, 张兵, 等. 完全预报(PP)方法在广东冬半年海面强风业务预报中的应用[J]. 应用气象学报, 2004, 15(4): 485-493. [11] 周伟隆, 陈往溪, 肖巍. 粤东海面冷空气强风的统计分析与预报[J]. 广东气象, 2005(4): 20-22. [12] TAMBKE J, LANGE M, FOCKEN U, et al. Forecasting offshore wind speeds above the North Sea[J]. Wind Energy, 2005, 8(1): 3-16. [13] 陈德花, 刘铭, 苏卫东, 等. BP人工神经网络在MM5预报福建沿海大风中的释用[J]. 暴雨灾害, 2010, 29(3): 263-267. [14] 薄文波, 高山红, 王永明. 基于WRF模式的渤海海面风的预报与订正[J]. 海洋湖沼通报, 2013(3): 37-44. [15] 荣艳敏, 阎丽凤, 盛春岩, 等. 山东精细化海区风的MOS预报方法研究[J]. 海洋预报, 2015, 32(3): 59-67. [16] PARVIZ L. A novel ensemble wind speed forecasting method using the differential weighting scheme and principal component analysis[J]. Időjárás, 2023, 127(1): 55-76. [17] 胡海川, 赵伟, 董林. 概率密度匹配方法在我国近海海面10 m风速预报中的应用[J]. 热带气象学报, 2021, 37(1): 91-101. [18] YANG Y Z, ZHA K W, CHEN Y C, et al. Delving into Deep Imbalanced Regression[C]// Proceedings of the 38th International Conference on Machine Learning. PMLR, 2021: 11 842-11 851[2024-03-06]. [19] CHAWLA N V, BOWYER K W, HALL L O, et al. SMOTE: synthetic minority over-sampling technique[J]. Journal of artificial intelligence research, 2002, 16(1): 321-357. [20] HE H, BAI Y, GARCIA E A, et al. ADASYN: Adaptive synthetic sampling approach for imbalanced learning[C]//2008 IEEE international joint conference on neural networks (IEEE world congress on computational intelligence). IEEE, 2008: 1 322-1 328[2024-04-01]. [21] GARCÍA S, HERRERA F. Evolutionary undersampling for classification with imbalanced datasets: Proposals and taxonomy[J]. Evolutionary computation, 2009, 17(3): 275-306. [22] HUANG C, LI Y, LOY C C, et al. Learning deep representation for imbalanced classification[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. [2024-04-01]. [23] DONG Q, GONG S, ZHU X. Imbalanced deep learning by minority class incremental rectification[J]. IEEE transactions on pattern analysis and machine intelligence, 2018, 41(6): 1 367-1 381. [24] CAO K, WEI C, GAIDON A, et al. Learning imbalanced datasets with label-distribution-aware margin loss[J]. Advances in Neural information processing systems, 2019, 32[2024-04-01]. [25] CHEN T Q, HE T, BENESTY M, et al. XGBoost: Extreme Gradient Boosting[CP/OL]. R package version 0.4-2, 2015. https://CRAN.R-project.org/package=xgboost .[26] DIEBOLD F X, MARIANO R S. Comparing Predictive Accuracy[J]. Journal of Business & Economic Statistics, 2002, 20(1): 134-144. [27] AKIBA T, SANO S, YANASE T, et al. Optuna: A Next-generation Hyperparameter Optimization Framework[A]. arXiv, 2019[2023-11-08]. [28] KOHAVI R. A study of cross-validation and bootstrap for accuracy estimation and model selection[C]//Ijcai. Montreal, Canada, 1995: 1 137-1 145. -
点击查看大图
图(12) / 表(3)
计量
- 文章访问数: 12
- HTML全文浏览量: 1
- PDF下载量: 2
- 被引次数: 0
粤公网安备 4401069904700003号