Comparative Machine Learning-Based Correction Experiment for a 10 m Wind Speed Forecast at a 100 m Resolution in Complex Mountainous Areas of the Winter Olympic Games
-
摘要: 本文以传统机器学习算法XGBoost和深度学习算法CU-Net为基础,针对北京快速更新无缝隙融合与集成预报系统(RISE系统)预报的北京冬奥会延庆及张家口赛区100米分辨率的冬季近地面10 m风速数据,进行每日逐小时起报的未来逐6小时间隔的冬奥高山站点及其周边地区风速预报偏差订正方法研究和对比分析。对于站点订正,首先将RISE系统预测的10 m风速插值到对应的自动气象站站点,然后根据风速等级表归类,针对每个分类单独构建XGBoost模型,每个区间模型合并后形成L-XGBoost,使用均方根误差和预报准确率作为评分标准,结果表明风速归类的L-XGBoost算法订正效果比不归类的原始XGBoost模型有一定提升,说明在传统机器学习中加入归类方法有助于改善复杂山地站点风速预报技巧。对于站点及其周边地区风速订正,本文在CU-Net模型基础上,通过引入不同深度的CU-Net子网络,构建了新的算法模型CU-Net++,并考虑了预报日变化误差和复杂地形对10 m风速的影响,以自动气象站为中心构建空间小区域样本数据,对RISE系统风速预报偏差进行订正。试验结果表明,CU-Net和CU-Net++均可以充分挖掘时间和空间维度的风场变化规律,且CU-Net++模型风速订正结果优于CU-Net模型,有效降低了RISE产品的格点风速预报误差,也发现预报误差和复杂地形的引入对10 m风速偏差订正起到重要的正向作用。Abstract: Based on a traditional machine learning algorithm (XGBoost), a deep learning algorithm (CU-Net), and the winter wind speed data from 10 m near the ground with a resolution of 100 m, this paper studied and compared the correction methods for wind speed forecast deviation in the mountainous stations and surrounding areas of the Yanqing and Zhangjiakou competition areas (Beijing Winter Olympic Games) using the rapid-refresh integrated seamless ensemble (RISE) system. For station correction, the 10-m wind speed predicted by the RISE system is interpolated to the corresponding automatic weather station. Subsequently, a separate XGBoost model is constructed for each classification according to the wind speed rating table. Afterward, each interval model was combined to form L-XGBoost, using the root mean square error and forecast accuracy as its scoring standard. Investigations revealed that the correction effect of the L-XGBoost algorithm for wind speed classification was better than the original XGBoost model without classification, indicating that introducing a classification method to traditional machine learning helped improve the wind speed prediction skills of the complex mountain stations. Subsequently, for the wind speed correction of the station and its surrounding areas based on the CU-Net model, this paper constructed a new algorithm model (CU-Net++) by introducing the CU-Net sub-networks with different depths, considering the influence of daily forecast errors and complex terrains on the 10-m wind speed. This paper also constructed spatial small-area sample data, considering the automatic weather station as the center, to correct the wind speed prediction deviation of the RISE system. The test results indicated that although both CU-Net and CU-Net++ fully mined the wind field change rules in time and space dimensions, the wind speed correction results of the CU-Net++ model performed better than those of the CU-Net model, effectively reducing the grid wind speed prediction error of RISE products. Hence, introducing prediction error and complex terrain plays an important positive role in the deviation correction of a surface 10 m wind speed-based investigation.
-
Key words:
- 100 m scale forecast /
- Complex mountain /
- Machine learning /
- Wind speed correction
-
图 6 预报时效为(a)6 h、(b)12 h、(c)18 h、(d)24 h的各风力区间10 m全风速订正结果。横坐标为风速区间,左坐标为RMSE,右坐标为误差减小率,绿色图例为模型订正后相比RISE(Rapid-refresh Itegrated Seamless Ensemble)原始预报RMSE减少百分比
Figure 6. Schematic showing the 10-m full wind speed correction results for each wind speed interval with forecast time effects of (a) 6 h, (b) 12 h, (c) 18 h, and (d) 24 h. Abscissa indicated wind speed interval, the left coordinate indicates RMSE, and the right coordinate indicates error reduction rate, the green legend shows the percentage reduction in RMSE compared to the original forecast of RISE after model correction
图 7 预报时效为(a)6 h、(b)12 h、(c)18 h、(d)24 h的各风力区间10 m全风速订正结果(横坐标为风速区间,纵坐标为预报准确率和准确率提升百分比,绿色图例为模型订正后相比RISE原始预报准确率增加百分比)
Figure 7. Schematic showing the 10 m full wind speed correction results for each wind speed interval with forecast time effects of (a) 6 h, (b) 12 h, (c) 18 h, and (d) 24 h (abscissa indicates wind speed intervals, the ordinate indicates wind prediction accuracy, The green legend shows the percentage increase in accuracy of the original forecast compared to RISE after model correction)
图 8 预报时效18 h情况下L-XGBoost子模型特征重要性权重分布。(a)−(g)分别表示风速等级1~7子模型的特征权重,横坐标为特征分数,纵坐标为特征名称
Figure 8. Weight distribution showing the feature importance of L-XGBoost sub-models after 18 h forecast. (a)−(g) represent the feature weights of sub models with wind speed levels 1−7, respectively. The horizontal axis represents the feature score, and the vertical axis represents the feature name
图 9 XGBoost和L-XGBoost两种模型10 m风速订正后的预报RMSE(左坐标)和误差减小率(右坐标):(a)A1489站;(b)A1491站;(c)A1703站;(d)B1649站
Figure 9. Forecasted RMSE (left coordinate) and error reduction rate (right coordinate) after a 10 m wind speed correction with the XGBoost and L-XGBoost models: (a) The A1489 station; (b) the A1491 station; (c) the A1703 station; (d) the B1649 station
图 11 二海陀站及其周边样本(A1491)在预报时效6 h、12 h、18 h、24 h下10 m风速订正结果对比:(a,c,e,g)RISE系统原始预报的RMSE;(b,d,f,g)CU-Net++订正后的RMSE
Figure 11. Comparison of 10 m wind speed correction results between the Erhaituo Station and its surrounding samples (A1491) under prediction time of 6 h, 12 h, 18 h, and 24 h: (a, c, e, g) Originally predicted RMSE by the RISE system; (b, d, f, g) the CU-Net++ corrected RMSE
图 12 竞速三号站及其周边样本(A1703)在预报时效6 h、12 h、18 h、24 h下10 m风速订正结果对比:(a,c,e,g)RISE系统原始预报的RMSE;(b,d,f,g)CU-Net++订正后的RMSE
Figure 12. Comparison of 10 m wind speed correction results between the Racing Station 3 and its surrounding samples (A1703) under prediction time of 6 h, 12 h, 18 h, and 24 h: (a, c, e, g) Originally predicted RMSE by the RISE system; (b, d, f, g) the CU-Net++ corrected RMSE
表 1 风力等级
Table 1. Wind grade
风力等级 风速/m s−1 描述 1 [0,1.5) 软风 2 [1.5,3.3) 轻风 3 [3.3,5.4) 微风 4 [5.4,7.9) 和风 5 [7.9,10.7) 清劲风 6 [10.7,13.8) 强风 7 ≥13.8 疾风 表 2 风速区间对应样本数
Table 2. Number of samples corresponding to wind speed range
风速等级 样本个数 1 70102 2 89019 3 63754 4 31125 5 10293 6 3840 7 1756 表 3 数据增强后风速区间对应样本数
Table 3. Number of samples corresponding to wind speed range after data enhancement
风速等级 样本个数 1 70102 2 89019 3 63754 4 31125 5 10293 6 11443 7 5287 表 4 单个样本组成
Table 4. Single sample composition
模型特征 标签 静态数据 实况数据 预测数据 误差 elev ws2at,wd2at,tst,
tdst,rhst,qst$ \text{usf}{\text{c}}_{t}^{\text{18}} $,$ \text{vsf}{\text{c}}_{t}^{\text{18}} $,
$ \text{uvf}{\text{c}}_{t}^{\text{18}} $$ {d}_{{t-24}}^{18} $,$ {d}_{{t-18}}^{\text{18}} $ ws2at+18 表 5 L-XGBoost模型和XGBoost模型对10 m风速订正结果RMSE指标对比
Table 5. Comparison of correction results of L-XGBoost and XGBoost models for 10-m wind speed in terms of RMSE
预报时效/h RMSE/m s−1 订正预报RMSE相对比原RMSE的变化百分比 RISE XGBoost L-XGBoost XGBoost_Rate L-XGBoost_Rate 6 1.839 1.408 1.331 23.38% 27.62% 12 1.869 1.467 1.405 21.5% 24.83% 18 1.916 1.491 1.429 22.13% 25.42% 24 1.927 1.493 1.442 22.53% 25.17% 表 6 L-XGBoost模型和XGBoost模型对10 m风速订正结果FA指标对比
Table 6. Comparison of correction results of L-XGBoost and XGBoost models for 10-m wind speed in terms of Forecast Accuracy (FA)
预报时效/h 预报准确率 订正预报准确率相比RISE原始准确率变化百分比 RISE XGBoost L-XGBoost XGBoost_Rate L-XGBoost_Rate 6 67.75% 78.15% 80.87% 10.40% 13.12% 12 67.13% 76.09% 79.00% 8.96% 11.87% 18 66.63% 75.30% 78.09% 8.67% 11.46% 24 65.61% 75.77% 78.44% 10.16% 12.83% 表 7 两组对照试验的单个样本组成
Table 7. Single sample composition of two groups of control experiments
特征 特征尺寸 标签 标签尺寸 (P, Wnow) (2×48×48) L (1×48×48) (P, Wnow, H, E1, E2) (5×48×48) L (1×48×48) 表 8 预报时效6 h下各种方法10 m 风速订正试验结果评分表
Table 8. Scoring table for 10 m wind speed correction test results of various methods under 6 h forecast time
模型 RMSE评分/m s−1 提升百分比 RISE原始 1.6600 / CU-Net(不加误差和地形) 1.2815 22.80% CU-Net(加误差和地形) 1.2467 24.90% CU-Net++(不加误差和地形) 1.2438 25.07% CU-Net++(加误差和地形) 1.2263 26.13% 表 9 预报时效12 h下各种方法10 m风速订正试验结果评分表
Table 9. Scoring table for 10 m wind speed correction test results of various methods under 12 h forecast time
模型 RMSE评分/m s−1 提升百分比 RISE原始 1.6908 / CU-Net(不加误差和地形) 1.3044 22.85% CU-Net(加误差和地形) 1.2931 23.52% CU-Net++(不加误差和地形) 1.2799 24.30% CU-Net++(加误差和地形) 1.2694 24.92% 表 10 预报时效18 h下各种方法10 m风速订正试验结果评分表
Table 10. Scoring table for 10 m wind speed correction test results of various methods under 18 h forecast time
模型 RMSE评分/m s−1 提升百分比 RISE原始 1.7285 / CU-Net(不加误差和地形) 1.329 23.11% CU-Net(加误差和地形) 1.3105 24.18% CU-Net++(不加误差和地形) 1.3097 24.23% CU-Net++(加误差和地形) 1.3026 24.64% 表 11 预报时效24 h下各种方法10 m风速订正试验结果评分表
Table 11. Scoring table for 10 m wind speed correction test results of various methods under 24 h forecast time
模型 RMSE评分/m s−1 提升百分比 RISE原始 1.7353 / CU-Net(不加误差和地形) 1.3485 22.29% CU-Net(加误差和地形) 1.3382 22.88% CU-Net++(不加误差和地形) 1.3328 23.19% CU-Net++(加误差和地形) 1.3267 23.54% -
[1] Benoit R, Schär C, Binder P, et al. 2002. The real-time ultrafinescale forecast support during the special observing period of the MAP [J]. Bull. Amer. Meteor. Soc., 83(1): 85−110. doi: 10.1175/1520-0477(2002)083<0085:TRTUFS>2.3.CO;2 [2] 常俊, 彭新东, 范广洲, 等. 2015. 结合历史资料的数值天气预报误差订正 [J]. 气象学报, 73(2): 341−354. doi: 10.11676/qxxb2015.021Chang Jun, Peng Xindong, Fan Guangzhou, et al. 2015. Error correction of numerical weather prediction with historical data [J]. Acta Meteor. Sinica (in Chinese), 73(2): 341−354. doi: 10.11676/qxxb2015.021 [3] Chen T Q, Guestrin C. 2016. XGBoost: A scalable tree boosting system [C]//Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. San Francisco: ACM, 785–794. doi:10.1145/2939672.2939785 [4] 丁瑞强, 李建平. 2007. 误差非线性的增长理论及可预报性研究 [J]. 大气科学, 31(4): 571−576. doi: 10.3878/j.issn.1006-9895.2007.04.02Ding Ruiqiang, Li Jianping. 2007. Nonlinear error dynamics and predictability study [J]. Chinese Journal of Atmospheric Sciences (in Chinese), 31(4): 571−576. doi: 10.3878/j.issn.1006-9895.2007.04.02 [5] Feng J, Ding R Q, Liu D Q, et al. 2014. The application of nonlinear local Lyapunov vectors to ensemble predictions in Lorenz systems [J]. J. Atmos. Sci., 71(9): 3554−3567. doi: 10.1175/JAS-D-13-0270.1 [6] 高登义, 邹捍, 周立波, 等. 2003. 中国山地环境气象研究进展 [J]. 大气科学, 27(4): 567−590. doi: 10.3878/j.issn.1006-9895.2003.04.010Gao Dengyi, Zou Han, Zhou Libo, et al. 2003. Advances of China mountain environment meteorology [J]. Chinese Journal of Atmospheric Sciences (in Chinese), 27(4): 567−590. doi: 10.3878/j.issn.1006-9895.2003.04.010 [7] 海滢, 陈光华. 2019. 利用人工神经网络模型预测西北太平洋热带气旋生成频数 [J]. 气候与环境研究, 24(3): 324−332. doi: 10.3878/j.issn.1006-9585.2019.18110Hai Ying, Chen Guanghua. 2019. Prediction of frequency of tropical cyclones forming over the western North Pacific using an artificial neural network model [J]. Climatic and Environmental Research (in Chinese), 24(3): 324−332. doi: 10.3878/j.issn.1006-9585.2019.18110 [8] Han L, Chen M X, Chen K K, et al. 2021. A deep learning method for bias correction of ECMWF 24−240 h forecasts [J]. Adv. Atmos. Sci., 38(9): 1444−1459. doi: 10.1007/s00376-021-0215-y [9] 贺圣平, 王会军, 李华, 等. 2021. 机器学习的原理及其在气候预测中的潜在应用 [J]. 大气科学学报, 44(1): 26−38. doi: 10.13878/j.cnki.dqkxxb.20201125001He Shengping, Wang Huijun, Li Hua, et al. 2021. Machine learning and its potential application to climate prediction [J]. Transactions of Atmospheric Sciences (in Chinese), 44(1): 26−38. doi: 10.13878/j.cnki.dqkxxb.20201125001 [10] Howard T, Clark P. 2007. Correction and downscaling of NWP wind speed forecasts [J]. Meteorological Applications, 14(2): 105−116. doi: 10.1002/met.12 [11] 胡海川, 黄彬, 魏晓琳. 2017. 我国近海洋面10 m风速集合预报客观订正方法 [J]. 气象, 43(7): 856−862. doi: 10.7519/j.issn.1000-0526.2017.07.009Hu Haichuan, Huang Bin, Wei Xiaolin. 2017. Objective correction method of ensemble forecast of 10 m winds on Chinese offshore [J]. Meteorological Monthly (in Chinese), 43(7): 856−862. doi: 10.7519/j.issn.1000-0526.2017.07.009 [12] Huang G, Liu Z, Van Der Maaten L, et al. 2017. Densely connected convolutional networks [C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, USA: IEEE, 2261–2269. doi:10.1109/CVPR.2017.243 [13] Isola P, Zhu J Y, Zhou T H, et al. 2017. Image-to-image translation with conditional adversarial networks [C]//Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, USA: IEEE, 5967–5976. doi:10.1109/CVPR.2017.632 [14] 贾春晖, 窦晶晶, 苗世光, 等. 2019. 延庆—张家口地区复杂地形冬季山谷风特征分析 [J]. 气象学报, 77(3): 475−488. doi: 10.11676/qxxb2019.033Jia Chunhui, Dou Jingjing, Miao Shiguang, et al. 2019. Analysis of characteristics of mountain–valley winds in the complex terrain area over Yanqing–Zhangjiakou in the winter [J]. Acta Meteor. Sinica (in Chinese), 77(3): 475−488. doi: 10.11676/qxxb2019.033 [15] Joe P, Lee G, Kim K. 2021. Measurement report: Strong valley wind events during the international collaborative experiment–Pyeongchang 2018 Olympic and Paralympic winter games project [J]. Atmospheric Chemistry and Physics Discussions, 2021: 1−32. doi: 10.5194/acp-2021-620 [16] Kingma D P, Ba J. 2015. Adam: A method for stochastic optimization [C]//Proceedings of the 3rd International Conference on Learning Representations. San Diego, CA, USA: ICLR. [17] 孔令彬, 赵艳茹, 王聚杰, 等. 2014. 基于支持向量机风速订正方法的研究 [J]. 西南大学学报(自然科学版), 36(5): 194−200. doi: 10.13718/j.cnki.xdzk.2014.05.029Kong Lingbin, Zhao Yanru, Wang Jujie, et al. 2014. Modification of wind speed based on SVM [J]. Journal of Southwest University (Natural Science) (in Chinese), 36(5): 194−200. doi: 10.13718/j.cnki.xdzk.2014.05.029 [18] Lameski P, Zdravevski E, Mingov R, et al. 2015. SVM parameter tuning with grid search and its impact on reduction of model over-fitting [C]//Proceedings of the 15th International Conference on Rough Sets, Fuzzy Sets, Data Mining and Granular Computing. Tianjin, China. Springer, 464–474. doi:10.1007/978-3-319-25783-9_41 [19] LeCun Y, Bottou L, Bengio Y, et al. 1998. Gradient-based learning applied to document recognition [J]. Proceedings of the IEEE, 86(11): 2278−2324. doi: 10.1109/5.726791 [20] Li H C, Yu C, Xia J J, et al. 2019. A model output machine learning method for grid temperature forecasts in the Beijing area [J]. Adv. Atmos. Sci., 36(10): 1156−1170. doi: 10.1007/s00376-019-9023-z [21] 黎静. 2016. 基于NWP-CFD的风资源评估和短期预报 [D]. 哈尔滨工业大学硕士学位论文.Li Jing. 2016. The wind resource assessment and short-term forecast based on NWP-CFD [D]. M. S. thesis (in Chinese), Harbin Institute of Technology. [22] 李林, 李卫林, 王振宇, 等. 2006. 青藏高原东边缘山区极值风速推算方法的研究 [J]. 高原气象, 25(6): 1028−1033. doi: 10.3321/j.issn:1000-0534.2006.06.008Li Lin, Li Weilin, Wang Zhenyu, et al. 2006. Research on reckoned method of maximum wind speed on the eastern edge of Qinghai–Xizang Plateau [J]. Plateau Meteorology (in Chinese), 25(6): 1028−1033. doi: 10.3321/j.issn:1000-0534.2006.06.008 [23] 李艳玲. 2016. 图像的最近邻缩放原理及实现 [J]. 长治学院学报, 33(5): 31−32. doi: 10.3969/j.issn.1673-2014.2016.05.010Li Yanling. 2016. Principle and implementation of nearest neighbor scaling of image [J]. Journal of Changzhi University (in Chinese), 33(5): 31−32. doi: 10.3969/j.issn.1673-2014.2016.05.010 [24] Lin Y J, Kruger U, Zhang J P, et al. 2015. Seasonal analysis and prediction of wind energy using random forests and ARX model structures [J]. IEEE Trans. Control Syst. Technol., 23(5): 1994−2002. doi: 10.1109/TCST.2015.2389031 [25] Lorenz E N. 1965. A study of the predictability of a 28-variable atmospheric model [J]. Tellus, 17(3): 321−333. doi: 10.3402/tellusa.v17i3.9076 [26] Lu C G, Yuan H L, Schwartz B E, et al. 2007. Short-range numerical weather prediction using time-lagged ensembles [J]. Wea. Forecasting, 22(3): 580−595. doi: 10.1175/WAF999.1 [27] 毛开银, 赵长名, 何嘉. 2020. 基于XGBoost的10 m风速订正研究 [J]. 成都信息工程大学学报, 35(6): 604−609. doi: 10.16836/j.cnki.jcuit.2020.06.004Mao Kaiyin, Zhao Changming, He Jia. 2020. A research for 10 m wind speed prediction based on XGBoost [J]. Journal of Chengdu University of Information Technology (in Chinese), 35(6): 604−609. doi: 10.16836/j.cnki.jcuit.2020.06.004 [28] McGovern A, Elmore K L, Gagne II D J, et al. 2017. Using artificial intelligence to improve real-time decision–making for high-impact weather [J]. Bull. Amer. Meteor. Soc., 98(10): 2073−2090. doi: 10.1175/BAMS-D-16-0123.1 [29] Mitchell M J, Ancell B, Lee J A, et al. 2020. Configuration of statistical postprocessing techniques for improved low-level wind speed forecasts in west Texas [J]. Wea. Forecasting, 35(1): 129−147. doi: 10.1175/WAF-D-18-0186.1 [30] Mu M, Duan W S, Wang J C. 2002. The predictability problems in numerical weather and climate prediction [J]. Adv. Atmos. Sci., 19(2): 191−204. doi: 10.1007/s00376-002-0016-x [31] 聂若鹰. 2007. 风力等级划分参考表[J]. 气象水文海洋仪器, (1): 67.Nie Ruoying. 2007. Reference table for wind grade classification [J]. Meteorological, Hydrological and Marine Instruments (in Chinese), (1): 67. [32] 任萍, 陈明轩, 曹伟华, 等. 2020. 基于机器学习的复杂地形下短期数值天气预报误差分析与订正 [J]. 气象学报, 78(6): 1002−1020. doi: 10.11676/qxxb2020.060Ren Ping, Chen Mingxuan, Cao Weihua, et al. 2020. Error analysis and correction of short-term numerical weather prediction under complex terrain based on machine learning [J]. Acta Meteor. Sinica (in Chinese), 78(6): 1002−1020. doi: 10.11676/qxxb2020.060 [33] Ronneberger O, Fischer P, Brox T. 2015. U-Net: Convolutional networks for biomedical image segmentation [C]//Proceedings of the 18th International Conference on Medical Image Computing and Computer-Assisted Intervention. Munich, Germany: Springer, 234–241. doi:10.1007/978-3-319-24574-4_28 [34] Shaw W J, Berg L K, Cline J, et al. 2019. The second wind forecast improvement project (WFIP2): General overview [J]. Bull. Amer. Meteor. Soc., 100(9): 1687−1699. doi: 10.1175/BAMS-D-18-0036.1 [35] Shi W Z, Caballero J, Huszár F, et al. 2016. Real-time single image and video super-resolution using an efficient sub–pixel convolutional neural network [C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, NV, USA: IEEE, 1874–1883. doi:10.1109/CVPR.2016.207 [36] 孙全德, 焦瑞莉, 夏江江, 等. 2019. 基于机器学习的数值天气预报风速订正研究 [J]. 气象, 45(3): 426−436. doi: 10.7519/j.issn.1000-0526.2019.03.012Sun Quande, Jiao Ruili, Xia Jiangjiang, et al. 2019. Adjusting wind speed prediction of numerical weather forecast model based on machine learning methods [J]. Meteorological Monthly (in Chinese), 45(3): 426−436. doi: 10.7519/j.issn.1000-0526.2019.03.012 [37] 王在文, 陈敏, Monache L D, 等. 2019. 相似集合预报方法在北京区域地面气温和风速预报中的应用 [J]. 气象学报, 77(5): 869−884.Wang Zaiwen, Chen Min, Monache L D, et al. 2019. Application of analog ensemble method to surface temperature and wind speed prediction in Beijing area [J]. Acta Meteor. Sinica (in Chinese), 77(5): 869−884. [38] Wilczak J M, Stoelinga M, Berg L K, et al. 2019. The second wind forecast improvement project (WFIP2): Observational field campaign [J]. Bull. Amer. Meteor. Soc., 100(9): 1701−1723. doi: 10.1175/BAMS-D-18-0035.1 [39] 乌日柴胡, 王建捷, 孙靖, 等. 2019. 北京山区与平原冬季近地面风的精细观测特征 [J]. 气象学报, 77(6): 1107−1123. doi: 10.11676/qxxb2019.059Wu Richaihu, Wang Jianjie, Sun Jing, et al. 2019. An observational investigation of fine features of near surface winds in winter over Beijing area [J]. Acta Meteor. Sinica (in Chinese), 77(6): 1107−1123. doi: 10.11676/qxxb2019.059 [40] 伍荣生, 谈哲敏, 王元. 2007. 我国业务天气预报发展的若干问题思考 [J]. 气象科学, 27(1): 112−118. doi: 10.3969/j.issn.1009-0827.2007.01.017Wu Rongsheng, Tan Zhemin, Wang Yuan. 2007. Discussions on the scientific and technological development of Chinese operation weather forecast [J]. Scientia Meteorologica Sinica (in Chinese), 27(1): 112−118. doi: 10.3969/j.issn.1009-0827.2007.01.017 [41] Xia J J, Li H C, Kang Y Y, et al. 2020. Machine learning-based weather support for the 2022 winter Olympics [J]. Adv. Atmos. Sci., 37(9): 927−932. doi: 10.1007/s00376-020-0043-5 [42] 杨璐, 陈敏, 陈明轩, 等. 2019. 高时空分辨率三维风场在强对流天气临近预报中的融合应用研究 [J]. 气象学报, 77(2): 243−255. doi: 10.11676/qxxb2019.010Yang Lu, Chen Min, Chen Mingxuan, et al. 2019. Fusion of 3D high temporal and spatial resolution wind field and its application in nowcasting of severe convective weather [J]. Acta Meteor. Sinica (in Chinese), 77(2): 243−255. doi: 10.11676/qxxb2019.010 [43] 杨璐, 宋林烨, 荆浩, 等. 2022. 复杂地形下高精度风场融合预报订正技术在冬奥会赛区风速预报中的应用研究 [J]. 气象, 48(2): 162−176. doi: 10.7519/j.issn.1000-0526.2021.092902Yang Lu, Song Linye, Jing Hao, et al. 2022. Fusion prediction and correction technique for high-resolution wind field in Winter Olympic Games area under complex terrain [J]. Meteorological Monthly (in Chinese), 48(2): 162−176. doi: 10.7519/j.issn.1000-0526.2021.092902 [44] 曾庆存. 1978. 计算稳定性的若干问题 [J]. 大气科学, 2(3): 181−191. doi: 10.3878/j.issn.1006-9895.1978.03.01Zeng Qingcun. 1978. Some aspects of the computational stability [J]. Chinese Journal of Atmospheric Sciences (Scientia Atmospherica Sinica) (in Chinese), 2(3): 181−191. doi: 10.3878/j.issn.1006-9895.1978.03.01 [45] 张延彪, 陈明轩, 韩雷, 等. 2022. 数值天气预报多要素深度学习融合订正方法 [J]. 气象学报, 80(1): 153−167. doi: 10.11676/qxxb2021.066Zhang Yanbiao, Chen Mingxuan, Han Lei, et al. 2022. Multi–element deep learning fusion correction method for numerical weather prediction [J]. Acta Meteor. Sinica (in Chinese), 80(1): 153−167. doi: 10.11676/qxxb2021.066 [46] 智协飞, 王田, 季焱. 2020. 基于深度学习的中国地面气温的多模式集成预报研究 [J]. 大气科学学报, 43(3): 435−446. doi: 10.13878/j.cnki.dqkxxb.20200219003Zhi Xiefei, Wang Tian, Ji Yan. 2020. Multimodel ensemble forecasts of surface air temperature over China based on deep learning approach [J]. Trans. Atmos. Sci. (in Chinese), 43(3): 435−446. doi: 10.13878/j.cnki.dqkxxb.20200219003 [47] Zhou Z W, Siddiquee M R, Tajbakhsh N, et al. 2020. UNet++: Redesigning skip connections to exploit multiscale features in image segmentation [J]. IEEE Transactions on Medical Imaging, 39(6): 1856−1867. doi: 10.1109/TMI.2019.2959609 -