冬奥会复杂山地百米尺度10m风速预报的机器学习订正对比试验.pdf

ata100mResolutionuntainousAreasoftheWinterOlympicic Games J.Chinese Journal of Atmospheric Sciences(in Chinese),47(3):Jingfeng,SONGLinye,CHENuan,etal.z0z3arative Machine Learning-Based Correction Experiment for a 10 m Wind SpeedForecastXU徐景峰，宋林烨，陈明轩，等.2 0 2 3.冬奥会复杂山地百米尺度10 m风速预报的机器学习订正对比试验 .大气科学，47(3)：8 0 5-8 2 4.May20232023年5月ChinesSciencesVol.47 No.3学科第47 卷第3期805-824.doi:10.3878/j.issn.1006-9895.2209.22117冬奥会复杂山地百米尺度10 m风速预报的机器学习订正对比试验宋林烨2徐景峰1,22陈明轩杨璐？韩雷！1中国海洋大学，青岛2 6 6 10 02北京城市气象研究院，北京10 0 0 8 9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10 0 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。关键词同百米尺度预报复杂山地机器学习风速订正文章编号号10 0 6-9 8 9 5(2 0 2 3)0 3-0 8 0 5-2 0中图分类号P45文献标识码Adoi:10.3878/j.issn.1006-9895.2209.22117Comparative Machine Learning-Based Correction Experiment for a 10 mWind Speed Forecast at a 100 m Resolution in Complex MountainousAreas of the Winter Olympic GamesXU Jingfeng2,SONG Linye,CHEN Mingxuan,YANG Lu,and HAN Lei1OceanUniversityofChina,Qingdao2661002 Institute of Urban Meteorology,China Meteorological Administration,Beijing 100089收稿日期2022-07-04；网络预出版日期2023-02-25作者简介徐景峰，男，1997 年出生，硕士研究生，主要从事人工智能与气象应用方向的研究。E-mail:通讯作者宋林烨，E-mail:资助项目北京市自然科学基金项目8 2 12 0 2 5、8 2 2 2 0 51，国家重点研发计划2 0 18 YFF0300102，北京市气象局科技项目BMBKJ202004011，国家自然科学基金项目42 2 7 50 12Funded byy Beijing Natural Science Foundation(Grants 8212025,8222051),National Key Research and Development Program(Grant2018YFF0300102),Science and Technology Project of Beijing Meteorological Bureau(Grant BMBKJ202004011),National NaturalScience Foundation of China(Grant 42275012)806Vol.47ChinesernahericSciences47卷学科Abstract Based on a traditional machine learning algorithm(XGBoost),a deep learning algorithm(CU-Net),and thewinter wind speed data from 10 m near the ground with a resolution of 100 m,this paper studied and compared thecorrection methods for wind speed forecast deviation in the mountainous stations and surrounding areas of the Yanqingand Zhangjiakou competition areas(Beijing Winter Olympic Games)using the rapid-refresh integrated seamlessensemble(RISE)system.For station correction,the 10-m wind speed predicted by the RISE system is interpolated to thecorresponding automatic weather station.Subsequently,a separate XGBoost model is constructed for each classificationaccording to the wind speed rating table.Afterward,each interval model was combined to form L-XGBoost,using theroot mean square error and forecast accuracy as its scoring standard.Investigations revealed that the correction effect ofthe 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 speedprediction skills of the complex mountain stations.Subsequently,for the wind speed correction of the station and itssurrounding areas based on the CU-Net model,this paper constructed a new algorithm model(CU-Net+)by introducingthe CU-Net sub-networks with different depths,considering the influence of daily forecast errors and complex terrains onthe 10-m wind speed.This paper also constructed spatial small-area sample data,considering the automatic weatherstation as the center,to correct the wind speed prediction deviation of the RISE system.The test results indicated thatalthough both CU-Net and CU-Net+fully mined the wind field change rules in time and space dimensions,the windspeed correction results of the CU-Net+model performed better than those of the CU-Net model,effectively reducingthe grid wind speed prediction error of RISE products.Hence,introducing prediction error and complex terrain plays animportant positive role in the deviation correction of a surface 10 m wind speed-based investigation.Keywordss100 m scale forecast,Complex mountain,Machine learning,Wind speed correction1引言风与人类社会密不可分，关系着人们的日常生活、各类大型活动举办和公共安全，提高风速风向的精细化预报水平具有重要意义。众所周知，许多冬季运动项目，尤其是冬奥会雪上项目赛事都在地形十分复杂的山区举行，气象条件不仅直接关乎赛事的顺利举办，还关乎运动员水平发挥和生命安全，而这其中，高分辨率风速的精准预报就是关键之一。受地形强迫、地面摩擦和日照辐射等的影响，山区各种边界层内的小尺度大气脉动、风速风向的空间变化和日变化特征及其影响机制复杂多变，导致山区小尺度风场预报难度和偏差远大于平原地区（Be n o i t e t a l.，2 0 0 2；高登义等，2 0 0 3；贾春晖等，2019;Wilczak et al.,2019;Shaw et al.,2019;Joe etal.,2021）。因此，研究适用于复杂地形下的高分辨率风速预报偏差订正方法、提升高山区风场预报的准确性就显得尤为重要（Howardand Clark,2007;Mitchell et al.,2020)。当前风速预测的主要方法之一是数值天气预报(Numerical Weather Prediction，NWP）模式。NWP模式通过在给定初始和边界值的情况下求解大气运动和热力学方程组，预测未来一段时间的风速变化。然而NWP模式描述的大气物理运动过程有限且初始场不可能绝对准确（Lorenz，196 5；曾庆存，197 8;Muetal.,2002；丁瑞强和李建平，2 0 0 7；伍荣生等，2007;Fengetal.,2014；常俊