收藏 | 设为首页 广告服务 | 联系我们 | 在线留言

干旱区地理 ›› 2026, Vol. 49 ›› Issue (11): 2255-2268.doi: 10.12118/j.issn.1000-6060.2025.773 cstr: 32274.14.ALG2025773

• 气候与水文 • 上一篇    下一篇

面向风电爬坡事件的WRF风速预报订正——XGBoost特征筛选与DNN预测

贾晓红1(), 徐丽娜1(), 韩丽娜1, 魏巍2   

  1. 1 内蒙古自治区气象服务中心, 内蒙古 呼和浩特 010051
    2 内蒙古自治区气候中心, 内蒙古 呼和浩特 010051
  • 收稿日期:2025-11-27 修回日期:2026-01-16 出版日期:2026-11-25 发布日期:2026-11-13
  • 通讯作者: 徐丽娜(1981-),女,学士,正研级高级工程师,主要从事风能太阳能预报技术研究与应用. E-mail: wutong2829@sina.com
  • 作者简介:贾晓红(1993-),女,硕士,工程师,主要从事新能源气象研究. E-mail: jiaxh22@sina.com
    第一联系人:

    论文构思:徐丽娜;数据管理:韩丽娜;规范分析:贾晓红;获取资助:徐丽娜;调查研究:韩丽娜;方法设计:贾晓红;项目管理:徐丽娜;提供资源:徐丽娜;软件编程:贾晓红;监督指导:徐丽娜;验证复核:贾晓红;可视化:魏巍;初稿撰写:贾晓红;修订编辑:贾晓红。

  • 基金资助:
    内蒙古自治区“揭榜挂帅”项目(2024JBGS0054);中国气象局创新发展专项(CXFZ2024J068);内蒙古自治区自然科学基金项目(2025ZDLH008);内蒙古自治区自然科学基金项目(2025QN04046)

WRF wind speed forecast correction for wind power ramp events: Feature selection with XGBoost and prediction with DNN

JIA Xiaohong1(), XU Lina1(), HAN Lina1, WEI Wei2   

  1. 1 Inner Mongolia Service Center of Meteorology, Hohhot 010051, Inner Mongolia, China
    2 Inner Mongolia Climate Center, Hohhot 010051, Inner Mongolia, China
  • Received:2025-11-27 Revised:2026-01-16 Published:2026-11-25 Online:2026-11-13

摘要:

风速的剧烈变化是导致风电场功率爬坡事件的直接原因,因此研发高精度的风速预测模型至关重要。选取内蒙古西部地区某风电场为研究对象,利用2023年中尺度天气预报模式(WRF)产品和实测风速,分析了风电场风速格点预报误差,通过筛选影响风速的特征因子和深度学习预测,最终提出了一种基于极限梯度提升-深度神经网络(XGBoost-DNN)的爬坡事件风速预测模型。结果表明:(1)1月、4月、7月和10月WRF模式预测与实测风速的平均偏差(ME)、平均绝对误差(MAE)、均方根误差(RMSE)和相关系数(r)分别为-0.68~0.37 m∙s-1、1.91~2.41 m∙s-1、2.40~3.08 m∙s-1和0.43~0.61,发现WRF模式对局部风速变化的捕捉能力有限,难以适应风速的快速波动。(2)考虑主导风向,对风速影响的气象因子特征贡献度与地理位置密切相关,处于下风向、近距离的格点预报值与风速实际值相关性最强。对于爬坡过程,湿度相关因子预报在各月体现出对风速明显不同的贡献。(3)特征因子数量对XGBoost-DNN模型的影响在不同月份有所区别,根据各月表现最好的特征数量对爬坡事件风速预测,以爬坡数据为训练样本的XGBoost-DNN模型上、下爬坡阶段MAE分别为1.24 m∙s-1、1.35 m∙s-1,较WRF模式分别减小41.5%、37.8%,其中4月误差最大,10月误差最小。研究结果可为进一步提高风电场风速和功率预报精度提供相关参考。

关键词: WRF模式, XGBoost-DNN模型, 风电爬坡, 风速预测

Abstract:

Dramatic changes in wind speed are the direct cause of wind power ramp events in wind farms, and thus the development of a high-precision wind speed prediction model is essential. This study focuses on a wind farm located in the western region of Inner Mongolia, China. We analyze grid point forecast errors in wind speed using forecast data from the weather research and forecasting (WRF) model alongside observed wind speed data from 2023. By identifying key factors affecting wind speed and utilizing deep learning techniques, we propose a ramping event wind speed prediction model based on an extreme gradient boosting-deep neural network (XGBoost-DNN). The results reveal the following: (1) The mean error (ME), mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient (r) for the predicted and observed wind speeds from the WRF model in January, April, July, and October are -0.68-0.37 m∙s-1, 1.91-2.41 m∙s-1, 2.40-3.08 m∙s-1 and 0.43-0.61, respectively. We found that the WRF model has limited capabilities in capturing local wind speed variations and struggles to adapt to rapid wind speed fluctuations. (2) Considering the prevailing wind direction, the influence of meteorological factors on wind speed is closely related to geographic location, with the strongest correlation between predicted and observed wind speeds occurring at grid points located downwind and nearby. For ramp events, humidity-related factors show significantly varying contributions to wind speed across different months. (3) The influence of feature factors on the XGBoost-DNN model differs by month. By selecting the optimal number of features for each month, the XGBoost-DNN model, trained on ramp data, achieves an MAE of 1.24 m∙s-1 and 1.35 m∙s-1 for the up- and down-ramp events, respectively, representing reductions of 41.5% and 37.8% compared to the WRF model. The highest error occurs in April, whereas the lowest occurs in October. These results provide valuable insights for enhancing the accuracy of wind speed and power forecasting in wind farms.

Key words: WRF model, XGBoost-DNN model, wind power ramp, wind speed prediction