高级检索
门晓磊, 焦瑞莉, 王鼎, 赵晨光, 刘亚昆, 夏江江, 李昊辰, 严中伟, 孙建华, 王立志. 基于机器学习的华北气温多模式集合预报的订正方法[J]. 气候与环境研究, 2019, 24(1): 116-124. DOI: 10.3878/j.issn.1006-9585.2018.18049
引用本文: 门晓磊, 焦瑞莉, 王鼎, 赵晨光, 刘亚昆, 夏江江, 李昊辰, 严中伟, 孙建华, 王立志. 基于机器学习的华北气温多模式集合预报的订正方法[J]. 气候与环境研究, 2019, 24(1): 116-124. DOI: 10.3878/j.issn.1006-9585.2018.18049
Xiaolei MEN, Ruili JIAO, Ding WANG, Chenguang ZHAO, Yakun LIU, Jiangjiang XIA, Haochen LI, Zhongwei YAN, Jianhua SUN, Lizhi WANG. A Temperature Correction Method for Multi-model Ensemble Forecast in North China Based on Machine Learning[J]. Climatic and Environmental Research, 2019, 24(1): 116-124. DOI: 10.3878/j.issn.1006-9585.2018.18049
Citation: Xiaolei MEN, Ruili JIAO, Ding WANG, Chenguang ZHAO, Yakun LIU, Jiangjiang XIA, Haochen LI, Zhongwei YAN, Jianhua SUN, Lizhi WANG. A Temperature Correction Method for Multi-model Ensemble Forecast in North China Based on Machine Learning[J]. Climatic and Environmental Research, 2019, 24(1): 116-124. DOI: 10.3878/j.issn.1006-9585.2018.18049

基于机器学习的华北气温多模式集合预报的订正方法

A Temperature Correction Method for Multi-model Ensemble Forecast in North China Based on Machine Learning

  • 摘要: 模式预报的订正是决定局地天气预报结果的一个重要步骤,基于机器学习的后处理模型近年来开始崭露头角。本文发展了基于岭回归(Ridge)、随机森林(Random Forest,RF)和深度学习(Deep Learning,DL)的3种后处理模型,基于中国气象局(CMA)的BABJ模式、欧洲中期天气预报中心(ECMWF)的ECMF模式、日本气象厅(JMA)的RJTD模式和NCEP的KWBC模式这4个数值天气预报模式2014年2月至2016年9月(训练期)近地面2 m气温预报和实况资料确定各模型参数,进而对2016年10月至2017年9月(预报期)华北地区(38°N~43°N,113°E~119°E)的逐日地面2 m气温预报进行了多模式集合预报分析。采用均方根误差对预报效果进行评估,这3种后处理模型的预报效果和4个数值天气预报模式以及通常的多模式集合平均(Ensemble Mean,EMN)的预报效果的对比表明:1)随着预报时长增加,4个数值预报模式及各种后处理模型的均方根误差均呈上升趋势;但区域平均而言,Ridge、RF和DL的预报效果在任何预报时长上都明显优于EMN和单个天气预报模式;特别是前几天的短期预报DL的预报效果更好,中后期预报Ridge的预报效果略好。2)华北地区的东南部均方根误差较小,其余格点上均方根误差较高,从空间分布而言,DL的订正预报效果最好,3种机器学习模型的误差在1.24~1.26℃之间,而EMN的误差达1.69℃。3)夏季各种方法的预报效果都较好,冬季预报效果都较差;但是Ridge、RF和DL的预报效果明显优于EMN,这3种模型预报的平均均方根误差在2.15~2.18℃之间,而EMN的平均均方根误差达2.45℃。

     

    Abstract: Post-forecast data processing is critical for obtaining reliable local weather forecast. In this study, the authors developed three post-processing models based on ridge regression (Ridge), random forest (RF), and deep learning (DL) methods. The post-processing models were trained by observational and forecast data of daily 2-m above surface air temperature in North China (38°N-43°N, 113°E-119°E) from four numerical weather forecast (NWF) models (BABJ model from China Meteorological Administration, ECMF model from ECMWF, RJTD model from Japan Meteorological Agency, and KWBC model from NCEP, respectively), for the training period from February 2014 to September 2016, and then applied to the forecast period from October 2016 to September 2017. The forecast results of the post-processing models together with those of commonly-used multi-model ensemble mean (EMN) and individual NWF models were evaluated according to the root-mean-square error (RMSE). The main results are as follows:1) For the region as a whole, with the increase in the forecast lead time, all the NWF models, EMN and the post-processing models exhibit increasing RMSEs, but the RMSEs of the three post-processing models are all significantly smaller than those of EMN and individual NWF models; especially, DL is slightly better for the short-term (the first few days) forecast and RF is slightly better for the longer-term prediction. 2) The RMSEs are relatively smaller in the southeastern part of North China, approximately in the range of (38°N-41°N, 115.5°E-119°E) than else where; on average, DL is slightly better, and the RMSEs of the three machine learning models are between 1.24℃ and 1.26℃, while the EMN error is 1.69℃. 3) There are seasonal differences:The results of all the models are relatively good for the summer, but poor in general for the winter. All the three post-processing models perform better than EMN and individual NWF models, with a smallest average RMSE of 2.15℃ for Ridge compared with 2.45℃ for EMN.

     

/

返回文章
返回