基于机器学习的格点气温预报订正方法

使用2017年9月至2021年3月国家级业务化运行的智能网格实况分析产品和欧洲中期天气预报中心全球模式(EC)产品,根据湖北省的地理分布特征构建6个分区,采用基于LightGBM机器学习算法建立的气温预报方法,生成湖北省0.05°×0.05°格点气温预报产品。利用2021年4—9月的预报产品和格点实况资料进行检验,结果表明:基于机器学习的气温预报方法(MLT)取得了较好的预报效果,其在0～72 h时效内优于中央气象台下发的气温精细化指导预报(SCMOC)和EC产品;MLT在山区的误差较平原大,但山区的订正幅度大于平原,日最高气温的订正幅度大于日最低气温的订正幅度;4—9月MLT、SCMOC、EC产品的平均绝对误差(MAE)日变化都呈现了白天偏高、夜间偏低、午后凸起的单峰特征,MLT的MAE值较SCMOC和EC产品的更低,并且在转折性天气中仍具有优势;站点检验与格点检验结论一致,基于格点建模的气温预报产品对站点预报同样得到了订正。机器学习在格点气温的模式订正方面可以作为一个行之有效的手段。

Gridded Temperature Forecast Correction Method Based on Machine Learning

Using the smart grid reality analysis product (CLDAS-V2.0) and the European Centre for 〖JP〗Medium-Range Weather Forecast Global Model (EC) product, which were operated at the national level from September 2017 to March 2021, six regions are constructed according to the geographical distribution characteristics of Hubei Province. The temperature forecast model established by LightGBM machine learning algorithm is used to generate 0.05°×0.05° gridded temperature forecast products in Hubei Province, and the forecast products are verified by the forecast products and grid data from April to September 2021. The results show that the temperature prediction method (MLT) based on machine learning has achieved a good forecast effect, and MLT is superior to SCMOC and EC model products in 0-72 h lead time. The error of MLT in mountain area is larger than that in plain area, but the correction amplitude of MLT in mountain area is larger than that in plain area, and the correction amplitude of daily maximum temperature is larger than that of daily minimum temperature. The diurnal variations of MAE in MLT, SCMOC and EC model products from April to September present the characteristics of higher in daytime, lower over night, and convex single-peak in afternoon. The MAE value of MLT is lower than that of SCMOC and EC model products, and still has an advantage in changeable weather. The results of site test and grid test are consistent, and the temperature forecast product based on grid modeling is also corrected. Machine learning can be used as an effective means to correct the pattern of gridded temperature.