基于全球预报系统气温的降尺度研究

Downscaling Study based on GFS Air Temperature

目前大多数数值预报的气温产品由于其粗略的空间分辨率,在局部尺度上具有局限性。为得到高分辨率的气温产品,利用美国国家环境预报中心(NCEP)2020年全球预报系统(GFS)气温预报数据,基于神经网络开发一种降尺度模型。将该方法应用于中国湖南省,成功将预测气温从0.25°×0.25°的网格数据降尺度为250 m×250 m的网格数据。利用97个国家级气象站对降尺度后的气温进行精度评估,其均方根误差为1.53℃,相关系数为0.982,偏差为-0.01℃。与降尺度前相比,降尺度后的气温与台站气温有更好的一致性,尤其是在高海拔地区。最后也将降尺度后的气温与ERA5-Land气温产品进行对比,其准确性优于ERA5-Land气温产品。

At present,most of the temperature products of numerical prediction have limitations on local scale due to their rough spatial resolution.In order to obtain high resolution temperature products,a downscaling model is developed based on neural network using the global forecast system(GFS)temperature forecast data from the National Environmental Prediction Center(NCEP)in 2020.The method was applied to Hunan Province,China,and the predicted temperature was successfully changed from 0.25°×0.25°grid data downscaling to 250 m×250 m grid data.The data used to evaluate the accuracy of the downscaling temperature include 97 national meteorological stations.The root mean square error was 1.53℃,the correlation coefficient was 0.982,and the deviation was-0.01℃.Compared with the temperature before the downscaling,the temperature after the downscaling is more consistent with the temperature at the station,especially at high altitude.Finally,the temperature after scaling down is also compared with the ERA5-Land temperature product,and the results show the accuracy is better than that of the ERA5-Land temperature product.