基于机器学习的多源实况分析产品和观测数据融合应用试验

Fusion and application experiment of machine learning based multi-sourcereal-time analysis products and observation data

利用中国气象局公共气象服务中心地面实况专业服务产品(CARAS_SUR1 km,简记为“CAR”)、国家气象信息中心多源融合实况分析产品(ART_1 km,简记为“ART”)、全国雷达反演降水产品(简记为“RAD”)、风云四号卫星反演降水产品(简记为“SAT”)以及全国气象观测站逐小时资料,应用机器学习方法建立了基于选定位置气温、降水、风向、风速要素的实况融合应用模型(简记为“GBDT模型”)。15 d逐时GBDT融合产品的全国分区域检验结果表明:GBDT气温融合产品在东北、华北、西北、华中、新疆、西藏6个区域较CAR和ART均有改进,在西藏的改进最明显,在华东和西南GBDT融合产品优于ART而逊于CAR,在华南和内蒙古GBDT融合产品误差较ART、CAR略有增加;GBDT降水融合产品在样本偏少的内蒙古较ART、CAR误差略有增加,其他区域有改进或基本相当;GBDT风速、风向融合产品较ART、CAR均有较大改进。试验结果表明,机器学习方法可应用于融合多源实况分析产品和观测数据,以开展选定位置气温、降水、风向、风速要素的实况气象信息服务。

Based on machine learning,an application model(GBDT model)of real-time fusion on temperature,precipitation,wind direction,and wind speed at selected locations is developed by using the professional service product(CAR)of Public Meteorological Service Centre of China Meteorological Administration,the multi-source fusion observation analysis data(ART)of National Meteorological Information Centre,the nationwide radar precipitation retrieval product(RAD),the Fengyun-4 satellite precipitation retrieval product(SAT),and the hourly data of nationwide meteorological observation stations.The regional inspection results of the 15-d hourly GBDT fusion product throughout the country are as follows.The GBDT temperature fusion product improves compared to CAR and ART in 6 regions:Northeast China,North China,Northwest China,Central China,Xinjiang,and Tibet,with the most significant improvement in Tibet.In East China and Southwest China,GBDT fusion product is superior to ART,but inferior to CAR,and its error slightly increases compared to ART and CAR in South China and Inner Mongolia.The error of GBDT precipitation fusion product has a slight increase compared to ART and CAR in Inner Mongolia,where there are fewer samples,while in other areas,there are improvements or they are basically equivalent.The GBDT wind speed and direction fusion products have significant improvements compared to ART and CAR.The experiment results indicate that the machine learning method can be applied to fuse multi-source real-time analysis products and observation data,providing real-time meteorological information service of temperature,precipitation,wind direction,and wind speed at selected locations.