摘要
为了扩展气象业务中历史预报资料在集合预报产品中的应用性,本文尝试基于机器学习的后处理模型,对欧洲中期天气预报中心(ECMWF)ECMF模式的数值天气预报模式生成的2019年5月—8月历史再预报(训练期)数据进行各模型参数训练,可对2019年5月—8月实时延伸期月降水预报(预报期)进行集合预报分析,均选取华东地区(23.0°N~38.5°N,113.0°E~123.0°E)。采用均方根误差和绝对误差插值空间分布对预报订正效果进行评估,通过K-近邻算法(KNN)、Bagging算法、随机森林(RF)、梯度提升回归树(GBRT)、极端随机树(ET)这5种后处理回归模型的预报订正效果的对比表明:5种后处理模型的均方根误差随预报时效的增长呈现波动变化;受地形分布的影响,华东地区的北部均方根误差较小,从空间分布清楚得出,K-近邻算法的订正预报效果最好,对应的平均绝对误差在10~20 mm,其他4种机器学习模型的误差在50~100 mm。
In order to expand the ensemble forecast product application of historical forecast data in the meteorological service,this paper tries to train the model parameters of the historical re-forecast(training period)data generated by the numerical weather forecast mode of European Centre for Medium-Range Weather Forecasts(ECMWF)ECMF mode based on the post-processing model in machine learning.The model can perform ensemble forecast analysis on the real-time extended monthly precipitation forecast(forecast period)from May 2019 to August 2019.East China(23.0°N~38.5°N,113.0°E~123.0°E)is selected for the analysis.The root mean square error(RMSE)and absolute error interpolation spatial distribution are used to evaluate the prediction correction effect.The comparison of the forecast correction effects of the five post-processing regression models,namely K-nearest neighbor(KNN)algorithm,Bagging algorithm,random forest(RF),gradient boost regression tree(GBRT)and extremely randomized trees(ET),shows that the RMSE of the five post-processing models fluctuates with the increase of forecasting effectiveness,and in terms of the influence of the terrain distribution,the RMSE in the north of East China is small.From the spatial distribution,it is clear that the KNN algorithm has the best correction forecast effect,with the corresponding average absolute error of 10~20 mm.The error of the other four machine learning models is within 50~100 mm.
作者
陆振宇
孔小翠
张恒德
黄威
LU Zhenyu;KONG Xiaocui;ZHANG Hengde;HUANG Wei(School of Electronics and Information Engineering,Nanjing University of Information Science&Technology,Nanjing 210044,China;National Meteorological Center,Beijing 100081,China)
出处
《现代电子技术》
2022年第7期86-91,共6页
Modern Electronics Technique
基金
国家自然科学基金资助项目(61773220)
江苏省自然科学基金资助项目(BK20150523)。
关键词
集合平均
历史再预报
延伸期降水预报
K-近邻算法
机器学习
均方根误差
插值分布
ensemble average
historical re-forecast
extended monthly precipitation forecast
KNN
machine learning
RMSE
interpolation distribution
作者简介
陆振宇(1976-),男,江苏常州人,博士,教授,主要研究领域为模式识别、智能控制;孔小翠(1994-),女,河南信阳人,硕士研究生,主要研究方向为集合预报订正分析;张恒德(1977-),男,安徽含山人,博士,正研级高级工程师,主要从事环境气象研究;黄威(1986-),男,安徽宿州人,硕士,高级工程师,主要从事中期天气预报。
