极限学习机神经网络在短期降水预报中的应用  

Application of extreme learning machine neural network to forecast of short-term precipitation

作　　者：虞双吉[1] 苗春生[1] 王新 

机构地区：[1]南京信息工程大学大气科学学院,江苏南京210044 [2]华风气象传媒集团,北京100081

出　　处：《云南大学学报（自然科学版）》2013年第4期507-515,共9页Journal of Yunnan University(Natural Sciences Edition)

基　　金：国家科技支撑项目(2012BAH05B01);气象公益性行业专项(201206068);国家自然科学基金面上项目(41276033);江苏高校优势学科建设工程项目(PAPD)

摘　　要：数值预报产品释用是天气预报现代化发展方向,也是我国预报业务建设的重点之一.结合主成分分析方法从中国气象局T213资料中选取降水预报因子,利用极限学习机(Extreme Learning Machine,ELM)神经网络方法进行衢州单站逐日降水的分级预报研究.结果表明:主成分分析方法可以保留原来预报因子矩阵大部分信息并大幅降低矩阵维数,使新的预报因子之间相互正交;ELM方法不仅具有很强的历史拟合能力,独立样本所有降水预报TS评分也从T213模式预报0.3提高到0.8,特别是对大到暴雨有了较强预报能力;与相同条件下BP方法预报模型相比,ELM方法网络参数设定简单,训练速度快,历史样本的训练结果和独立样本的预报结果也都要优于BP方法.The interpretation and application of numerical forecast products is the direction of the modernization of the weather forecast and the key to improving the level of weather forecast. Using principle component analysis to choose the precipitation predictors in data of T213 Model. The predictors are used to daily categorical forecast of precipitation in Quzhou by Extreme Learning Machine （ELM）. The results show that the method of principle component analysis can simplify the matrix of predictors and get new predictors. The new predictors which are mutually orthogonal contain most of the information of original predictors. ELM neural network has a valid prediction ability to training sample, and the TS score of the rainfall forecast for testing sample is improved from 0.3 to 0. 8. Particularly for a heavy rain, ELM neural network shows strong prediction ability. To contrast with the method of Back Propagation （ BP）, ELM has more simple network setting and more fast computation speed. The prediction ability to training sample and testing sample of the ELM neural network are also better than BP neural network.

关 键 词：人工神经网络 极限学习机(ELM) 主成分分析 降水分级预报 

分 类 号：P456[天文地球—大气科学及气象学]