摘要
从业务需求出发，提出了面向气候模式产品释用的神经网络。选用主分量作为网络的输入和输出，大大减少了其节点数，重点突出了大尺度影响变化关系，提高了实际预测的稳定性；用全局寻优的遗传算法取代经典BP算法，为高质量的网络学习训练提供了保证；针对实际设计代价函数，保证了网络学习训练能适应气候模式产品释用的基本要求，学习目的更明确，针对性更强。分别以夏季（6～8月）NCEP／NCAR500 hPa高度场、国家气候中心海气耦合模式500hPa高度预测场主分量为外界输入信号，同期中国降水场、华中区域降水场主分量为网络输出信号，进行了拟合预测和独立预测试验。结果表明：用模式500hPa高度预测场主分量为外界输入信号，网络输出（降水场主分量）反演的中国、华中地区降水场预测距平与实况同号率，有可能接近用NCEP／NCAR500 hPa高度场主分量为外界输入信号相当的技巧水平。
Artificial neural networks for reexplanation and reanalysis of products of climate model are proposed according to professional requirements. As principal components are selected for the input and output of the networks, number of the nodes are decreased greatly, a large scale influence is stressed and prediction is stabilized. Because of the substitution of global genetic algorithms （GA） for classical BP the neural networks obtain perfect knowledge through enough practice. The cost function designed for vocational work ensures the practiced neural networks suitable for reexplanation and reanalysis of products of climate model. Practice of neural networks have a definite purpose exactly. Principal components of JJA （June, July, August） NCEP/NCAR 500 hPa geopotential height fields and 500 hPa geopotential height fields of National Climate Center coupled ocean-atmosphere model are selected for the input, and principal components of same time rainfall fields over both of China and the middle areas of China are selected for the output of the networks, both of imitative and untrained samples have tested. The results show thatusing principal components of 500 hPa geopotential height fields of model as the input, the fields of rainfall recovered by the output （principal components of rainfall fields） may approach the skill level of forecast using principal components of 500 hPa geopotential height fields of NCEP/NCAR as the input.
出处
《气候与环境研究》
CSCD
北大核心
2008年第5期681-687,共7页
Climatic and Environmental Research
基金
中国气象局气候研究开放实验室开放课题
武汉区域气象中心科技发展基金项目
关键词
模式产品释用
神经网络
遗传算法
可预测性
reexplanation and reanalysis of products of model, artificial neural networks, genetic algorithm, predictability
