自组织网络与广义回归网络耦合的副热带高压指数预测

PREDICTING THE SUBTROPICAL HIGH INDEX BY COUPLING SELF-ORGANIZING FEATURE MAP AND GENERALIZED REGRESSION NEURAL NETWORK

利用亚洲夏季风系统中各成员变化活动与西太平洋副高存在的不同程度的时延相关性,从1995～2004年NCEP/NCAR逐日再分析资料中,提取了亚洲夏季风系统各成员变化活动的特征指标及其对应的超前三候的西太平洋副高(简称副高)面积和脊线指数。在此基础之上,建立了自组织网络与径向基函数网络串级耦合的副高指数预测模型。该模型首先用自组织网络对各指标样本按其自身相似原则进行无监督分类,随后用广义回归网络分别对分类出的各指数样本子集进行有监督的训练建模和预测。模型的预测试验结果表明:副高指数的预测结果与其实际值之间的相关系数达到0.89,明显优于单一的神经网络模型预测效果。

Based on time-lag correlation between the activity of individual members of the Asian monsoon system and the Western Pacific Subtropical High, character index of the members＇ variation and the corresponding subtropical high three pentads ahead were extracted from the NCEP/NCAR daily reanalysis dataset during the ten years from 1995 to 2004. The subtropical high index predicting model was built that coupled Self-Organizing Feature Map（SOFM） and Generalized Regression Neural Network（GRNN） in series. The model sorted the index samples without supervision according to the characteristic similarity by the SOFM Neural Network, and then built and trained the model by the GRNN based on the index sample subset. The prediction experiment results show that the correlation coefficient between the predicted result and the actual value reached 0.89 and that experiment results were much better than that by single Neural Network prediction model.