欢迎访问中国科学院大学学报,今天是

中国科学院大学学报 ›› 2015, Vol. 32 ›› Issue (5): 588-593.DOI: 10.7523/j.issn.2095-6134.2015.05.003

• 数学与物理学 • 上一篇    下一篇

利用灰色关联极限学习机预报日长变化

雷雨1,2,3, 蔡宏兵1,2, 赵丹宁1,3   

  1. 1. 中国科学院国家授时中心, 西安 710600;
    2. 中国科学院时间频率基准重点实验室, 西安 710600;
    3. 中国科学院大学, 北京 100049
  • 收稿日期:2014-11-04 修回日期:2015-03-12 发布日期:2015-09-15
  • 通讯作者: 雷雨
  • 基金资助:

    中国科学院西部之光联合学者项目(201491)资助

Prediction of length-of-day variation using grey relational analysis and extreme learning machine

LEI Yu1,2,3, CAI Hongbing1,2, ZHAO Danning1,3   

  1. 1. National Time Service Center, Chinese Academy of Sciences, Xi'an 710600, China;
    2. Key Laboratory of Time and Frequency Primary Standards, Chinese Academy of Sciences, Xi'an 710600, China;
    3. University of Chinese Academy of Sciences, Beijing 100049, China
  • Received:2014-11-04 Revised:2015-03-12 Published:2015-09-15

摘要:

针对日长变化难以用精确模型进行预报的问题,将一种新型人工神经网络——极限学习机(extreme learning machine, ELM)用于日长变化预报中.首先针对时间序列预测问题中存在的嵌入维数选取和网络结构设计问题,提出一种基于灰色关联分析(grey relational analysis, GRA)的ELM算法(GRA-ELM),该算法将灰色关联分析输入节点选取嵌入到ELM网络的训练过程中,同时完成嵌入维数和隐层节点规模的确定.然后根据日长变化数据的特点对其进行预处理,建立一种能够高精度、近实时预报日长变化的GRA-ELM预报模型.最后将GRA-ELM模型的预报结果同标准ELM、反向传播神经网络、广义回归神经网络和地球定向参数预报比较竞赛的结果进行比较.结果表明,通过本方法得到的日长变化较其他方法在精度上有较大改善.

关键词: 日长变化, 预报, 灰色关联分析, 极限学习机, 神经网络

Abstract:

Due to the time-varying characteristics of length-of-day (LOD), it is difficult to model LOD variations with a deterministic model. We employ a new type of artificial neural networks (ANN)-extreme learning machine (ELM) to predict LOD variations. In order to solve the problems of embedding dimension selection and network topology design, a training algorithm for ELM based on grey relational analysis (GRA) is first proposed. It optimizes the input and hidden layers simultaneously. Secondly, the values of LOD variation are preprocessed and a GRA-ELM model is then set up to accurately forecast LOD variation in near real-time. Finally, the prediction results are analyzed and compared with those obtained by the back propagation neural networks, generalization regression neural networks and Earth orientation parameters prediction comparison campaign. The results show that the prediction accuracy of our method is equal to or even better than those of the other prediction methods. The developed method is easy to use.

Key words: length-of-day (LOD) variations, prediction, grey relational analysis (GRA), extreme learning machine (ELM), neural networks (NN)

中图分类号: