多元混沌时间序列的加权极端学习机预测  

Multivariate chaotic time series prediction based on weighted extreme learning machine

作　　者：韩敏[1] 王新迎[1] 

机构地区：[1]大连理工大学电子信息与电气工程学部,辽宁大连116023

出　　处：《控制理论与应用》2013年第11期1467-1472,共6页Control Theory & Applications

基　　金：国家自然科学基金资助项目(61074096)

摘　　要：针对多元混沌时间序列具有强非线性,难以建立数学模型进行准确预测的问题,本文提出一种加权极端学习机预测算法.首先对多元混沌时间序列进行相空间重构,并根据相空间中输入数据对预测误差的影响施加不同的权重.然后,提出一种支持向量极端学习机预测模型,具有支持向量机的核映射表达能力以及极端学习机的一步快速训练能力,因此训练简便且具有较好的泛化性能.所提算法具有和训练样本三次方成正比的计算复杂度,因此适用于102?103样本规模的平稳时间序列.基于Lorenz混沌时间序列和年太阳黑子和黄河年径流混沌时间序列预测的仿真结果证明所提算法的有效性.Considering the strong nonlinear property of the multivariate chaotic time series and difficulties in mathe- matical model building for accurate prediction, we propose a weighted extreme learning machine （ELM） prediction model. Firstly, the multivariate chaotic time series are reconstructed in the phase space, and the instances in prediction window are weighted according to their influence on the prediction errors. A support-vector extreme learning machine prediction model is proposed, which combines both the kernel mapping ability of support-vector machines and the one-step fast train- ing advantage of extreme learning machine, so that it is easy to conduct and has good generalization performance. The computational complexity of the proposed algorithm is proportional to the cube of the training sample. Therefore, it is suitable for stationary time series with 10^2 ~ 10^3 sample size. The effectiveness of the proposed model are demonstrated by simulation results based on Lorenz chaotic time series, annual sunspots, and the runoff of the Yellow River chaotic time series.

关 键 词：时间序列 预测 极端学习机 支持向量机 样本加权 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]