Abstract: To improve the on-line prediction precision of time serials under small initialization sample，a structure adaptive sequential regularized extreme learning machine(SA-SRELM) is proposed. For this new method，at the stage of on-line sequential learning，the output weight of SA-SRELM is renewed by different recursive methods according to training sample size. Meanwhile，when training sample size is added to a destined value，new samples are added and the oldest samples are removed to enhance the dynamic adaptability of prediction model. Based on above process，the prediction model is built. Three typical chaotic time series prediction examples are used to verified the validity of proposed method. In the end，the proposed method is used to predict aeroengine exhaust gas temperature. The results show that SA-SRELM can get better generalization performance compared with regularized extreme learning machine(RELM) and sequential regularized extreme learning machine(SRELM). The prediction accuracy for SA-SRELM is almost six times and twice as high as RELM and SRELM，respectively.

Key words: Regularized extreme learning machine；Small sample；Time series prediction；Dynamic adaptability；Generalization performance

摘要： 为提高初始小样本情况下时间序列在线预测的精度，提出了一种结构自适应序贯正则极端学习机(SA-SRELM)。该方法在在线序贯学习阶段，针对不同训练样本规模选择不同的递推方式对输出权值进行更新；同时，在训练样本达到一定规模后，为提高预测模型对系统的动态适应性，在加入新样本的同时对旧样本进行剔除，完成预测模型的训练。利用3种混沌时间序列预测实例对所提方法的有效性进行了验证。最后，将所提方法用于航空发动机排气温度预测中，结果表明该方法相对正则极端学习机(RELM)和序贯正则极端学习机(SRELM)方法具有更好的泛化性能，预测精度分别是二者的约6倍和2倍。

关键词: 正则极端学习机；小样本；时间序列预测；动态适应性；泛化能力