具有选择与补偿机制的加权集合序贯极端学习机及其应用

针对集成在线序贯极端学习机(EOS-ELM)预测精度不高和动态适应性差的问题,提出一种具有选择与补偿机制的加权集合序贯极端学习机.该加权集合序贯极端学习机在序贯学习过程中,通过对当前预测模型精度的判断决定是否进行递推更新操作,同时为提高预测模型的动态跟踪能力,在加入新样本的同时对旧样本进行剔除;然后,利用EMD对残差序列处理后进行预测,并将初始预测结果与残差预测结果相加得到最终预测模型.通过对上证指数的预测,结果表明所提方法具有更好的泛化性能,预测精度相比EOS-ELM提高了近36.1%.

Weighted ensemble sequential extreme learning machine with selection and compensation and its application

To solve the problem of low accuracy and worse dynamic adaptation of ensemble on-line sequential extreme learning machine (EOS-ELM), a new weighted ensemble sequential extreme learning machine with selection and compensation (WESELM) is proposed. The WESELM excecutes recursive renewing judging by current model prediction accuracy in sequence learning stage, and removes the oldest sample when new sample joined for the sake of improving the dynamic tracking capability; then, residual error series are predicted after preprocessing by empirical mode decomposition (EMD), the original prediction results and the residual error prediction results are added as ultimately prediction model. In the end, WESELM is used to predict Shanghai composite index, and result shows that the proposed method could get better generalization performance, whose prediction accuracy has increased almost 36.1% than EOS-ELM.