Deep Learning Algorithms Based On ELM And Their Applications In Time Series Prediction

Extreme Learning Manchine(ELM),as a novel Single-hidden Layer Feedforward neural Networks(SLFNs),has proved to be well suited to different kinds of classification and regression problems.However,when confronted with many complex tasks,the shallow architecture ELM still has some limitation owing to failing to seek deep representation of raw data.Recent years,deep ELM models like Hierarchical ELM(H-ELM),deep representation learning via ELM(Dr-ELM)have been proposed to be applied in multiple application in machine learning.In this paper,a novel double deep ELMs heterogeneous ensemble system(DD-ELMs-ES)is proposed to focus on the problem of time series forecasting.In DD-ELMs-ES,besides H-ELM and Dr-ELM are utilized as the basic models,a novel Constarined H-ELM(CH-ELM)is presented and serves as another basic model as well.CH-ELM intends to constrain the hidden neurons' input connection weights,so that the bettern feature mapping can be obtained.What's more,a self-adaptive ReTSP-Trend pruning technique is proposed to implement ensemble pruning in DD-ELMs-ES.In DD-ELMs-ES,the deep models guarantee the useful features can be extracted from original data while ensemble scheme improves the robustness of whole system.The experimental resultes demonstrate that compared to basic models and other state-of-art algorithms,DD-ELMs-ES is capable of achieving better generalization performance.Then,in this work,a new deep ELM models called Cascading Deep Architecture based on Weak Results(CDA-WR)is also proposed to tackle with time series prediction of financial stock.CDA-WR changes the traditional time series prediction model.It is divided into three stages.In the first stage,relative weak results can be obtained through constructed weak predictor.In the second stage,reorganize the weak result with original sample to construct new sample,and then employ the Stacked Auto encoders(SAEs)to extract useful features from new samples.In the third stages,OS-ELM will be utilized to complete the final prediction.The results on five financial stock time series datasets show that CDA-WR has better predictive performance than other ELM variants.