ELM-Based Improved Layered Ensemble Architecture for Time Series Forecasting

Abstract

In this paper, an extreme learning machine (ELM)-based improved layered ensemble architecture (EILEA) for time series forecasting is proposed. Compared with multilayer perceptron (MLP)-based layered ensemble architecture (named LEA), our proposed structure has improved in two aspects. First of all, we proposed an inflection point estimation-based density peaks clustering algorithm to replace K-means algorithm used by LEA, which can automatically determine the number of clusters without the influence of the choice of initial cluster centers on the clustering results. Second, EILEA employs ELM as individual learners, taking advantage of its high learning speed to greatly improve the prediction speed. The proposed architecture has been tested on the time series data of neural network (NN)3 competition. In terms of forecasting accuracy and speed, the experimental results have revealed clearly that our proposed structure is better than LEA.