Empirical Mode Decomposition, Extreme Learning Machine and Long Short-Term Memory for Time Series Prediction: A Comparative Study

Abstract

The use of models that combine empirical mode decomposition (EMD) and artificial neural networks (ANN) to time series prediction has been attracted much research interest in several areas of great relevance. However, the way the two methods are combined can vary. Thus, a comparison between different combinations of models is presented in this work. The first objective is to verify if the use of EMD improves the prediction results. The second objective is to compare whether it is better to group the intrinsic mode function (IMFs) and then perform the prediction, or predict each IMF separately and then aggregate the results. The methods were tested for six different time series and the results show that EMD improves the prediction for the most of the investigated series, especially if one predictor is used for each IMF separately.