Time Series Forecasting with an EMD-LSSVM-PSO Ensemble Adaptive Learning Paradigm

In this study, an empirical mode decomposition (EMD) based least square support vector machine (LSSVM) ensemble adaptive learning paradigm is proposed for time series forecasting. For this purpose, the original time series are first decomposed into several intrinsic mode functions (IMFs) and one residual component. Then phase space reconstruction (PSR) is done in each component, where the samples are put into LSSVM for training. Particle swarm optimization algorithm (PSO) is used to achieve the adaptive optimization of forecasting models in different components, which makes LSSVMs better describe the signal characteristics in different scales, thus greatly improves the efficiency and accuracy of the learning and training; Finally, the forecasting values of the original series are obtained through the reconstruction of the forecasting values in each component. For illustration and verification, the spare parts cost series for a vessel is used to test the effectiveness of the proposed EMD-LSSVM-PSO ensemble adaptive learning methodology. Empirical results obtained demonstrate the attractiveness of the proposed method.