Thesis title: Decomposition-based Forecasting for Shipboard Electric Loads
Variational Mode Decomposition (VMD) was presented in 2014 as a general tool for function decomposition that subdivides functions into band-limited components. Although it was not specifically designed for forecasting, researchers and practitioners have made significant efforts to employ VMD as a preprocessing step for machine-learning-based predictors. In this viewpoint, the original signal is decomposed into band-limited subsignals, a machine learning-based system predicts the future behavior of each subsignal, and the components' predictions are recombined to generate the forecast of the original signal. Initially proposed for univariate cases, VMD was later extended to multivariate contexts. While this step expanded the application of VMD to forecasting with multivariate input, its usage as a preprocessing step for a predictor is less straightforward than the univariate case, and extra care needs to be employed to properly deal with the added complexity. The present work critically reviews the theory and implementation of VMD as a preprocessing step, analyzing these aspects in relation to forecasting. Furthermore, established facts on the nature of machine learning-based predictors are discussed to better refine potential new research avenues, especially in the building of practical tools. Finally, this work introduces new algorithmic concepts with broad applicability for enhancing forecasting systems.