Stacked Generalization Concept for Electrical Load Prediction

Short-term power load forecasting has been widely investigated, as it provides crucial and on-demand information for power planning and operation. In literature, different statistical and mathematical methods along with machine learning and data-driven based approaches have been employed and considered for this matter. In terms of the latter approaches, various benchmark models are found in literature for electrical load forecasting, such as support vector regression (SVR) and artificial neural network (ANN). Further steps has been already taken by combining such models in the form of ensemble learning schemes, in particular bagging and AdaBoost ensembles. Different from other methods, this paper investigates the stacked generalization ensemble concept where the electrical load has been analyzed in a time series fashion. The results have been compared with the individual underlying benchmark models, which have shown that the ensemble scheme outperforms the individual models. Furthermore, the introduced scheme is rather robust against error propagation, as the estimated load at a certain future time slot is also utilized to estimate further slots.