Master's Thesis : Ensemble methods applied to electricity demand forecast

Abstract

This work studies performances of an ensemble method approach based on stacking to forecast energy demand in order to get day-ahead nominations for industrial consumer. This ensemble method's study is conducted with four different models as meta learner, with five learning algorithms as base models. Model training and evaluation are done using historical data composed of one full year meter measurements. This study shows that despite being widely used, ensemble methods need refinement to claim better results than a naive algorithm alone on the case study considered in this work. Summing seasonal naive predictions is the most naive setup evaluated in this work and it shows better performances than almost all other tested methods. However, the best one seems to be the combination of WaveNet as base learner as well as meta learner. Overall, this WaveNet aggregator achieves good performances whatever the base learner we chose to use and except in the seasonal naive case it outperforms any base/meta model combination we tested in this work.