Warm-Start Meta-Ensembles for Forecasting Energy Consumption in Service Buildings

Abstract

Energy Management Systems are equipments that normally perform the individual supervision of power controllable loads. With the objective of reducing energy costs, those management decisions result from algorithms that select how the different working periods of equipment should be combined, taking into account the usage of the locally generated renewable energy, electricity tariffs etc., while complying with the restrictions imposed by users and electric circuits. Forecasting energy usage, as described in this paper, allows to optimize the management being a major asset.This paper proposes and compares three new meta-methods for forecasts associated to real-valued time series, applied to the buildings energy consumption case, namely: a meta-method which uses a single regressor (called Sliding Regressor – SR), an ensemble of regressors with no memory of previous fittings (called Bagging Sliding Regressor – BSR), and a warm-start bagging meta-method (called Warm-start Bagging Sliding Regressor – WsBSR). The novelty of this framework is combination of the meta-methods, warm-start ensembles and time series in a forecast framework for energy consumption in buildings. Experimental tests done over data from an hotel show that, the best accuracy is obtained using the second method, though the last one has comparable results with less computational requirements.