?
Feature and model selection for day-ahead electricity-load forecasting in residential buildings
The need for accurate balancing in electricity markets and a larger integration of renewable sources ofelectricity require accurate forecasts of electricity loads in residential buildings. In this paper, we considerthe problem of short-term (one-day ahead) forecasting of the electricity-load consumption in residentialbuildings. In order to generate such forecasts, historical electricity consumption data are used, presentedin the form of a time series with a fixed time step. Initially, we review standard forecasting methodologiesincluding naive persistence models, auto-regressive-based models (e.g., AR and SARIMA), and the tripleexponential smoothing Holt-Winters (HW) model. We then introduce three forecasting models, namelyi) the Persistence-based Auto-regressive (PAR) model, ii) the Seasonal Persistence-based Regressive (SPR)model, and iii) the Seasonal Persistence-based Neural Network (SPNN) model. Given that the accuracy ofa forecasting model may vary during the year, and the fact that models may differ with respect to theirtraining times, we also investigate different variations of ensemble models (i.e., mixtures of the previ-ously considered models) and adaptive model switching strategies. Finally, we demonstrate through sim-ulations the forecasting accuracy of all considered forecasting models validated on real-world datagenerated from four residential buildings. Through an extensive series of evaluation tests, it is shown thatthe proposed SPR forecasting model can attain approximately a 7% forecast error reduction over standardtechniques (e.g., SARIMA and HW). Furthermore, when models have not been sufficiently trained, ensem-ble models based on a weighted average forecaster can provide approximately a further 4% forecast errorreduction.