Feature Extraction for Day-ahead Electricity-Load Forecasting in Residential Buildings

In the context of electricity demand response, an important task is to generate accurate forecasts of energy loads for groups of households as well as individual consumers. We consider the problem of short-term (one-day-ahead) forecasting of the electricity consumption load of a residential building. In order to generate such forecasts, historical energy consumption data are used, presented in the form of a time series with a fixed time step. In this paper, we first review existing (one-day-ahead) forecasting methodologies including: a) naive persistence models, b) autoregressive-based models (e.g., AR and SARIMA), c) triple exponential smoothing (Holt-Winters) model, and d) combinations of naive persistence and auto-regressive-based models (PAR). We then introduce a novel forecasting methodology, namely seasonal persistence-based regressive model (SPR) that optimally selects between lower- and higher-frequency persistence and temporal dependencies that are specific to the residential electricity load profiles. Given that the proposed forecasting method equivalently translates into a regression optimization problem, recursive-least-squares is utilized to train the model in a computationally efficient manner. Finally, we demonstrate through simulations the forecasting accuracy of this method in comparison with the standard forecasting techniques (a)-(d). Copyright (C) 2020 The Authors.