Day-Ahead Building Load Forecasting with a Small dataset

Abstract

Abstract A new method is presented, to derive an algorithm that provides a forecast of one-day-ahead electricity consumption of a building. The approach aims to obtain high accuracy with a small dataset of 1-2 weeks, motivated by practical situations where the building is new or subject to relatively frequent changes, and/or limited local computation and memory are available. The method introduces a fictitious input signal that captures the prior information on the periodic behavior of building load time series. Moreover, the use of a linear model structure enables the derivation of guaranteed accuracy bounds on the forecast error, which can be used in day-ahead energy scheduling and optimization. Using an experimental dataset with measurements collected from an office building, it is found that the fictitious input can largely improve the prediction accuracy of the model, outperforming linear predictors and scoring a performance similar to that of nonlinear ARX models, such as recurrent neural networks, while retaining the capability to provide guaranteed accuracy bounds.