Short-term electricity consumption forecasting with NARX, LSTM, and SVR for a single building: small data set approach

ABSTRACT

Nowadays, there is an undoubted change of trend toward a decentralized and decarbonized electric grid, where the electric generation based on local resources will take on special relevance. In this context, the encouragement of collective self-consumption (CSC) becomes one of the key issues. One of the aspects that will contribute to this aim is the development of power consumption-forecasting tools. This article proposes the comparison of three models to perform a day ahead consumption forecasting of ESTIA 2 building: nonlinear autoregressive neural network with exogenous inputs (NARX), long-short-term memory cell (LSTM) and support vector regression (SVR). First, the model structure has been designed by selecting the suitable-input combination and the optimal time window (TW) for the three models. Then, parameters of each model have been adjusted to achieve the most accurate prediction. After forecasting separately winter and summer seasons, experiments reveal that the proposed NARX neural network is the one that predicts with the highest accuracy in both winter and summer months, obtaining a mean absolute percentage error (MAPE) of 14,1% and 12%, respectively. Likewise, regardless of the model, better results have been obtained in summer predictions, which is closely related to the dependence of the building’s consumption on the heating, ventilation, and air conditioning (HVAC) system.

KEYWORDS:

• Collective self-consumption
• artificial neural networks
• nonlinear autoregressive exogenous
• Long-Short-Term memory cell
• support vector regression

Acknowledgments

This research study carried out in the frame of EKATE project has been supported by FEDER Interreg POCTEFA program. In addition, Meteo France has provided the meteorological historical data used during the whole process.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/15567036.2022.2104410