Open AccessComparative study of continuous hourly energy consumption forecasting strategies with small data sets to support demand management decisions in buildings
Author(s) -
MarianoHernández D.,
HernándezCallejo L.,
Solís M.,
ZoritaLamadrid A.,
DuquePérez O.,
GonzalezMorales L.,
AlonsoGómez V.,
JaramilloDuque A.,
Santos García F.
Publication year - 2022
Publication title -
energy science and engineering
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.638
H-Index - 29
ISSN - 2050-0505
DOI - 10.1002/ese3.1298
Subject(s) - mean absolute percentage error , consumption (sociology) , computer science , electricity , energy consumption , decision tree , demand forecasting , time horizon , random forest , artificial neural network , operations research , data mining , artificial intelligence , mathematical optimization , engineering , mathematics , sociology , electrical engineering , social science
Buildings are one of the largest consumers of electrical energy, making it important to develop different strategies to help to reduce electricity consumption. Building energy consumption forecasting strategies are widely used to support demand management decisions, but these strategies require large data sets to achieve an accurate electric consumption forecast, so they are not commonly used for buildings with a short history of record keeping. Based on this, the objective of this study is to determine, through continuous hourly electricity consumption forecasting strategies, the amount of data needed to achieve an accurate forecast. The proposed forecasting strategies were evaluated with Random Forest, eXtreme Gradient Boost, Convolutional Neural Network, and Temporal Convolutional Network algorithms using 4 years of electricity consumption data from two buildings located on the campus of the University of Valladolid. For performance evaluation, two scenarios were proposed for each of the proposed forecasting strategies. The results showed that for forecasting horizons of 1 week, it was possible to obtain a mean absolute percentage error (MAPE) below 7% for Building 1 and a MAPE below 10% for Building 2 with 6 months of data, while for a forecast horizon of 1 month, it was possible to obtain a MAPE below 10% for Building 1 and below 11% for Building 2 with 10 months of data. However, if the distribution of the data captured in the buildings does not undergo sudden changes, the decision tree algorithms obtain better results. However, if there are sudden changes, deep learning algorithms are a better choice.