PREDICTION OF ELECTRICITY CONSUMPTION IN A PUBLIC BUILDING USING MACHINE LEARNING

Abstract

Given that people spend a significant portion of their time in buildings, it is essential to design indoor environments that promote their health, well-being, and productivity. This could be achieved at the expense of the energy consumed by heating, ventilation, and air conditioning systems, which account for up to 56% of total building energy consumption. To reduce this consumption, strategies such as precooling or preheating could be beneficial, provided there is advanced information on the building's energy performance. In this work, machine learning techniques were applied to address gap filling in data series and predict electricity consumption at the Rector's Office building of the National University of Salta. Meteorological data from own stations, the National Meteorological Service, and satellite data were used. Using neural networks, missing data in the temperature and solar radiation series were filled, and 16 models based on deep learning were trained to predict electricity consumption at different time periods. Performance, prediction horizon, and computational cost were evaluated. The results obtained demonstrate a good predictive capacity of the selected model, with correlation values greater than 0.95 for 15-minute predictions and 0.75 for 6-hour predictions.