Deep forecasting of electricity consumption and production

Abstract

Buildings stock represented 41% of the final European energy consumption in 2016. It is, therefore, critical to improve their energy efficiency. Forecasting the electricity consumption of buildings would help users in saving energy, as it can support energy efficiency and reveal building system faults. The focus of this thesis is to apply machine learning for forecasting the electricity consumption and production. The data includes electricity consumption, solar power generation and weather times series that were previously collected hourly in the campus of Lappeenranta-Lahti University of Technology. It is analysed over a 36-hour prediction horizon with a state-of-the-art model, called the Interpretable Multi-Variable Long Short-Term Memory Neural Network, that was selected among the best forecasting models. An accurate forecast with a MAAPE of 2.45% is achieved for the electricity consumption data. Forecasting the solar power generation data was less successful, reaching a MAAPE of 55.86% at best.