A study of deep learning-based multi-horizon building energy forecasting

Artikel i vetenskaplig tidskrift, 2024

Building energy forecasting facilitates optimizing daily operation scheduling and long-term energy planning. Many studies have demonstrated the potential of data-driven approaches in producing point forecasts of energy use. Despite this, little work has been undertaken to understand uncertainty in energy forecasts. However, many decision-making scenarios require information from a full conditional distribution of forecasts. In addition, recent advances in deep learning have not been fully exploited for building energy forecasting. Motivated by these research gaps, this study contributes in two aspects. First, this study has adapted and applied state-of-the-art deep learning architectures to address the problem of multi-horizon building energy forecasting. Eight different methods, including seven deep learning-based ones, were investigated to develop models to perform both point and probabilistic forecasts. Second, a comprehensive case study was conducted in two public historic buildings with different operating modes, namely the City Museum and the City Theatre, in Norrköping, Sweden. The performance of the developed models was evaluated, and the predictability of different scenarios of energy consumption was studied. The results show that incorporating future information on exogenous factors that determine energy use is critical for making accurate multi-horizon predictions. Furthermore, changes in the operating mode and activities held in a building bring more uncertainty in energy use and deteriorate the prediction accuracy of models. The temporal fusion transformer (TFT) model exhibited strong competitiveness in performing both point and probabilistic forecasts. As assessed by the coefficient of variance of the root mean square error (CV-RMSE), the TFT model outperformed other models in making point forecasts of both types of energy use of the City Museum (CV-RMSE 29.7% for electricity consumption and CV-RMSE 8.7% for heating load). When making probabilistic predictions, the TFT model performed best to capture the central tendency and upper distribution of heating load of the City Museum as well as both types of energy use of the City Theatre. The predictive models developed in this study can be integrated into digital twin models of buildings to discover areas where energy use can be reduced, optimize building operations, and improve overall sustainability and efficiency.