High-precision energy consumption forecasting for large office building using a signal decomposition-based deep learning approach

Accurate long-term energy consumption forecasting is crucial for efficient energy management in large office buildings. Recent research highlights that deep learning approaches, including RNN, LSTM, and transformer-based models, are at the forefront of promising advancements. They are unified in obtaining more discriminative representations. The challenges lie in the complexity of data influenced by diverse factors such as weather, building characteristics, and occupant behavior, etc., and the need to accurately model the intricate patterns of time-series periodicity and trends. In this paper, we introduce SPAformer, an innovative end-to-end deep learning model adept at unraveling and forecasting the intricate components of energy consumption data. It is motivated by the hypothesis that decomposing energy consumption into detailed functional categories and isolating trends and periodic components can significantly enhance forecasting accuracy. In response, we propose spectra-patch attention (SPA) mechanism, which combines time and frequency signals, to better capture the repeating patterns in lengthy data sequences. We have evaluated our approach on a real-world granular dataset from a large commercial office building and demonstrated SPAformer's superior performance. By achieving a 12% improvement in prediction accuracy over state of the art attention-based models, SPAformer marks a significant stride in energy forecasting. This work contributes to better-informed decision making about energy saving strategies, emphasizing the model's usefulness in the ongoing planning and fine-tuning of building energy systems.