Large-scale comparison and demonstration of continual learning for adaptive data-driven building energy prediction

Data-driven models have been increasingly employed in smart building energy management. To avoid performance degradation over time, data-driven models need to be continually updated to adapt to the changes in building operations. However, several critical issues in the model update process raised wide concerns, especially the concept drift and catastrophic forgetting issues. The concept drift issue happens when the statistical properties of target variable change over time in unforeseen ways. The catastrophic forgetting issue refers to the process that the previously learnt knowledge or patterns may be diluted and eventually lost in model update. Although a few model update methods were proposed, there is a lack of comprehensive comparison of the methods for adaptive data-driven building energy prediction. This paper conducted a comprehensive investigation on the performance of three conventional model update methods and five emerging continual learning methods using 2-year data of 100 buildings extracted from an open-source dataset. The results show that continual learning methods are more effective in ensuring long-term accuracy while cutting down on computation time and data storage expenses. The CV-RMSE of Elastic weight consolidation and Gradient episodic memory decreased by around 14% and 8% on average compared with static model and accumulative learning. The comparison results are valuable to the development of adaptive data-driven building energy prediction models which are more reliable over time and robust against changing operation conditions, thus more practically applicable in smart building energy management.