Elastic weight consolidation-based adaptive neural networks for dynamic building energy load prediction modeling: Fid-Bau Portal

Elastic weight consolidation-based adaptive neural networks for dynamic building energy load prediction modeling

Zhou, Yangze / Tian, Xiangning / Zhang, Chaobo / Zhao, Yang / Li, Tingting

Highlights • An elastic weight consolidation-based method is proposed for building load prediction. • The proposed method has higher accuracy than conventional prediction methods. • The proposed method costs low computational and storage source. • The proposed method has higher model stability than sliding window fining-tuning.

Abstract Data-driven building energy load prediction models should be updated dynamically to adapt to building performance degradation and changes of outdoor environment. Conventional dynamic modeling methods usually have to either increase the costs of model training and data storage for obtaining high accuracy, or decrease accuracy for reducing the costs of model training and data storage. Therefore, this paper presented an elastic weight consolidation-based sliding window fine-tuning method for dynamic building energy load prediction modeling with high accuracy, low computational costs, and low data storage costs. It adopted elastic weight consolidation to make a model be able to remember historical knowledge and learn new knowledge without storing historical data. Three common dynamic modeling methods (sliding window retraining, accumulative retraining, and sliding window fine-tuning) and a static modeling method were adopted for performance comparisons with the proposed method using the five-year operational data of a public building. According to the results, the mean absolute error of the proposed method decreased by 66.58%, 9.06%, and 8.70% on average compared with sliding window retraining, sliding window fine-tuning, and static modeling, respectively. It was further demonstrated that the proposed method had better stability than sliding window fine-tuning. Moreover, the proposed method showed similar accuracy to accumulative retraining, while it had lower costs of model training and data storage than accumulative retraining.