Data-centric or algorithm-centric: Exploiting the performance of transfer learning for improving building energy predictions in data-scarce context. (1st February 2022)

• Record Type: Journal Article
• Authors: Fan, Cheng; Lei, Yutian; Sun, Yongjun; Piscitelli, Marco Savino; Chiosa, Roberto; Capozzoli, Alfonso
• Is Part Of: Energy. Volume 240(2022)
• Publication Date: 2022-02-01
• Subjects: Transfer learning -- Building energy predictions -- Data scarcity -- Data science -- Predictive modeling; Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2021.122775
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: Data-driven methods have gained increasing popularity due to their high-convenience and high-accuracy in practice. Considering the wide discrepancies in data availability across different buildings, transfer learning can be applied to improve the feasibility and robustness of data-driven solutions for individual buildings. In principle, the performance of transfer learning can be enhanced from two perspectives, i.e., the algorithm-centric and data-centric perspectives. The algorithm-centric perspective highlights the adoption of advanced learning algorithm, while the data-centric perspective emphasizes the preparation of proper data for cross-building sharing. At present, there is a lack of studies to systematically compare the performance of the above-mentioned strategies for building energy predictions in a broad range of building types. This study, therefore, investigates the actual performance of transfer learning in data-scarce context, i.e., target buildings have insufficient/extremely limited operational data for model calibrations and domain adaptations. Various transfer learning methods, using different learning algorithms and source data utilization schemes, have been developed and applied for performance comparisons. Comprehensive data experiments have been designed using 600 actual buildings to draw statistically significant conclusions. The results are helpful for quantifying the behavioral patterns of transfer learning, and providing practical guidelines to develop cost-effective data-driven solutions for building energy predictions. Highlights: Transfer learning is used for building energy predictions in data-scarce context. Data- and algorithm-centric perspectives are integrated for predictive modeling. Data experiments have been conducted using actual data from 600 buildings. The performance of various learning strategies has been quantitatively assessed. Insights on cost-effective transfer learning strategies are obtained for practice.