DOI number:
10.1016/j.jobe.2025.114695
Journal:
Journal of Building Engineering
Key Words:
Building energy consumption forecasting; Cross-building; Deep learning; Transfer learning
Abstract:
Precise forecasting of building energy consumption (BEC) plays a crucial role in enabling intelligent energy management and supporting sustainable urban planning. Nevertheless, achieving high predictive accuracy often depends on access to extensive historical consumption records, which presents significant challenges in collecting such data from both new or renovated buildings. In order to address the challenge of reduced predictive accuracy due to limited training samples, this work proposes a cross-BEC forecasting framework that based on a deep transfer learning approach that combines CNN and Transformer architectures. The framework utilized a CNN-Transformer feature extractor to capture shared spatio-temporal characteristics between source and target buildings. A systematic analysis was performed to evaluate the impact of varying data volumes from source and target buildings on model performance, and multiple transfer strategies were compared to assess their effectiveness. Experimental results demonstrated that optimal transfer performance is achieved when source-target building pairs are constructed with 12 months of data from the source and 2 months from the target. The strategy involving frozen Transformer blocks with fine-tuned CNN layers and fully connected layers exhibited exceptional robustness and generalization, achieving a maximum transfer improvement rate of 0.670. This study offers actionable insights for selecting source and target buildings data sizes in deep transfer learning applications, thereby contributing to the construction sector's efforts to enhance energy efficiency and reduce emissions.
Paper Publications
