Answers are grounded in this paper and cite their sources.
Building energy consumption prediction plays a critical role in intelligent energy management, demand response, and low-carbon operation of smart buildings. This paper proposes a novel physics-guided multi-scale temporal Transformer (PMSTFormer) for building energy consumption forecasting. First, a dynamic key variable selection (DKVS) module is developed to adaptively estimate the contribution of heterogeneous input variables at each prediction horizon. Second, a multi-scale temporal fusion encoder (MTFE) is designed to capture temporal dependencies at multiple resolutions. Finally, a physics-guided thermal consistency learning (PTCL) strategy is proposed by embedding physical constraints into the optimization process. PTCL is implemented as a set of differentiable soft constraints that jointly optimize data fidelity and simplified thermal-response consistency. It does not require an explicit thermal model or solver in the loop, and is therefore complementary to methods that enforce physical consistency by construction. Extensive experiments are conducted on publicly available benchmark datasets Building Data Genome 2. Comparative studies demonstrate that the proposed PMSTFormer consistently outperforms state-of-the-art forecasting models. Furthermore, comprehensive ablation studies verify the effectiveness of each proposed module. The developed methodology not only improves forecasting accuracy but also offers a practical and reliable decision-support tool for intelligent building energy management systems.
No takes yet. Share an insight, caveat, or question.
Dai et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: