Researchers have developed a novel machine learning framework called the Physics-guided Multi-scale Temporal Transformer (PMSTFormer) to advance building energy consumption forecasting. Building energy consumption prediction plays a critical role in intelligent energy management, demand response, and low-carbon operation of smart buildings.
The PMSTFormer architecture is built upon three primary technical components designed to handle the complexities of energy usage data. Central to the system is a dynamic key variable selection (DKVS) module, which is responsible for adaptively estimating the contribution of various heterogeneous input variables at each specific prediction horizon.
Beyond the selection module, the researchers implemented a multi-scale temporal fusion encoder (MTFE). This component is engineered to capture complex temporal dependencies across multiple resolutions.
Finally, a physics-guided thermal consistency learning (PTCL) strategy is proposed by embedding physical constraints into the optimization process. This is implemented as a set of differentiable soft constraints that work to 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.
To validate the efficacy of the PMSTFormer, the researchers conducted extensive experiments using the publicly available Building Data Genome 2 benchmark dataset. This rigorous testing phase allowed for a direct comparison between the proposed model and current state-of-the-art forecasting methodologies. The results consistently demonstrated that the PMSTFormer outperforms these existing models in accuracy. Furthermore, the team performed comprehensive ablation studies to verify the individual contribution of each proposed module, confirming that the DKVS, MTFE, and PTCL components all play a necessary role in the overall predictive success of the framework.
The developed methodology not only improves forecasting accuracy but also offers a practical and reliable decision-support tool for intelligent building energy management systems.
Subject of Research: Deep learning-based energy forecasting for intelligent building management systems.
Article Title: Physics-guided multi-scale temporal transformer with dynamic key variable selection for building energy consumption forecasting
Article References: Dai, F., Zeng, Y., Yuan, C., Huang, L., & Mei, M. (2026). Physics-guided Multi-scale Temporal Transformer with Dynamic Key Variable Selection for Building Energy Consumption Forecasting. Frontiers in Environmental Science, 14, 1971486. https://doi.org/10.3389/fenvs.2026.1971486
Image Credits: AI Generated
DOI: 10.3389/fenvs.2026.1971486
Keywords: PMSTFormer, energy consumption forecasting, Transformer models, physics-guided machine learning, smart buildings, temporal dependency, thermal consistency learning, dynamic key variable selection, Building Data Genome 2, demand response, low-carbon operation, environmental systems engineering, predictive analytics, multi-scale temporal fusion, data-driven modeling
Cite Scienmag News
Katie Riggs. (November 18, 2026). Physics-Guided Transformer Improves Accuracy in Building Energy Consumption Forecasting. Scienmag. https://scienmag.com/physics-guided-transformer-improves-accuracy-in-building-energy-consumption-forecasting/
Katie Riggs. "Physics-Guided Transformer Improves Accuracy in Building Energy Consumption Forecasting." Scienmag, 18 November 2026, https://scienmag.com/physics-guided-transformer-improves-accuracy-in-building-energy-consumption-forecasting/. Accessed 21 November 2026.
Katie Riggs. "Physics-Guided Transformer Improves Accuracy in Building Energy Consumption Forecasting." Scienmag. November 18, 2026. https://scienmag.com/physics-guided-transformer-improves-accuracy-in-building-energy-consumption-forecasting/








