Thermal-GEMs: Generalized Models for Building Thermal Dynamics

A team of researchers has introduced Thermal-GEMs, a framework for creating generalized, data-driven models to predict building thermal dynamics. This is crucial for enabling energy-efficient building operations through advanced control systems and fault detection. The research tackles a major industry hurdle: obtaining accurate models traditionally requires months or years of measurement data from a specific target building. The study's core innovation is a first-of-its-kind, comprehensive assessment comparing multi-source transfer learning (TL) architectures—which are pretrained on data from multiple other buildings—against emerging time series foundation models (TSFMs) that are trained on vast, diverse datasets.

The findings are highly practical for engineers and building operators. The team demonstrated that multi-source TL models are highly effective in real-world applications, achieving up to 63% lower forecasting errors than models trained on data from just a single source building. However, the research also reveals a critical data threshold: for a multi-source TL model to consistently outperform a general-purpose TSFM, it needs to be pretrained with data from 16 to 32 source buildings, collected over at least one year. This provides a clear, data-driven roadmap for selecting the most accurate modeling strategy based on the amount of source building data available, directly impacting the feasibility and cost of deploying AI for smart building management.