Chul-Hong Park, Seongkwon Cho, Tae Yong Song, Seon-Young Heo, Junu Lee & Cheol Soo Park

Received 13 Jan 2025, Accepted 28 Jul 2025, Published online: 05 Aug 2025

https://doi.org/10.1080/19401493.2025.2542352

Abstract

Achieving robust extrapolation and generalizability of physics-based models and the convenience of data-driven models have been one of the major goals of building energy modelling. In this study, a novel physics-embedded hybrid modelling approach that combines the advantages of the two approaches for predicting the thermal behaviour of the building is proposed. It incorporates an RC model and neural networks by utilizing custom-designed layers and a technique called ‘Siamese neural network’. ANN is used to predict various parameters in the RC model, and the embedded 1st-principles-based equations are directly used for predictions, achieving physically consistent results. The model allows flexible structures, simultaneous training of both time-invariant and varying parameters, and multiple-timestep calculation, all with a limited number of measured states. The model was applied to an office area in an existing building, and it was found that the model was good enough to predict room air temperature with high accuracy.

Key Figures