A Hybrid Approach to Building Thermal Modeling Using Physics-Based Machine Learning

Buildings account for approximately 30 % of primary energy consumption, mainly due to Heating, Ventilation, and Air Conditioning (HVAC) systems. Reactive controllers can be used to manage these systems optimally, however, their performance depends strongly on the accuracy of building thermal models. In this study, a hybrid physics-informed machine learning (PIML) approach is proposed to improve indoor temperature prediction by combining a data-driven model with a simplified resistance-capacitance (RC) network. The framework integrates a 4R3C thermal model with a multi-layer perceptron (MLP). The MLP is used both to identify thermal parameters and to predict indoor temperature and heating/cooling loads. The hybrid PIML model is applied to a building with limited information on HVAC operation and system characteristics. Results show that the hybrid PIML model captures thermal dynamics more accurately than the physics-only model, achieving a mean absolute error (MAE) of 0.454° C and a root mean square error (RMSE) of 0.547° C, compared with 0.851° C and 0.974° C for the physicsonly model. The proposed framework is lightweight, reproducible, and compatible with model-based control applications.