Sensitivity analysis of physical regularization in physics-informed neural networks (PINNs) of building thermal modeling

Abstract

Building energy consumption constitutes over 40 % of the total primary energy consumption, and buildings play an essential role in carbon-neutrality and energy transition. To unlock their potential, an accurate control-oriented thermal model is crucial for energy management and efficiency. However, developing such a model remains a daunting task due to the complexity and time-consuming of physics-informed models, as well as the generalization and interpretability problems of purely data-driven models. Thus, we propose a Physics-informed Neural Network (PINNs) model based on a simplified Resistance-Capacitance (2R2C) thermal model and a fully connected neural network architecture. Three factors—physical regularization conditions, data lengths, and prediction durations—are proposed and investigated. First, the parameters in the 2R2C thermal model are estimated based on the historical data of an actual case building. Second, the 2R2C model is integrated into a fully connected neural network by reconstructing the constrained loss function. Finally, the temperature and load demand prediction performance of the proposed PINNs model is analyzed using physics-informed conditional regularization with respect to different data lengths and prediction durations. The results show that the PINNs model (temperature CVRMSE of 4.31 % and RMSE of 0.94 ℃) outperforms the pure neural networks (NNs) (temperature CVRMSE of 5.91 % and RMSE of 1.27 ℃) and physic-based grey-box 2R2C model (temperature CVRMSE of 8.76 % and RMSE of 1.91 ℃) in multi-step predictions in buildings. The results also emphasize the importance of introducing physical knowledge in conventional data-driven models. The proposed PINNs model can be conveniently and efficiently deployed in the energy management system of buildings, owing to its simplicity and generalization.