jismart2023008: Physics-Informed Neural Networks for Building Load Prediction: A Review

Abstract

Building load prediction plays a key role in achieving energy efficiency and building comfort. Traditional data-driven methods have achieved significant advancements in building load prediction, but challenges remain regarding the generalization capability of conventional neural networks. To address these issues, the research community has recently shown interest in physics-informed data-driven methods, which have introduced a novel predictive approach known as Physics-Informed Neural Networks (PINN). This review aims to provide a comprehensive summary of the application of PINN in building load prediction. Firstly, the fundamental principles of PINN and its advantages over traditional artificial neural networks are briefly introduced. Subsequently, several representative application cases are presented, highlighting the modeling approaches of PINN, including the selection of base networks, design of physics-based loss functions, and the training process of the networks. Finally, the advantages and challenges of PINN are summarized, along with prospects for future research. It is expected that this review will stimulate further research and development in the field of building load prediction, providing valuable insights and guidance for the realization of intelligent and efficient building energy management.