A Physics-Informed Graph Neural Network Approach for Multi-Time-Step Prediction of Indoor Temperature Fields

Building energy consumption has increased significantly, with public buildings, being major contributors to carbon emissions. Optimizing air conditioning efficiency is crucial for energy conservation, and accurate indoor temperature prediction is key to reducing energy use. While computational fluid dynamics (CFD) has been used to model airflow and temperature fields, its high computational costs and reliance on grid quality limit its practical efficiency. Recent advances in deep learning, particularly neural network-based methods, have shown promise in solving fluid dynamics problems more efficiently. However, existing models often fail to incorporate the underlying physical principles, affecting predictive accuracy in complex scenarios. To address this, we propose a novel framework combining CFD with physics-informed graph neural network (GNN). This framework leverages GNN to model fluid dynamics in air-conditioned environments while embedding physics-based loss functions to ensure the model respects fluid dynamics principles. The proposed model not only achieves accurate temperature field predictions but also significantly reduces computational time. This work aims to optimize air conditioning systems in public buildings, contributing to energy efficiency and carbon emission reduction.