Physics-Informed Neural Network for Building Energy Demand Prediction

Abstract

The optimization of energy demand in the built environment has become a critical priority in the global drive for sustainability. Despite building energy simulation tools can provide valuable performance insights, their computational demands and complexity often impede practical implementation, especially for small-scale operational.

This study presents the development of a novel surrogate modeling framework designed to predict residential building energy demand while maintaining computational efficiency.

The methodology starts with data generation through EnergyPlus simulations across diverse building configurations, while, for modeling, a physics-informed neural network (PINN) is employed, which integrates domain knowledge from an additional simplified physical model with data-driven predictions. Differently from traditional machine learning approaches, the PINN framework helps to promote thermodynamic consistency by embedding fundamental heat transfer equations and energy conservation principles directly into the learning process, thereby fostering physically meaningful predictions even in scenarios not represented in the training data. The simplified physical model implemented within the loss function is a resistance-capacity model, which captures essential thermal dynamics through a compact representation of heat storage and transfer.

The developed framework tries to effectively fill the gap between high-fidelity simulation tools and practical implementation requirements by enabling accurate and physical-informed evaluation, while accounting for complex system interactions. The PINN’s ability to align with physical constraints helps to ensure that the predictions remain reliable even under uncertain or previously unseen conditions, a crucial advantage for real-world applications. Finally, the additional simplified model acts as a physics-informed regularization that is proven to be more effective than common regularization methods in generalizing the model.