Physics-Informed Modularized Neural Networks for Building Dynamic Modeling: A Smart Buildings Hackathon Case Study

Building Energy Modeling (BEM) plays a critical role in energy analysis, retrofit design, and control optimization for buildings.However, traditional physics-based BEM approaches are laborintensive and require extensive metadata and calibration, while purely data-driven models often struggle to generalize under unseen conditions.To address these limitations, Physics-Informed Machine Learning (PIML) integrates prior knowledge into datadriven models, combining the strengths of both approaches.To demonstrate the effectiveness of PIML, in this study, we applied our previously developed Physics-Informed Modularized Neural Network (PI-ModNN) to a real-world multizone HVAC dataset from the Smart Buildings Challenge Hackathon, under a one-day development constraint.Results show that PI-ModNN outperformed a classical LSTM baseline in both generalization and physical consistency, achieving a six-month average MAE of 5.71 C compared to 7.33 C for LSTM.However, performance is still below our expectations, which might be because of simplifying the multizone problem to a single-zone formulation, the absence of occupancy and solar radiation data, lack of data preprocessing, and the computational complexity of modeling inter-zone heat transfer.This study highlights both the potential and limitations of rapid-deployment PIML models in complex building environments and suggests key areas for future improvement.