Dual-phase evaluation framework of physics-informed machine learning model for building dynamics modeling and advanced control

• Dual-phase evaluation of physics-informed ML models for building control. • Real-world and simulation-based comparison of PCNN, PCICNN, and ModNN. • Control performance evaluated using MPC and RL in a building testbed. Physics-informed machine learning (PIML) models have emerged as promising approaches for modeling indoor temperature dynamics by combining data-driven techniques with physical principles. However, a systematic comparison of different PIML methods and their practical control performance is still lacking. This study presents the first comprehensive evaluation of three representative PIML models: Physics-Consistent Neural Network (PCNN), Physics-Consistent Input Convex Neural Network (PCICNN), and Modularized Neural Network (ModNN). We develop a dual-phase evaluation framework that includes controlled 1-month simulation-based comparisons and 1.5-month real-world implementation in an office building testbed under cooling-dominant test conditions, assessing prediction accuracy, physical consistency, training efficiency, and control performance. Simulation results indicate that ModNN achieves the highest prediction accuracy (3.02% MAPE) and the strongest physical consistency, effectively capturing thermal lagging effects. PCICNN demonstrates stable performance, enabling smooth control operation, while PCNN shows fundamental limitations such as poor accuracy (5.03% MAPE) and failure to learn solar radiation effects. The real-world implementation reveals significant energy savings (35.9%-88.73%) across various advanced control strategies, though energy savings alone are insufficient for comprehensive evaluation. Although PCNN records the highest apparent savings (88.73%), it displays problematic control behaviors like cycling signals and excessive temperature violations. PCICNN offers the most reliable operation with smooth signals, making it ideal for model predictive control applications. The study confirms that physical consistency directly influences control effectiveness and operational reliability. Our gradient analysis highlights how modeling deficiencies translate into control issues, providing valuable guidance for model selection. This work advances the field by emphasizing control quality metrics alongside accuracy in evaluation.