ScholarWorks@성균관대학교

초록

This study analyzed integration and transferability of Physics-Informed Neural Network (PINN) models for occupant-centric ventilation control in occupied buildings. Balancing indoor air quality and energy efficiency remains a major challenge, as conventional control strategies often fail to adapt to dynamic occupant behaviors and environmental variations. Model predictive control improves performance through predictive optimization but still depends on model accuracy. Recently, PINNs, which embed physical laws into neural networks, have emerged as a promising approach for interpretable and adaptive control. However, most studies remain limited to single-zone modeling, whereas real buildings are subdivided into multiple zones with varying occupants and behavior dynamics. Evaluating model integration across similar rooms and transferability between zones is essential for scalable occupant-adaptive control. In this work, two physically similar university classrooms were monitored for 12 weeks to collect indoor and outdoor CO₂, temperature, and occupant activity data. Three types of PINN models, room-specific, integrated, and transfer were developed and evaluated for predictive robustness and control applicability. While all room-specific models satisfied the ASTM D5157 criteria, the integrated model showed a slight systematic underprediction (Fractional Bias = −0.27), and the transfer model exhibited significantly higher prediction errors (mean daily Normalized Mean Squared Error = 0.37) compared with the room-specific model (0.17). SHAP-based interpretation revealed how each model internalized physical and behavioral dynamics differently, reflecting a trade-off between generalization and control fidelity. These findings highlight the need for zone-customized learning strategies addressing spatial and behavioral variations, providing empirical evidence for adaptive and interpretable ventilation control.

키워드