A novel hybrid modeling approach developed by researchers at the Center for Environmental Energy Engineering (CEEE) integrates artificial intelligence with physics-based models to optimize energy performance in HVAC systems. The image features lead author Po-Ching Hsu alongside the outdoor unit of a VRF system used to gather experimental data for model development.
Heating and cooling account for nearly half of all energy consumed in buildings—a share that machine learning could help reduce. However, the effectiveness of machine-learning models depends heavily on the quality and quantity of training data. To overcome this limitation, a research team from the University of Maryland's CEEE has introduced a hybrid modeling framework that merges data-driven techniques with physical principles.
Published in a recent edition of Energy & Buildings, the study presents a hybrid approach tailored to variable refrigerant flow (VRF) systems, aiming to boost energy efficiency while maintaining indoor comfort levels.
The proposed model builds upon earlier machine-learning work by the team, which used HVAC operational data, building conditions, and short-term weather forecasts to enhance system performance. Those earlier models, however, struggled under extreme weather conditions—events that occur infrequently in College Park and were thus underrepresented in the training data.
"Data-driven models perform well with large datasets, but in practice, such data volumes are rarely available," explains Po-Ching Hsu, lead author and a Ph.D. candidate in mechanical engineering and graduate research assistant at CEEE. The paper was co-authored by Research Professor Yunho Hwang, who leads CEEE's Energy Efficiency and Heat Pumps consortium.
"Data-driven models are highly accurate when sufficient data exist—but that's often not the case in real-world applications."
Po-Ching Hsu, CEEE graduate research assistant and mechanical engineering Ph.D. candidate
Physics-based models, by contrast, require less data but demand greater computational resources and modeling effort. "After comparing both approaches," Hsu says, "I considered combining them to develop a model that is both accurate and computationally efficient."
A typical VRF system includes one outdoor unit connected to several indoor units, each serving different thermal zones. For this study, the researchers constructed a virtual VRF system based on empirical data collected from a setup in Glenn L. Martin Hall on campus, comprising one outdoor unit and seven indoor units.
*The hybrid model demonstrated strong predictive capability for indoor-unit capacity and total power consumption—key factors in optimizing energy use. It maintained reliable performance even in data-limited scenarios, surpassing conventional machine-learning models. Predictions aligned closely with actual system measurements, with typical errors of only 5–6%. Moving forward, the team plans to investigate the model's scalability and reliability across different VRF configurations and geographic locations.*
Download the paper: “Hybrid machine learning–physics-based modeling and model predictive control of variable refrigerant flow systems in buildings.”
If you meet any problem in download it, please send email to wind@raccexpo.com for the paper.
Resource: https://ceee.umd.edu
