Abstract
Heating, Ventilation, and Air-Conditioning (HVAC) systems are significant energy consumers, particularly in critical environments like hospitals. This paper details the real-world deployment of a data-driven Model Predictive Control (MPC) system using Gaussian Processes (GPs) to manage the HVAC of a surgery center in Brazil. The system achieved up to 4.76% energy savings while maintaining strict thermal comfort standards in a complex, multi-zone environment subject to real-world disturbances.
Introduction
Real-world experimental evidence for advanced HVAC control is rare due to non-linear building dynamics and unmeasured disturbances such as occupancy fluctuations and solar radiation. The authors addressed these challenges at Hospital São Julião in Campo Grande, Brazil, developing a controller capable of learning complex dynamics without a perfect physical model, handling uncertainty, and optimizing energy use within strict safety limits.
Methodology
The core approach utilizes Gaussian Process (GP) based Model Predictive Control. Unlike standard neural networks that provide a single prediction, GPs are non-parametric Bayesian models that provide both a mean prediction and a measure of uncertainty (variance). The controller uses these uncertainty estimates to "robustify" the system, maintaining a safety margin based on how uncertain the model is at any given moment.
System Architecture
The system uses a distributed setup where local controllers (National Instruments myRIO) manage individual air-handling units (AHUs), while a central computer runs the GP-MPC optimization using CasADi and the IPOPT solver.
Key Insights and Implementation
- Uncertainty as a Feature: GP variance is integrated directly into the MPC constraints. By adding a term proportional to the standard deviation to the temperature constraint, the controller automatically becomes more conservative when it encounters unfamiliar conditions.
- Domain-Informed Feature Selection: Building geometry was used to prune GP features. For example, only adjacent rooms were allowed to influence each other's temperature models, which kept the computational load manageable.
- Handling Non-Convexity: GP-MPC leads to non-convex optimization problems. The authors addressed this by initializing the solver with a "warm start" trajectory derived from a simple virtual PI controller.
Experimental Results
The system was tested during regular hospital use, facing high solar radiation and varying occupancy. Key findings include:
- Energy Efficiency: The GP-MPC achieved energy savings of 2.29% (mild weather), 3.13% (warm weather), and 4.76% (hot weather) compared to standard PI and ON/OFF controllers.
- Thermal Comfort: The controller successfully kept temperatures below the 21°C limit during experimental runs, even when faced with unexpected equipment faults.
- Computational Feasibility: Median solve times were kept around 18.7 seconds, well within the 10-minute sampling window required for real-time control.
Conclusion
This research provides rare experimental evidence that data-driven MPC can effectively manage complex building dynamics in real-time. While scaling to larger buildings remains a challenge due to the computational complexity of GPs, the results provide a strong foundation for future data-driven climate control in healthcare settings.
Summary updated