While data center electricity demand is growing globally thanks to expanding technologies like AI and cloud computing, Ján Drgoňa, associate professor in the Whiting School of Engineering’s Department of Civil and Systems Engineering, is working to reduce electricity consumption by optimizing its usage in complex systems.

Ján Drgoňa

With a secondary appointment in the Department of Electrical and Computer Engineering, and a member of the Ralph O’ Connor Sustainable Energy Institute and Data Science and AI Institute, Drgoňa’s research focuses on scientific machine learning for dynamic systems, optimization, and control. Working with collaborators Ján Boldocký and Martin Gulan from the Slovak University of Technology, and the Pacific Northwest National Laboratory’s Cary Faulkner, Elad Michael, and Aaron Tuor, the team published a study testing an AI-enabled control method designed to make data center chiller plants run more efficiently and reliably in real time than current industry-standard, rule-based strategies.

The team’s approach, known as Differential Predictive Control, or DPC, appears in Control Engineering Practice.

“Data centers use a huge amount of electricity and around 40% of that goes into cooling to prevent equipment from overheating,” says Drgoňa. “Even small efficiency improvements matter at the scale of data centers, and depending on the size of the facility, an approximately 10% reduction in cooling energy can translate into more than $1 million in savings and less overall demand on power grids.”

The team’s approach addresses a long-standing problem, which is that complex systems, like chiller plants, require making a multitude of decisions, such as determining how many chillers should be on or off, selecting appropriate temperatures, and determining flow rates, but the decisions must be made in real-time so that systems can operate more efficiently and effectively. The team says that current rule-based control strategies waste energy, and advanced optimization techniques are too expensive and cumbersome to execute.

To bridge the gap, the team developed DPC, an AI methodology combining physics-based modeling with scientific machine learning to identify optimal decisions prior to deployment.

“Rather than re-solving a complex optimization problem every few minutes, we train our physics-based AI agent ahead of time to learn how to make those decisions efficiently,” Drgoňa says. “In simple terms, we simulate how the cooling system behaves, teach the agent what good decisions look like, and then deploy it as a fast, safe, real-time controller.”

DPC can be applied to both continuous decisions, like temperature adjustment, and discrete decisions, like when to turn chillers on and off. The study evaluated DPC in a high-fidelity chiller plant model referred to as a digital twin, which simulates the data center environment and operating conditions. The simulation results indicate that the method is capable of achieving an average of about 12% energy savings when compared with current rule-based controls. In addition to lowering electricity consumption and operating costs, DPC has the potential to extend mechanical equipment life by reducing the wear associated with inefficient cycling.

As the next step in their research, the team plans to use DPC in an operational data center and apply the approach to related problems, such as grid-responsive compute resource scheduling.

The team will also work with Johns Hopkins University’s Yury Dvorkin, associate professor of civil and systems engineering and electrical and computer engineering, to coordinate data center facility control with power grid needs as part of a broader effort to maintain energy reliability for all grid users. This coordination is anticipated to realize greater savings for energy consumers rather than simply adding more energy generation to the power grid.

“Our work points to a new way of controlling complex energy systems,” says Drgoňa. “Instead of solving computationally expensive optimization problems repeatedly, we can learn how to solve them once and then implement that intelligence at scale. This idea extends beyond data centers to buildings, power systems, energy storage, and broader industrial processes.”

This story was written by Danielle McKenna and originally appeared on the JHU Department of Civil and Systems Engineering website.