Researchers have proposed a relax-and-round mathematical method to help electrical grids better manage the surging power demand from data centers.

Researchers at the Department of Energy’s Oak Ridge National Laboratory (ORNL) have developed a computational method to help electric grid operators and data center developers manage power more efficiently as artificial intelligence drives unprecedented growth in electricity demand. The research, presented at the 2026 IEEE Power & Energy Society General Meeting, addresses a growing operational challenge: large data centers are increasingly supplementing utility power with on-site generation. Managing these additional generators significantly complicates power system operations and stretches the limits of current grid optimization tools.

To address this, ORNL researchers developed a “relax-and-round” optimization method that dramatically reduces the time required to determine which generators should operate at any given moment. The approach first “relaxes” the problem by temporarily allowing traditionally binary (yes-or-no) decisions to take on continuous values, enabling rapid solutions through efficient mathematical techniques. Once a solution is found, the method “rounds” those values back to practical, discrete decisions regarding generator status. This two-step process sidesteps the heavy computational burden of traditional mathematical optimization while still delivering high-quality, reliable results.

The team validated the approach using a baseline system of 46 generators with varying operating characteristics to represent different generator types. By creating slightly modified copies of these units, researchers simulated internal variations and progressively scaled the test systems in increments of 46 generators. On a 276-generator system, the new method reduced computation time from approximately 10 hours to less than one second, with no increase in cost or loss of feasibility. While larger systems became impractical to solve using traditional methods, the new approach scaled efficiently to systems comprising tens of thousands of generators. This scalability positions the method as a practical tool for planning future power grids accommodating large AI data centers and other distributed energy resources.

“The scientific community has developed powerful solvers for continuous-variable problems, but not an efficient way to directly solve integer-variable problems,” said ORNL research scientist Shaked Regev, who led the project. “Instead of an all-purpose solver, we focused our attention on the specific problem of unit commitment and used powerful, problem-specific heuristics to reduce the integer-variable problem to solving multiple continuous-variable problems, which is faster and scales better.”

Industry analysts project that U.S. data centers will add tens of gigawatts of new electricity demand over the next several years. As utilities expand grid capacity, many developers are investing in natural gas generators and other on-site power resources to guarantee reliable operations. These newer generating assets are expected to be smaller, more flexible, and faster to respond than traditional power plants, necessitating updated scheduling and dispatch strategies. ORNL’s optimization method enables operators to quickly identify cost-effective operating plans, reducing fuel costs while maintaining grid reliability.

Designed for parallel computing and compatible with graphics processing units (GPUs), the algorithm opens the door to even faster performance on leadership-class computing systems. The work underscores ORNL’s expertise in power systems, applied mathematics, and high-performance computing, supporting DOE efforts to modernize national energy infrastructure while enabling continued growth in AI and advanced computing.

“Working in a multidisciplinary team allows us to solve more complicated problems,” Regev noted. “I enjoyed learning a lot about power systems and teaching others about the quirks of mathematical optimization solvers. We were able to challenge assumptions about the correct way of doing things, and this led us to the breakthrough.”