Home electrification to stabilize the grid

Electric heat pumps can lend a helping hand to stabilizing the power grid. 

“It's the combination of having a technology that consumes less energy, is more flexible, and much more controllable that allows for grid stability to be increased if used properly,” said Mario Bergés, a professor of civil and environmental engineering at Carnegie Mellon University. 

As an electric heating source for homes, heat pumps also support reducing household carbon emissions, a priority area of Trane Technologies, a Grand Challenge Partner at CMU’s Scott Institute for Energy Innovation. 

Bergés received support from the Scott Institute’s seed grant program to refine a model called Ibex-RL, an upgrade of a previous model that uses both machine learning and physics-based concepts to control heating, ventilation, and air conditioning (HVAC) functions in buildings. Specifically, Ibex-RL includes thermostat controls for room temperature and use of backup heating when electric heat pumps aren’t enough. Bergés worked with a research team that included his PhD student, Ozan Baris Mulayim, who was a key contributor to the studies. 

Taking the algorithm from a perfect simulation into a real, occupied home forces you to confront the messiness of the real world.

Ozan Baris Mulayim, PhD candidate, Department of Civil and Environmental Engineering

As its name suggests, Ibex-RL uses reinforcement learning (RL), a method that improves a model’s decision making by interacting with an environment—an appropriate choice for tailoring a home heating system. In the case of such systems, one of the most important metrics to determine is the set point, or the base room temperature that the heating and cooling systems work to maintain. 

Ibex-RL also carves a path toward a scalable solution that easily puts energy efficiency into homes—in 2022 it was estimated that 13 percent of U.S. greenhouse gas emissions were from direct residential and commercial sector emissions. 

“Typically, the problem you're trying to solve is that you want to keep the occupants comfortable and lower their energy costs,” said Bergés. “One can let an algorithm learn through interactions how to make this work, but using a model allows the algorithm to learn much faster. In our previous work we used a very simple model to great success, but for more complex scenarios, such as using heat pumps, we need more complex models like what Ibex-RL is designed to leverage.” 

Keeping a home comfortable not only saves money but lessens the strain on the power grid—a particular concern when usage spikes during heat waves and cold snaps as occupants crank the air conditioning or heating in response to extreme temperatures. 

The seed phase of this research culminated with Ibex-RL being deployed in an all-electric testbed home that belongs to Purdue University. During the testrun, the researchers paid attention to how well Ibex-RL’s controls maintained the set point and how often backup heat was used—something preferable to avoid due to the cost increase. 

“Taking the algorithm from a perfect simulation into a real, occupied home forces you to confront the messiness of the real world,” said Mulayim, who studies civil and environmental engineering. “Seeing it learn intelligent strategies on its own—and ultimately deliver a practical 22 percent in energy savings over the existing controller—validates our physics-informed approach as a robust and effective solution for real-world building control.”

“Tracking occupant comfort and activity will be a likely second chapter of this work,” said Bergés of how this project will grow. 

The researchers published a paper with the results from the testrun. Bergés, Mulayim and their collaborators presented additional findings at ACM BuildSys 2025 in Golden, CO, in November 2025.