A new machine learning approach developed by researchers at the Johns Hopkins University and the University of Mons could help property owners better predict and manage energy usage and costs, potentially reducing both electricity bills and greenhouse gas emissions. The team’s Decision-Focused Learning (DFL) model can accurately forecast the cost of heating, ventilation, and air conditioning (HVAC) in commercial and residential buildings.
Their results appear in E-Energy ’25: Proceedings of the 16th Association for Computing Machinery International Conference on Future and Sustainable Energy Systems.
Buildings account for about 13% of U.S. greenhouse gas emissions, according to the Environmental Protection Agency, with this figure climbing to 31% when building electricity consumption is included. The majority of this energy demand stems from HVAC systems working continuously to maintain comfortable indoor temperatures.
“Optimizing energy use in buildings is one of the most impactful ways we can reduce carbon emissions, but HVAC systems are some of the most complex systems to design in terms of optimal energy usage,” said Yury Dvorkin, study co-author, associate professor in the departments of Civil and Systems Engineering (CaSE) and Electrical and Computer Engineering at Johns Hopkins Whiting School of Engineering, and the U.S. director of the National Science Foundation’s Electric Power Innovation for a Carbon-free Society Center (EPICS).
Dvorkin says that each building has distinct characteristics, such as their age, size, local climate, and occupancy levels that affect energy consumption and cost. He noted that even an imperceptible temperature adjustment on the scale of just one degree Fahrenheit can lead to substantial changes in energy use and utility costs.
Traditional methods of predicting heating and cooling use in buildings rely on physics-based models, which are highly accurate, but difficult to scale due to each building requiring significant customization to address unique characteristics. Their complexity also makes physics-based models hard to scale quickly, the researchers say.
In contrast, simpler AI-driven data models promise faster forecasting, but often ignore the physical realities of how buildings work, sometimes recommending impractical or even impossible settings. Pietro Favaro, study co-author and doctoral candidate at the University of Mons, also points out that historical building data used to train the AI models prevents the models from learning where they could have the greatest impact, as the historical data often repeats daily routines.
By combining physics-based and AI models, the team developed DFL, which allows users to optimize decisions that reduce energy usage and cost, rather than simply aiming for a specific temperature or temperature range.
“What makes DFL unique is that it directly connects predictions to outcomes,” said Dvorkin. “It doesn’t just model temperatures—it helps users decide how to run HVAC systems to meet their goals, whether that be a lower level of emissions or a lower electricity bill, by taking the downstream effects into account.”
Favaro, who was a previous year-long visiting student with Dvorkin’s group in the Whiting School of Engineering’s Department of Civil and Systems Engineering and Ralph O’Connor Sustainable Energy Institute, validated DFL using real data from buildings and simulated experiments on a multi-zone building. Instead of looking for a single “correct” temperature, the researchers focused on the model’s margin of error and tested how each suggestion ultimately affected energy use and cost. Over time, the DFL method not only obeyed the laws of physics but also learned which choices led to the most energy-efficient outcomes.
Dvorkin and Favaro say their approach has far-reaching implications. DFL could help reduce the environmental impact of heating and cooling, while also freeing up electricity for other uses—or for export to regions facing power shortages. As global demand for electricity and computing power grows, driven by the rise in AI and data centers, such energy savings could bolster grid resilience. The team also found that energy savings translate directly into cost savings—a win for property owners and their electricity bills.
“While we focused this study specifically on HVAC management, the methodology we developed can be applied to a broad range of systems, including those that are more complex, like data centers, which is something we would like to explore in collaboration with CaSE’s Jan Drgona,” said Favaro.
The team’s paper was awarded Runner-Up in this year’s Best Paper competition at the ACM e-Energy conference. Favaro continues to collaborate with Dvorkin on the refinement and applications of DFL.
Thanks to a recent Nexus Award from Johns Hopkins, Dvorkin is now working with fellow Hopkins researchers Magdalena Klemun and Ben Link on Project INTERSECT—a research initiative applying these findings toward affordable and community-responsive data center expansion. The group will also work with Jan Drgona, an expert in building energy management and machine learning, to develop solutions for thermal management of data centers.
“With technologies like DFL, we have an exciting opportunity to translate academic research directly into solutions that reduce energy costs, improve air quality, and develop a more reliable electrical grid,” said Dvorkin. “This is the kind of real-world impact our communities need.”
Additional study co-authors were Jean-François Toubeau and François Vallée of the University of Mons.