New AI Cuts Building Energy Bills Without Any Custom Setup

Heating, cooling and ventilation usually eat up a huge share of a building's energy bill. For years, engineers have known that a smarter controller — one that looks ahead and plans, rather than just reacting to the current temperature — can cut that waste. The catch has always been setup: you need months of data from your specific building, plus an expert to build a custom model. That cost and hassle is why most offices, schools and apartment blocks still run on simple thermostats.

This paper attacks the setup problem directly. The researchers pretrained an AI on many buildings at once, so it learns general patterns about how buildings respond to heating and cooling. Then they drop it into a brand-new building and let it run with no extra training at all — what engineers call "zero-shot." That's like hiring a building manager who has worked in hundreds of similar buildings and can start on day one without a handover period.

The clever part is how they gathered the training data. Instead of only recording buildings during normal, boring operation, they deliberately poked and prodded the systems — turning things up and down on purpose — so the AI saw how each building reacts across a much wider range of conditions. That "excitation" data made the difference. Models trained only on everyday operation performed poorly once actually wired into a control system, even when their predictions looked accurate on paper.

Testing across 32 simulated buildings, the excited-data model delivered the best control of any method tested, beating an online adaptive controller by 6.4% and a standard PI controller by 36.9%. The important caveat: this is all simulation, not real buildings, and real-world plumbing, sensors and occupant behavior are messier. Still, if it holds up, building owners could get smarter, cheaper climate control — and everyone else could see it in lower energy costs and emissions.