Trane Technologies, a climate-control technology company, has developed a predictive HVAC control system that recorded energy savings of up to 19% in early testing.
The system forecasts how a building will respond to different heating and cooling decisions, then compares thousands of possible operating strategies against energy costs, emissions and comfort requirements. Trane said the energy saving was more than 3% higher than that achieved by BrainBox AI’s existing control algorithms.
The technology remains in the exploratory phase. Trane plans to conduct field trials before considering commercialisation.
The system is based on research carried out at the BrainBox AI Trane Technologies AI Lab, with support and contributions from researchers at Mila, the Quebec Artificial Intelligence Institute.
Field trials are the next stage of development.
HVAC operation planned hours in advance
Most building controls wait for sensors to detect a change before responding. Trane’s system forecasts the building’s thermal behaviour and plans its HVAC operation ahead of time.
It can pre-cool a building in summer or pre-heat it in winter before electricity prices rise. Heating or cooling can then be reduced during the most expensive period while indoor temperatures remain within limits set by the building operator.
The system combines model predictive control with Neural Ordinary Differential Equations, known as Neural ODEs. The Neural ODE model predicts indoor temperature and HVAC power demand. The controller uses those forecasts to decide how the equipment should run.
Jean-Simon Venne, president, founder and chief technology officer of BrainBox AI and head of the AI lab, said the system could simulate thousands of control strategies within seconds. It then selects one based on its likely effects on costs, emissions and occupant comfort.
The calculation is repeated during each control cycle using updated sensor readings, forecasts and operating conditions. Building managers can specify acceptable temperature ranges, limits on equipment cycling and the emissions profile of the electricity grid.
They can also configure the controller to reduce total energy use, operating costs or peak demand. When grid-emissions data are available, the system can take changes in the carbon intensity of electricity into account when scheduling heating and cooling.
Neural ODEs need less building data
Trane said the Neural ODE model requires less training data and computing power than traditional Long Short-Term Memory models. The company said this would speed up the onboarding of new buildings while reducing the cloud computing required.
BrainBox AI has previously said Neural ODEs reduced the data needed to train its models tenfold. Models that had required weeks of building data could reach the same accuracy with several days of readings, according to the company.
BrainBox AI also said the models could be trained using central processing units rather than graphics processing units, reducing the computing resources and energy involved.
Trane completed its acquisition of BrainBox AI in January 2025. The companies had previously worked together for more than two years. Trane announced the BrainBox AI Lab in August 2025 to work on autonomous controls, predictive models and other uses of artificial intelligence in building energy management.
