Towards Predictive Controls
Predictive maintenance analyzes sensor data and operational patterns to detect early signs of wear or malfunction, allowing technicians to service or replace HVAC components before they break down. A related area is predictive control: adjusting equipment proactively to keep a system running more efficiently. It was the subject of Dr. Bursill’s PhD thesis.
“It’s more machine learning than modern AI,” Dr. Bursill says. “It would’ve been very flashy and exciting in the ‘90s, but everything in HVAC kind of lags behind other industries.”
Predictive controls use models to predict where a system is headed, and then react before it gets to that state. It’s an area that was opened up by supply chain shortages that forced many control vendors to purchase new types of chips—and, in buying new chips, many systems were upgraded.
“You’re now seeing gigabytes of RAM on a controller that controls something as simple as a VAV box,” Dr. Bursill says. “So you’re able to do these machine learning fits on the controllers, which changes the game. The LLM didn’t do this, but it can interact with the result, so you get more sophisticated supervisory systems, which is very exciting.”
It’s still early days. Predictive controls in building HVAC systems are still emerging. Adoption is limited: no commercial vendor is doing this at a huge scale on HVAC systems, yet. But the technical capability is there.
“I think it’s coming,” Dr. Bursill says. “There are multiple vendors researching this right now. The bigger issue isn’t really the ROI—you’re using sensors that are already there, and building models that can leverage the data, so it’s not really a hardware cost. The cost is integration.”
Before predictive controls can be deployed, someone has to program them for a specific building’s control system: mapping the algorithm to the correct data points. But each system has thousands of data points: temperature sensors, airflow measurements, valve positions, fan speeds.
The naming and organization of those data points (also known as input/output (I/O) points) at the programming level are often inconsistent: they might’ve been programmed long ago by someone whose process was not standardized or intuitive. Some older systems limited the number of ASCII characters that could be used, so that instead of a temperature data point being named Room101_IndoorTemp, it might be something more cryptic, like Z_T2M89.
“It makes it difficult to rapidly deploy,” Dr. Bursill says. “You have to have someone with a lot of subject matter knowledge in both building controls and HVAC to come and look at this system and figure out what’s what.”
Efforts are underway to address the problem through new industry standards: ASHRAE’s proposed standard 223P focuses on metadata and classification for building control systems, defining the way sensors and data points are described, so software systems can understand them more easily.
Another proposed standard, 231P, would standardize how control algorithms themselves are described and integrated into building systems. Together with updates to BACnet, a communication protocol, these standards would make building automation systems far easier for AI tools to interpret, thereby lowering the integration barriers that currently slow the adoption of predictive controls. It doesn’t fix the existing destandardized data problem, but it will hopefully prevent it in the future.
“The right way to do it is to do it right at the beginning,” Dr. Bursill says. “We could try to fix the past, but that stuff is out there, and people have been working on it a lot already. But going forward, we have solutions to make this work better in the future.”
