「英中对照」AutomatedBuildings编辑委员会探讨：当建筑运行数据出错时AI应当如何应对？

From the AutomatedBuildings Editorial Board

AI in Real Buildings: When the Data Doesn’t Agree
From the AutomatedBuildings Editorial Board

Real buildings rarely provide perfect information.

Sensors drift, occupancy changes, forecasts move, equipment is overridden and systems sometimes disagree. Some data is missing, while other data looks perfectly reasonable even when it is not.

As AI takes on a larger role in analysing and acting on building data, the important test is how it performs under those normal operating conditions.

Take weather as a simple example. A system might use a forecast to decide when to pre-cool a building or shift a load. The forecast may change, local conditions may differ, occupancy may not be what was expected, and a sensor inside the building may be giving a believable reading that does not accurately represent the space.

The system then has to decide what to do with information that does not line up neatly.

Can it recognise the conflict? Can it seek another source, compare the information with previous behaviour, reduce its confidence or adjust the decision? At what point should it fall back to a known operating condition or ask someone to intervene?

Those are important questions for any system expected to act autonomously.

Real buildings will always contain some uncertainty, so the answer cannot simply be to wait until every input is perfect. The more useful question is whether the system can recognise uncertainty and respond appropriately when it encounters it.

It also changes what should count as proof.

A successful demonstration shows that something worked in that instance. A pilot provides more evidence. Deployment across different buildings, seasons and operating conditions provides more again.

Trust develops over time, as owners and operators see how a system performs when conditions change, equipment develops problems, sensors drift, overrides are introduced and the many exceptions of day-to-day building operation begin to appear.

AI may also be able to do things an operator cannot easily do, such as compare large amounts of information continuously, detect inconsistencies across systems or identify patterns that would otherwise be difficult to see. Those capabilities could be valuable, but they still need to be demonstrated in operation.

Owners and operators should be able to understand enough about the system’s behaviour to answer some practical questions. What information did it rely on? Did it recognise uncertainty? How did it respond when sources disagreed? What happened as a result?

Trust in an autonomous system will come partly from the technology itself, but also from the evidence accumulated through real use.

As AI moves from analysing building data to influencing or making operational decisions, that evidence will matter more.

The question is not simply whether a system can perform when the conditions are right. It is how well it responds when they are not.

If you are already using AI in live building operations, we are interested in the exceptions. When inputs conflicted, a sensor was wrong or reality did not match the model, what did the system actually do? We would like to hear the examples that worked, and the ones that did not.

Join us in Chicago at AHR2027 to continue the conversation. 

The post When Building Data Is Wrong, What Should AI Do? appeared first on AutomatedBuildings.com.