AI for Building Automation: What’s Real and What’s Hype

What an Analytics Layer Actually Reads

The distance between "the building has data" and "something can analyze it" is mostly a question of how much each source says about itself. A value on its own is a number. What makes it analyzable is knowing what it measures, in what units, on which piece of equipment, and how that equipment relates to everything around it. That description exists somewhere on every project. The only real question is whether it lives in the station, where the people making changes to the building can see it, or gets rebuilt by hand inside the analytics tool, once per site, and maintained there.

Where you read from also carries a cost that has nothing to do with modeling. An analytics layer polling the field network directly becomes another client competing for the same bus, and on BACnet MS/TP over RS-485 the token passing rate, together with Max_Master and Max_Info_Frames, puts a hard ceiling on how much can be requested per second. A second poller there can slow the loops the controllers are actually executing. A station that already holds current values and histories is the cheaper read, and it does not put analytics traffic on wire that control depends on.

Sources differ sharply in how much of the description they carry:

• A Niagara station with a tagged model. Points resolve by path, facets carry units and type, history extensions exist where collection was configured, and tags and relations record which equipment a point belongs to and what that equipment serves. A question resolves to a specific point rather than to a string that resembles one.
• A Niagara station without tagging. The points and histories are there and the names are usually consistent, but consistent is not semantic. Retrieval falls back to matching name fragments, so every question becomes a question about the naming convention, and any equipment commissioned by a different hand is a separate special case.
• Exports and reports. A CSV or a report export is whatever columns the exporter chose, frozen at the moment it ran. Units appear in the header if they appear at all, and timestamps are frequently written in site local time with no UTC offset.
• Direct protocol polling. Modbus TCP on port 502 returns 16-bit register values with no self-description at all: type, scale, sign and meaning live in the vendor's register map document and nowhere on the wire. BACnet/IP on UDP 47808 tells you considerably more, since Object_Name is required and unique within the device, analog objects carry a required Units property of type BACnetEngineeringUnits, and a device may already hold Trend Log or Trend Log Multiple objects readable with ReadRange. What no BACnet read gives you is equipment relations: which air handler that discharge temperature belongs to, and which terminal units it serves. The standard's optional Structured View object can express a hierarchy, but you cannot count on finding one.

History Collection Decides the Answer Before Anyone Analyzes Anything

An interval history samples, and the sample rate is a ceiling on everything downstream of it. A status point collected on a fifteen minute Boolean interval extension cannot show a compressor short cycling on a five minute period; the series looks calm because the collection is aliasing the behavior, not because the equipment is stable. Nothing in the analytics layer recovers information the collection never captured, which makes the interval a design decision about the questions you intend to ask, taken before anyone asks the first one.

Change-of-value collection has the opposite property. Records appear only when the value moves past the configured delta, so a genuinely steady value and a dead sensor produce the same thing, which is nothing. COV series are also irregular, and most analysis wants a regular grid, so a resampling step exists somewhere whether or not anyone chose it. Whether that step holds the last value across a gap, interpolates across it, or marks it missing changes the number the analysis reports.

Capacity and transport set the rest. A Niagara history extension with a record count capacity and a Full Policy of Roll discards its oldest record once full, so retention is a function of collection rate: at one minute a point writes 1,440 records a day, so a 10,000 record capacity holds roughly a week, where the same capacity at fifteen minutes holds roughly three months. Histories collected on a JACE and archived to a Supervisor transfer by record, so a gap that opens on the Supervisor during a link outage generally backfills once the link returns. A gap still sitting there after the link is healthy is a different problem, and worth separating from the first one before anybody starts explaining the data.

Three symptoms that get misread as equipment or sensor problems:

• Two points on the same equipment disagree about when an event happened. Check the clock before the sensors: history records are stamped by the station that collected them, so a JACE whose time has drifted from the Supervisor stamps everything it owns with that offset.
• A value sits perfectly constant for hours. Read the status, not just the value. A driver holding a last known value, a point left out of service, and a manual override all present as clean data to anything reading value alone; Niagara status flags and the BACnet Status_Flags and Reliability properties are where the difference is visible.
• One hour appears twice, or disappears entirely, once or twice a year. That is a timestamp written in local time with no UTC offset meeting a daylight saving transition, and it quietly corrupts any comparison spanning the boundary.

Advisory Is a Wiring Decision, Not a Description

If anything ever writes back, the mechanism decides whether the word advisory means anything. On BACnet, writes to a commandable object arbitrate through a sixteen level priority array in which the lowest numbered active entry wins and Relinquish_Default applies when every level is NULL. The standard assigns 1 and 2 to life safety, 5 to critical equipment control, 6 to minimum on and off timing, and 8 to manual operator, which leaves 9 through 16 available for supervisory and optimization writes. A recommendation written at 14 is therefore a recommendation in the literal sense: an operator at 8, and every interlock below that, take it away without anyone touching the analytics layer.

The failure that follows write-back is the write nobody relinquishes. A supervisory process that commands a value at some priority and then stops, crashes, or loses its network path leaves that entry populated indefinitely, and the local sequence has no way to reclaim a level it does not own, so the equipment stays commanded to a decision made from data that is now days old. There is a useful asymmetry here: a COV subscription is issued with a Lifetime and lapses when it is not renewed, while a property write carries no expiry at all. Anything that writes needs either a heartbeat that relinquishes to NULL on timeout, or logic in the station that clears the level when the supervisory link is absent.

Four decisions have to be made before write-back is enabled, and none of them are questions about the model:

• Which priority level a supervisory write uses, and which levels must always beat it.
• What happens to that level when the writer stops: a timeout that relinquishes to NULL, or station logic that clears it.
• The clamp, meaning the range outside which a recommended value is refused at the station rather than written to the equipment.
• Who is told when a write is refused, clamped or relinquished, and where that appears, since a silently rejected recommendation looks exactly like one that was accepted and did nothing.

Where a Rule Already Does the Job

A large share of what gets sold as machine learning fault detection is a threshold and a timer. Stuck valve, short cycling, simultaneous heating and cooling, a sensor reading outside its physical range, a damper commanded one way while the temperature moves the other: these are deterministic rules over a handful of points, and they beat a learned model on the thing that matters operationally, which is that they name the fault instead of scoring the hour. ASHRAE Guideline 36 goes further and writes a set of these alarms into the sequences of operation themselves, so on a job built to it, part of the fault list is already specified and already executing in the controller rather than waiting on a layer above.

What a rule cannot do unaided is catch slow drift, because drift needs a reference and a lone sensor has none. The available references are physical rather than statistical, and if none of them exists on the equipment in question, the honest answer is instrumentation rather than analytics:

• A mixed air temperature that has to fall between return and outside air, weighted by damper position.
• Two sensors on the same header, or supply and return on a loop where flow is known.
• A main meter against the sum of the submeters beneath it.
• A relationship the equipment cannot physically violate, such as a coil that keeps cooling with its valve commanded closed, which turns any violation into a sensor question rather than a mystery.

When the Honest Answer Is Not Yet

Some buildings are not ready for an intelligent layer, and they are recognizable in advance. If the existing alarm console already produces findings that nobody acts on, a layer producing more findings raises the ignore rate rather than the response rate, because the bottleneck is the response process and not detection. If the points that would answer the question were never installed, no amount of modeling substitutes for a sensor. And if collection only recently started, anything weather dependent has no baseline yet, since comparing a chiller against outside air temperature requires having seen a range of outside air temperatures.

The other thing worth knowing before committing is that the ongoing cost of an intelligent layer is not the model, it is the mapping. Every controller replacement, retrofit, renamed point and added piece of equipment breaks part of the correspondence between the building and its representation, and it breaks quietly: the analysis keeps running against the stale version and keeps looking plausible. Someone has to own that maintenance. When the model lives in the station, in tags and relations rather than in an external spreadsheet, the maintenance at least happens in the same place as the change that caused it.