Removing output, or removing waste

There is a version of sustainability that is really just subtraction. Set the cooling two degrees warmer. Cut the lighting levels. Turn the ventilation down out of hours and a little during them. Consumption falls, the number on the report improves, and the people inside the building quietly begin to hate it.

It works, in the narrow sense that the meter agrees with you. It is also the reason so many energy programmes are silently abandoned in their second year — not reversed by decision, but eroded by occupants who find the override switch, prop the door, bring in a desk fan, and put the whole thing back where it was.

The building-level meter tells you almost nothing

Most buildings know their consumption the way most people know their bank balance: one number, once a month, after the fact.

That number is enough to tell you the bill went up. It is not enough to tell you anything you could act on. A single main meter cannot distinguish a chiller running against a stuck valve from a genuinely hot month, cannot separate the tenant who works weekends from the plant that never stops, and gives you no way to check whether last quarter’s retrofit actually did what the vendor said it would.

Sub-metering changes the question from how much to where. Split the load by circuit and by system — HVAC, lighting, sockets, lifts, kitchen, IT — and by zone, and consumption becomes a map rather than a total. Render that map over time and the patterns arrive without anyone looking for them:

• The floor whose lighting load never quite reaches zero.
• The plant whose baseline crept up eleven percent over eight months.
• The riser drawing power on a floor that has been vacant since April.

Hunting the black holes

The most valuable window is the one nobody is in the building for.

Overnight and weekend consumption is close to a pure diagnostic signal. Whatever a building draws at three in the morning is, almost by definition, serving nobody. Some of it is legitimate — refrigeration, servers, security, freeze protection — and the legitimate portion is stable and predictable, which is exactly what makes the rest of it stand out.

So the analysis is worth automating rather than performing. The system profiles each circuit’s unoccupied baseline, watches for drift against it, and when a circuit departs from its own history it raises the finding — with the deviation converted into currency, because a report that says this AHU has been drawing 4.2 kW more than its baseline since 14 November, costing approximately X per month gets acted on, and a report that says anomaly detected does not.

Peak, not just volume

Volume is only half the bill. The other half is the peak — contract capacity, demand charges, the fifteen minutes in August that set your tariff for the year.

Managing that is a control problem rather than an analysis problem, and it has to be solved in real time. As the load approaches a defined fraction of contract capacity, the system executes a shedding sequence in a pre-agreed order: pre-cool ahead of the anticipated peak, drift setpoints within the comfort band rather than outside it, stage non-critical plant, throttle the loads that hold their value under interruption before touching those that don’t.

Two things make this succeed or fail.

The first is prediction. Shedding reactively, after the peak has begun, is already late. Occupancy and load forecasting — this floor fills at 08:40 on weekdays, the demand curve on days above 32°C looks like this — lets the sequence start early enough to be gentle. A peak avoided by pre-cooling at eight is invisible to the occupants. The same peak handled by cutting cooling at two in the afternoon is a complaint.

The second is the exclusion list, and it is the point where this article and the WELL article have to be read together. Fresh air delivery to occupied zones is not a sheddable load. Nor is task lighting below the level people need to work. A demand strategy that reaches its target by degrading air quality has moved the cost rather than removed it, from the electricity account to the productivity and health account where nobody is measuring.

From kilowatt-hours to the reports people actually need

Consumption data has to leave the engineering domain to be worth anything, and it leaves in two directions.

Where the AI earns its place

None of the above requires a language model. Sub-metering, baselining, shedding sequences and emission factors are ordinary engineering, and they should be done ordinarily.

What the AI layer changes is who can ask the questions.

An energy manager with SQL access and a fortnight can find the eleven percent baseline creep. Nobody else in the organisation can, and the energy manager has other work. Put a language model over the historian and the questions open up to everyone who has a stake in the answer.

The second one returns a comparison rather than a vendor claim. That is the real shift — not that the analysis becomes possible, but that it becomes cheap enough to do casually, and the findings that matter are usually the ones nobody had budgeted a fortnight to look for.

Where to go from here