Appliance signatures in meter data

How can a single electricity meter reveal the moment you boiled the kettle?

There is exactly one meter on a house, and it sits where the supply comes in, long before the wiring fans out to sockets and lights. It cannot see the kettle. It has no sensor in the kitchen, no idea which room is which, no list of what you own. All it measures is the total flowing past a single point.

And yet an energy app can plausibly tell you that you boiled the kettle at 07:12, that the immersion heater ran for forty minutes, that the fridge is cycling more often than it did last summer. That should feel suspicious. How can a measurement of one aggregate quantity be turned back into a statement about individual devices that were never measured?

Begin by asking what the meter is really recording. Not a smooth quantity but a sum — at every instant, the total of whatever happens to be drawing power. And crucially, appliances do not fade in and out. They switch. A kettle goes from zero to about three kilowatts in a fraction of a second and back to zero when it boils.

So the total is a staircase. Each step up is a device turning on; each step down is one turning off. Nobody records the kettle, but the step is a faithful trace of the kettle, because the sum changed by exactly the kettle's demand at exactly the moment it switched. That is the whole idea: you never observe the appliance, you observe a change in the aggregate that only the appliance could have produced.

Now push on the obvious objection. A step of three kilowatts could be a kettle, a shower, or an oven element. Height alone is not identifying. What else does the step carry?

Rather more than you might think, because the supply is alternating and there are two components to what a device draws. A purely resistive load — a kettle element, an immersion heater — draws current exactly in step with the voltage and consumes what it takes. A motor is inductive: its magnetic field pulls current out of phase, so it also draws reactive power, which sloshes back and forth without being consumed. Measure both and every step becomes a point on a plane rather than a number on a line. A three-kilowatt step with almost no reactive component is a heating element; a two-hundred-watt step with a large one is a motor. That two-dimensional signature is the core of the method George Hart developed through the 1980s and published in the Proceedings of the IEEE in 1992, and it is still the backbone of non-intrusive load monitoring.

Further clues stack on top. A fridge compressor draws a brief surge many times its running current as it starts. A washing machine is not one load but a sequence — fill, heat, tumble, spin — so it appears as a recognisable pattern of steps over an hour. Every on-step must eventually be matched by an off-step of about the same size, which lets an algorithm pair events and reject spurious ones. Duration helps too: three kilowatts for three minutes at seven in the morning is far more likely to be a kettle than a shower.

Now the constraint that decides whether any of this is possible at all. All of it depends on seeing the step. If the meter reports one number per half hour, the kettle's three minutes are averaged into thirty and arrive as a modest bump indistinguishable from a hundred other explanations. Great Britain's smart meters record half-hourly data for settlement, and the privacy framework makes coarser data the default with consent required for finer — but the meter can also broadcast a near-live reading roughly every ten seconds over its home network to a device you connect yourself. That gap in sampling rate, not any cleverness in the software, is what separates "your usage was high on Tuesday" from "you boiled the kettle at 07:12".

Be straight about where the method weakens, because confident labels in an app hide real uncertainty. Two appliances of similar size and character are hard to tell apart, and simultaneous switching merges steps. The central assumption — that appliances are on or off — is eroding: inverter-driven fridges, heat pumps and washing machines vary their power continuously by design, so they ramp rather than step and offer no clean edge. Small electronics sit below the noise floor of everything else. Published accuracy varies widely by appliance and dataset, and identification is best for large, distinctive, on-off loads — exactly the kettle, and much less so the router.

One consequence follows from the mechanism rather than from any extra data collection. If a step is a proxy for an appliance, a meter is not only measuring energy; it is producing a record from which occupancy, sleeping hours and mealtimes can be inferred, none of which anyone chose to report. That is a consequence of resolution, not intent — which is why the argument about data granularity is really an argument about what a proxy quietly implies.

Three refinements.

First, disaggregation is inference, not measurement. Every appliance label is a best explanation of an event, and it can be wrong in a way that a direct plug-level monitor cannot. Treat the numbers as estimates with a confidence attached, whatever the interface implies.

Second, the interesting quantity is the derivative, not the level. What identifies a device is the change at the moment it switches, which is why one high reading tells you very little and a well-sampled trace tells you a great deal.

Third, the method must learn what lives in your house. Some systems ask you to confirm a guess, others train on public datasets of labelled household traces — and accuracy depends on whether your appliances resemble those.

• Why inverter-driven appliances are steadily making this inference harder.
• How reactive power differs from real power, and why only one is billed to households.

Every term the collection defines is gathered in the glossary.