You make a great argument - the timescale / interval we’re using to look at power usage really affects how we perceive it. And depending on what trends we’re looking for, the on-signature of a motor, vs the on-signature of an EV, vs the typical daily use of AC, vs. the typical seasonal use of AC, vs degradation of solar production from PV panels, all require different analysis timescales.
On the flip side, I see Sense’s definition of Always On to be somewhat of an absolute measurement, rather than a type of follow-on analysis. They attempt to quantify and categorize a subset of usage that will never be detectable using the Sense ML techniques. The thing that’s a bit confusing to users and makes house-level Always On a bit incongruous, is that it needs to be “measured” using different techniques than all the other real devices, and measured on a different time scale in order to filter out all the daily cyclic activity. House-level Always On is also a little discordant in that it represents components from many different devices, although you can also get Always On from a single device using a smartplug.
Really, Always On is an orthogonal measurement to Sense ML device detection, meant to find real-time detectable power/energy usage that the Sense ML techniques never could. And in fact, it finds opponents of your refrigerator and furnace usage, that’s never identified, even if Sense picks up the regular cycles of your fridge or HVAC.
So again, the current Always On represents a real fundamental but orthogonal measurement, not an scale-based analysis as you describe above. The 24-48 hour measurement window Always On uses, may make it seem like an analysis rather than a measurement, but that filter was carefully picked to remove daily cycles of things that Sense should pick up (recirc pumps, lights, etc. on timers) on the low end and anomalous events on the longer end (you don’t want to use 1 month because that could let a network outage dominate your Always On for a month).
But I’ll admit there is a gray area in the Always On measurement, for transitions at low end of the power scale for specific devices on smartplugs (devices that Sense is either incompletely or never detecting using ML). Consider the profiles of measured devices like my modem, Time Machine, or AppleTV. All of them have a relatively low power operating mode, but look like they have a still lower standby mode as well, that would likely be more clearly defined with a tighter data sample window. The current Sense Always On algorithm for that device is going to home in on the lower end of that distribution (1% bin), but given Sense wouldn’t likely detect the changes, the whole of the device usage really fits into the house level definition of Always On. The Always On for the house and the Always On for a device are calculated similarly, but have somewhat different meaning because of the scale you highlighted. That’s where your notion of a threshold makes sense. But I would rather that Sense do some mode analysis for us, to help us make that decision.