I’m curious if a high always-on load (real-time, not the bubble) gets in the way of device detection, or at least makes it take longer than it normally would. Some of the always-on devices, like servers, use various amounts of power based on what they’re up to at the time, causing some noise in the graph. I could also imagine the high load also makes the algorithm ignore smaller changes, for example LED lights and smaller TVs.
In general, maybe not that much. But yes, sort of.
Too long answer:
A 0.5-1% inaccuracy in readings (typical) across the current range can translate to noise that would drown out lower load devices. Or you could say “swamped” as others have:
However, I think you are conflating Always-on and device detection. Always-on is a list of devices that Sense determines are always-on (ok, sorry) in addition to (if it exists, I believe) a calculation done on the Power Meter. With Smart Plugs, some Always-on lists can be quite accurate but without those it’s an ongoing calculation done on the Power Meter (the overall waveform) and with weighting (I assume) given to detected devices.
An example, for arguments sake: Let’s say you ALWAYS have your OLED TV on and never let it go to standby … depending upon what you’re watching it might use anywhere between 50-300W. Imagine how the algorithm is going to determine whether it’s Always-on and then how much it would list as the Always-on usage … it’s a calculation done on the overall device usage and not exactly easy to determine although the TV is Always-on.
In some ideal case you would say “It’s been ON for a year and used a total of X Wh so its Always-on is X/8,760”. You can see though that this is an ongoing calculation that will vary and when the device goes off decisions need to be made!
Meanwhile if you have a 10W LED bulb that is being switched on and off (which is well within the OLED TV “noise”), it’s going to take some real learning for Sense to ever detect the bulb. Or imagine you have a second identical TV. Fun!
That said, different devices with different electrical signatures are going to have better/worse chances of detection regardless of whether a high load is ON. e.g. If a 4kW hot water tank is on (fairly stable resistive load) it shouldn’t significantly affect the detection of a non-inverting AC unit switching on and off (inductive load). The difficultly with most Always-on devices is they are all similar … AC-DC power transformers … and low wattage. You are correct in thinking that a server with varying usage will affect detection of low wattage devices … it’s similar to the OLED TV in that regard.
I think the trick, if you can do it, is to assemble your Always-on (literally) stuff onto a power strip and put that on a single Smart Plug (or use the HS300 strip). If you go with the former you are, in a way, choosing to cripple Sense’s ultimate ability to distinguish between low-wattage devices, at least from the human-input standpoint. The algorithm, though, won’t care and can plod along toward ultimate ML perfection.