"Sense" home energy monitor?

[cvanc]

https://sense.com/

...Comments?

Seems a bit, well, impossible to me.  How could it possibly 'learn' what devices are consuming energy in the house?  It seems to be basically a pair of clamp-on ammeters and a little intelligence.  But to a monitor like this, does my coffeemaker look any different from the space heater in my basement?

[ataradov]

I think we already discussed similar product here, or it may have been this exact thing. It may monitor specific noise patterns or on/off patterns. But mostly it is a lot of BS.

[CustomEngineerer]

Its clearly not BS. Its from the same team that brought speech recognition technology to the market. What more proof do you need?

[cvanc]

Its from the same team that brought speech recognition technology to the market.

Interesting; I hadn't caught that before.  That (to me) makes it seem more likely to be a somewhat real product and not crap.

Let's run with the 'speech recognition' idea for a bit.  Suppose this little box does its' thing at a sufficiently hi frequency and bit depth to resolve the transient behavior, as measured at the power panel, of my coffeemaker turning on each morning.  It's always plugged into the same outlet in my kitchen.  The turn-on (and turn-off?  Hey, why not) behavior as measured at the breaker panel is probably consistent & repeatable.  It (might) have a kind of 'signature' that (might) be unique enough to identify out of all the other transients that (might) be occurring at the same time.

Maybe.  Right?  Might it be able to identify at least some stuff in your house by this method?  Sorting out the voice of my coffeemaker from all the other voices- a different kind of speech recognition, if you will.

[coppice]

Load disambiguation is an interesting topic, that a number of capable teams have worked on for about 30 years. A few products have been around for a while, and they do provide some workable functionality. The snag has always been that there is a big gulf between what people are lead to expect, and what is achievable. It started with people noticing that its easy to see when various appliances switch on and off, by looking at a simple plot of household power consumption versus time. The snag is that getting from that observation to a solid product that reliably disambiguates more than a handful of the house's biggest consuming appliances is really tough - in fact, its probably impossible - and its getting tougher all the time, as appliances get more "electronic" in their load characteristics.

[dorin]

The snag is that getting from that observation to a solid product that reliably disambiguates more than a handful of the house's biggest consuming appliances is really tough - in fact, its probably impossible - and its getting tougher all the time, as appliances get more "electronic" in their load characteristics.

Which is probably why they are using machine learning to deal with this. The capability of ML to classify complex patterns that also evolve in time makes this a promising product.

[coppice]

The snag is that getting from that observation to a solid product that reliably disambiguates more than a handful of the house's biggest consuming appliances is really tough - in fact, its probably impossible - and its getting tougher all the time, as appliances get more "electronic" in their load characteristics.

Which is probably why they are using machine learning to deal with this. The capability of ML to classify complex patterns that also evolve in time makes this a promising product.

Everyone doing this is using some form of machine learning. Sometimes with some human guidance, but there are strong reasons to only use that as a last resort - the average person can't be relied on to give any useful guidance at all. Current results are still pretty weak, though. You only have a few meaningful parameters from which to extract information about the loads - basically active power, reactive power, harmonic patterns, and the start up and stop curves. When you consider the variability of these from instance to instance of an appliance running its pretty hard to achieve reasonable reliability detecting minor loads among the big ones. Its easy to create good demos, though, and I think a lot of people find it technically interesting. Practical and usable is still open to question.

[coppice]

As PFC enabled devices are coming into prevalence, I see device pattern detection harder and harder.
Imagine if all devices emit little to no conducted noise and harmonics and no current lagging or leading. That would be very hard to tell the difference.

Its not just PFC. The load curve as things start and stop is one of the more interesting things you can latch on to for identifying loads. Motors with soft start, and various other refinements in modern equipment, are making these curves less distinctive. I suspect disambiguation systems will eventually be identifying whose algorithms have started and stopped rather than which appliances. E.g. "One of the several things in this house taking a couple of hundred watts, and using one of those TI C2000 based PFCs just came on. Now one of the several things with an NXP custom PFC IC just switched off.".