What's new in v23 and Web App v5.5: Smarts Plugs and Themes


My 2c. In your case, just plug the outlet strip that feeds them all into the smart plug. No way that Sense will find the TiVo, Blu Ray, or Yamaha AV at this point in time.


I was hoping to get an accurate reading on what the TV was using power wise since I really don’t have any idea at the moment. I was also hoping that by isolating the TV, since it creates a lot of noise, might lead to other devices being detected. As far as the other AV equipment, I am curious about the Yamaha and Tivo since again I really have no idea. I might isolate the TV and put the other equipment on a different plug just to get an idea of usage. Does anyone have their Tivo detected? It i always on and the power patterns should be consistent. The Yamaha is a whole different story and probably similar to the plasma as far as power usage goes. Does anyone have an AV receiver detected?


Isolating the TV would work as well. If it’s a smart TV of the right model, Sense might find it. I wouldn’t bet on the TiVo or the the Yamaha though. I have had a Sense for nearly 2 years and have had a TiVo, a Tivo mini and the Yamaha Aventage RX2040 for the whole time - none detected. But that’s fairly easy to understand because all of them have always on baselines with content-dependent active power. All have highly varying on and off signatures when they go into and out of different active modes.


@RyanAtSense instead of scoping the problem:

to combining multiple smart plugs into a single device, how about Sense just looks at the larger UI/UX issue of the bubble screen as it gets better at identifying unique devices? When Sense gets really good, or when we buy a bunch of smart plugs, that screen (especially on a phone) is going to get out of control quickly.

So, perhaps give users a way of grouping devices. Not merging them, but for the purpose of display, reporting, etc - group them. By room, by type, by consumption level, by whatever… Add user-defined tags for devices, let us use those and other things programmatically obtained by Sense to filter/group/etc.

Just my few cents-


@RyanAtSense Any chance we can see support for these next? :slight_smile:

I’d love to get a couple of these if they would be supported. Without them being supported it’s no really worth the expense to me though as regular power strips work fine. As far as I know each outlet on these strips + the USB port do report individual usage so these would be awesome for my electronics closet, office setup, Entertainment setup, etc to get each individual device reported :slight_smile:


Looks like a cool power strip. Based on some reviews, it does monitor energy usage by plug. But it also uses a very different pair of chips for power monitoring (STMP33’s from ST Microelectronics x2). Hopefully the Kasa firmware abstracts all the differences between the underlying hardware (sample/update rates, device/plug identification) away, so it is easy for Sense to integrate.


Piling on here too id like to see that


Yeah that’s what I’m hoping too. I’d be a very cool power strip to have like 4 off :stuck_out_tongue:


Just an FYI to all, Sense collects data from the HS110’s 2 times/second (I’ve seen closer to 3.) Pretty hard to detect an accurate signature at that, but hey it’s nice to get a general idea.


They are still monitoring the mains at much higher rate. I believe HS110 just give them additional datapoints; they aren’t used solely for device identification.


Are you really seeing 2-3 unique datapoints per second ? How are you measuring that rate ? I don’t think the underlying V1 HS110 measurement hardware is capable of greater than one measurement update per second. Could you be seeing the same data sampled twice per second ?

ps: OK the hardware is capable of faster updates, but the default reference configuration is 1 measurement update per second.


Compared to how well you handle data, I am quite crude :slight_smile:. I meant I get one measurement or update every 2-3 seconds, I’m sorry, very big difference.

My measurements come from me counting between updates. My well pump controller changes very fast during ramp up and down which allows me to count cause its literally always going to be a different wattage.


This AI stuff is way beyond me. But I might be able to understand the variables involved if someone laid them out. Maybe others are in the same boat.

I do understand that the problem is to detect individual device consumption. But there is a big leap for me to just say do it with AI.

At base, what does the AI look at? I can imagine that power consumption is a complex wave captured by the CTs. Maybe there are myriad “frequency patterns?” discernible in the wave? Do such patterns correspond to devices? Do they contain information on how much each device draws?

Or maybe the devices correspond to patterns and consumption is separate and the problem is to correlate pattern start/stop with changes in consumption? - as well as detect such patterns at all?

I’m so lost. I’m hoping for a leg up here.


I love the integration. I just installed 2 Wemo plugs and have 3 more on the way. This goes a long way solving the outstanding issues with un-detected devices and “Other”.

I am working on a efficiency improvement project and I just installed an Aeotec heavy duty switch.

It would be great if you could integrate with Aeotec too. This looks like a good solution for difficult to detect high power applications 240v - 40 A.



Maybe the best thing to help you understand is to look at how AI “recognizes” things that we humans also “recognize”. One of the challenges envisioning how Sense works is that it deals with recognition within data that our brain hasn’t been programmed to recognize.

I find diagrams like this to be helpful in understanding how Convolutional Neural Network (CNN) learns how to recognize faces. After training, each successive layer in neural networks is able to identify increasingly complex visual structures that are similar, but not necessarily the same as the ones used in training.

A “time domain” example that might help is how RNNs/LSTMs (RNN/LSTM are the buzzword acronyms for time domain neural networks) can be used to learn how to parse the keywords, syntax and basic semantics of either natural languages or structured languages without the need to actually build a parser, simply by supplying a large enough dataset. Take a look at the examples at the end of this paper to see how neural networks can “learn” how to “understand” Shakespeare, LaTex algebraic proofs, Wikipedia articles, etc., and write similar, based on a stream of text from any of these sources. Fascinating to see how “neurons” in the network learn to fire, just like syntactic and semantics productions in a parser.


I would guess that Sense has a set of similar neural networks that work similarly in the time domain, that do similar for short time windows and perhaps for much longer time windows.

Inputs/features coming into to a Sense classification detector are likely:

  • voltages on each leg of the power, or maybe just the delta since the last time period
  • the same for leg currents
  • the same for power on each leg
  • the phase angle between the current and voltage for each leg - that has to be computed on some kind of sliding window basis.
  • perhaps more sophisticated frequency domain analysis of the waveforms (harmonics)

Plus ground truth feedback when available:

  • voltage, or delta voltage for a given device
  • current, or delta current for a given device
  • power, or delta for a given device

Sense produces prediction outputs each time period or every few time periods of:

  • Device on
  • Device off
  • Device power consumption

When this data is used for training back at the mothership, error data, representing the difference between what Sense predicted (probably every few microsecond time ticks) and the ground truth, is used to update hundreds of thousands of parameters/weights within the neural networks, to make the network more sensitive to the ground truth waveforms. To understand this process in greater detail, you’ll need to go deep on linear algebra, feedforward and feedback in neural networks, plus gradient descent optimization.


That would be a really cool device to support if it talked WiFi instead of Z-wave.


Yes Z-Wave is a bit of a pain, but I got the Samsung hub and it works well enough. I guess it should be OK to integrate in a similar way as for the Wifi plugs. The switch has energy monitoring and on off trough the mobile app.


There seems to be a problem if a device is detected after a plug is installed. The new device you add to the plug doubles the wattage.

For example here is my fridge. It was detected on Sat and merged with the plug, everything before then was doubled, everything after is merged as expected.

Same for my furnace


Thanks Kevin.

It would be nice to dive into AI, but I expect it would be years before I could appreciate it. But understanding the problems it is trying to solve seems within reach.

One thing that popped out for me in your explanation is the recognition of on and off patterns. I had been thinking that Sense was recognizing device consumption by some corresponding wiggles in the inputs. Maybe that’s not it at all and what it is really trying to discover is the ons and offs. That explains how it could sometimes miss one rather than just noticing that it is currently on or off.

A working model for me then becomes that Sense uses AI magic to detect ons and offs from the wiggles in its input. Then correlates them to changes in power consumption.


Will this new feature of sense work with other models of the same brand? I just bought this Wemo plug: