Frustrations & Questions

I’ve been running Sense on a fairly standard configuration for 6 weeks. My electricity bill is through the roof on a new house, hoping to use it to figure out causes, monitor usage, the usual stuff. I just can’t seem to get it trained to a useful point, wondering if anyone has any insight.

  1. Why does ‘Always On’ fluctuate so much? I would expect this to change when I buy a new device, but things that cycle frequently should be under ‘Other’, no?

  2. Is there a way to see which phase of the incoming electric a device is connected to? I’ve spent hours throwing circuit breakers trying to identify devices.

  3. Is there a way to stop devices from being changed when they’re detected correctly? Devices keep getting combined incorrectly, driving me nuts.

  4. I have an oil-fired boiler in the basement, runs up to a heat exchanger/blower in the attic. Sense correctly identifies the blower, but keeps identifying the boiler/pump/etc as varying other things. Shouldn’t an ML algorithm correlate the two?

  5. I have a 1500W electric heater for a particularly cold room. Still hasn’t been detected. Given the characteristics of this device, would think it would be one of the first things detected. Not many items draw this much and cycle so frequently.

  6. Half my consumption is still “other” (~375kWh/month). After 6 weeks I’m not seeing any new devices. Is this likely to get better?

  7. I’ve read the forums, seen the extensive discussion of “Why can’t I train my sense”. I understand why it needs to independently extract a device power signature, but really don’t see why you can’t offer “hints”. Tell it occasionally when you trigger a device. If Sense detects a device that highly correlates to your hints, it could tag it with the appropriate device name.

  8. I have 5 refrigerators, 5 heating systems, and 4 stoves. I’m confused as to how these devices are tagged. I can see a fridge (or any motor) having a high initial starting current (especially a compressor), making the pattern fairly unique. Stoves (electric) must be hard, variable draw, but being one of the few 220 devices should simplify the pattern. Why does Sense think there are so many?

I know these are a lot of specific examples that the community may not be able to help with, hoping more than anything to get feedback from people who have been using it 3-6 months. I love the idea, but without more granular and accurate data, I find myself using it solely as a “whole house meter”, for which much less expensive options are available.

Hello @jtagen
I’m a user like you but I’ll try to answer some because I’ve had the same issues.
1: Always on is a rolling average for the last 48 hours or so of things that are always on. It’s actu much more complicated than that but beyond me to explain. Search for @kevin1 explanations and you get a solid answer.
2: No, you can not see which side of the phase a device is using. You can only view total watts on each leg in the app but not individually.
I posted a way to use breakers for identification by going backwards and turning off everything except for sense and internet breakers and turning one on at a time to identify. One only at a time.
3: when you get a detection go and either change the name to what you want or flip the slider off for “this is a guess”, sense will not change the name on you if either of these are done. Combine you can’t control. If you had something combined that is t correct, contact support at ASAP or use the report a problem button in the device page for them to help.
4: so if your oil burner is on but instead of that device showing on shows something like heat 1. Go to the device page and heat 1 (shows on but actually oil) and use the setting to report problem and click “this device is not on”. It may take a few times but will help with improving the detection on both devices.
5: I also have a 1500 watt heater that is used a lot that took 2 months to be detected. I also thought because of how it’s used and how easy it is for me to identify that sense would also. Just be patient on this one, it will come.
6: I’m over 2 months in and still ha e a lot of “other”. Detection rates are different for everyone. How many do you have now? If you have half a dozen or more then I’d say you are doing okay. I have had a very fast detection rate but it stopped and it’s been at least a week since I’ve seen one. This is normal as sometimes they come in waves. A few then a period of nothing. Patience is key.
7: I’m avoiding this one
8: This is where you come in and “Train” sense. Sense identified what it thought it was but is not always correct. You have to help identify when it’s calling a fan a fridge or A/C a pump. It takes some observation on your part but is part of the enjoyment for me. Also remember that sense can see a component of a device. So my fridge has 12 different electric components in it. So far sense has detected the ice/water dispenser, defrost element, Compressor and fridge light when the door opens (opens 700 times a month).
So each detection may be one component that you can choose to merge with other components of the same device or keep separate. Like my fridge I keep separate so I can see when the fridge is opened. It lets me know when someone is home and now I can even tell who it is by patterns.
So I haven’t been at it for 3-6 months but close.
Anytime you can’t find your answer here then support will be glad to help, just report a problem in app or shoot an email. I prefer email and thinks it’s a little faster.
Most of all is being patient. A lot of us had higher expectations than sense currently delivers. Things may start a little muddy but hang in there and be a part of the community for the best experience. If you can’t find what your looking for on here then you can PM me anytime and also the moderators are where to get the best help. The main moderator and the manager is a sense employee he can be a huge help. He is @RyanAtSense
Welcome to the community
And here are a couple links you may want to visit
Community Moderators

Understand your frustrations. I’ll throw you a few quick answers.

Long treatise on how Always On works here, though Sense may have tweaked the calculation since then.

Anything you do to give Sense guidance on when a device is going on or off really doesn’t have enough time accuracy or repetition to aid machine learning. But your negative feedback, once a device has been found (Device is not on) is helpful. Theoretically, Device1 is conflated with Device2 now, would also be useful if both were detected devices, but that option isn’t available yet.

Your suggestion that Sense allow you to “lock” a Model is an interesting one, but comes with lots of complications. It means that Sense would need to maintain individual house-based models, you wouldn’t get the benefits of (sometimes) greatly improved detection plus changes in your house, and changes in other models your house uses, could still “dislodge” your correct detections.

1 Like

Can you define “so much”? It recalculates every 48h, based on your baseline usage. The linked post above goes into some more detail on the calculation.

At this time, no, but feel free to add your opinion to this thread Track Use by Phase

From auto-combination/deletion/renaming, yes! Sense will only do those things if the device is marked as a guess or hasn’t been renamed from Sense’s initial detection. Find more info here: Device Detection Update: Device Updates!. If it’s still happening, regardless, shoot a message to Support.

I’d second @samwooly1’s suggestion to make use of the Device is Not On feature. You might also want to write into Support about that. Occasionally, they can do some work around conflation.

Sometimes the most obvious devices to the eye aren’t the most obvious to Sense, for a litany of reasons. Could be the time it’s used, the frequency of use, other devices that are commonly run alongside with it, something weird with the on or off transients, and more. Space heaters are a pretty good candidate for smart plugs though, especially because you gain the ability to remotely turn it off.

We’re always improving our models, so yes, I would certainly say it should improve.

Training isn’t impossible and we really hope to implement some sort of training functionality into the future, but it’s far from simple or obvious. “Hints” could certainly be useful, but correlation just wouldn’t be simple to parse out. Some quick reasons — devices overlap and mask each other, signatures look different in different contexts (including when turn on/off in rapid succession), nailing the actual start/stop of a device would be incredibly tough to do.

This kinda relates to the above. Stove is the simplest and most obvious one, with the variable draw, but even heating systems and refrigerators can look different depending on environmental factors and other concurrent devices. You can try deleting the “worst” models, or try merging all of the models together.

Ultimately, device detection can vary across homes for many reasons. 6 weeks is also fairly early in the process. Sense needs to see a lot of device cycles, and that takes time. Trust me, I feel your pain. There’s things in my home that haven’t been detected (dang 2nd leg of my water heater…), and that can be frustrating. But, I know firsthand that we’re improving all the time — I sit next to the folks who are working on those water heater models. :+1:

Hope that helps a bit! Don’t hesitate to reach out to Support which home-specific questions either. They can look at your data and really help to parse out particular issues in your home.


Even in my very straightforward panel (only 9 breakers) with 2 Sense devices things are still not stable or known after a couple of years. We are, I suspect, some further years away from rapid & reliable device identification. As the ML matures (and the database grows), so do the hardware integration options like adding per-circuit or device monitors like the TP-Link Smart Plugs (I haven’t tried that myself but it’s tempting). I think for some complex (large) circuits it is inevitable that more sophisticated hardware will be needed. This is why there are moves toward per-breaker monitoring solutions.

I had early enthusiasm for a very fine-grained per-device monitoring but what I eventually settled with (while awaiting inevitable upgrades!) is a fine-grained but macro view of usage. Big ticket items like a 4kW water heater eventually get identified well enough that you can create alerts and monitor usage and discover things you didn’t expect.

Maybe there should be a METAPHORS topic?: Imagine yourself as a listener (the Sense) and all your devices as books. In your mind there are people constantly reading aloud the books and you are listening (Fahrenheit 451!). Ideally all the people reading have different voices and they are reading different books. With serious concentration maybe you can follow along a single thread. Maybe a learning machine can actually follow all the threads in this case. This would be the hope for Sense. The reality is more like many of the readers have quite similar voices; many of the books are similar if not identical (certainly in the same language!); and they are all wandering around in a small house, sometimes reading the newspaper during breakfast and sometimes not, while you are confined to the sofa trying to follow along.

1 Like

Please explain what you mean by 2 sense devices. Do you only have 2 detected devices or do you have side by side monitors installed? Or something different?
Btw, enjoyed your analogy in the last paragraph. Had not heard it’s before but made a lot of sense to a layman.

Terminology mixup there, I intended to say “2 Sense monitors” as referred to in response to your post:

Always good to have one’s analogies’ mixed metaphors appreciated. Thanks.

I would add three more things to your really fun analogy.

  • Volume - some readers are much louder than other readers and some readers ramp wildly up and down in volume.
  • Speed - some readers are speaking at 300 words per minute, some are speaking at 2 words per minute (Sense has to look for “signals” in the range from 500msec - motor on, to several minutes - EV ramp)
  • And some readers are just quietly humming - constant draw devices.

I think how Sense does such a good job reporting, tagging and notifying about already detected devices is something like this.
While that detected device is still operating in the same room with all the different voices and sound levels, it works like a parent and a child.
Once I learned my children’s cries and voices, I would go to day care or school where there would be 30 wild kids yelling, crying and talking. I could walk in the room and without looking, know exactly where my child wasif they were making noise. My sensitivity to their sounds only got better with time. I think this is how Sense works also.

@kevin1 : And then there are the whisperers: “What? Watt! Speak up!” – those pesky low wattage LEDs.

@samwooly1 : Totally in agreement there and digging you, but, um, so what you’re saying is that 30 wild devices (=kids) is going to take 30 well-connected Senses (=parents).

Running with these analogies reminds me of thoughts I had when first watching the Sense identify (or not) the devices … if only the devices themselves could send out truly unique signals (kiddie screams). If nothing else you can resort to switching things on and off more frequently to encourage the identification. If devices could send a high frequency “I am on/off” then a Sense could become almost 100% accurate. Given that is not going to happen anytime soon (or ever) does it make sense to at least be able to tell Sense all the devices you have up front? There’s a big difference between walking into a room of 30 kids when you know your kid is there versus not.

Not that it would take 30 sense monitors because I have 2 kids and have no problem identifying each. The thing is that was only after detection ( I had to learn them first to recognize). My thought was not on detections but on accurate identification of each signature there after.
I’m pretty sure the signal going out from smart plugs to sense is exactly what is happening now. The integration seems to me to be like Sense uncovering it’s ears to that signal. The smart plugs are constantly sending out the signal from what I understand, waiting for a listening devi e to hear it.
Maybe I’m misunderstanding and sense is querying the plugs.

There’s a thread somewhere on the forum suggesting a list of appliances per home. In it, as I remember, there was also feedback from Sense that a list wouldn’t be all that useful to detection since there is also a unique home component to each signature (between home wiring differences plus background noise). The big benefit they identified was that a list would give them a better set of targets for model development as they move down the list.