By Dependence on electricity has become a way of life. Laptops, mobile phones, air conditioners, stoves, televisions, electric vehicles … you get the idea … our daily lives are using an ever increasing amount of electricity, but we tend not to think about how much energy each of these things use. At least until we get our electric bill. Out of sight, out of mind. But what if you could tap into the power of machine learning to help you monitor your energy use? To find out where it’s coming from and where it’s going? To potentially let you know if one of those devices is failing and in need of being replaced? Would you be willing to add that artificial intelligence to your home?
When you get a car that requires charging, whether it’s a plug-in hybrid or full blown EV, it always seems to send you down a path of wanting to know exactly where your energy use is coming from. When I got my first plug-in hybrid car about six years ago, I started down this path myself. And it only got more intense when I had solar installed and got my Tesla Model 3. About two years ago I installed a Sense whole home energy monitor, which not only helps me track how much energy my solar panels are generating, but also how much energy I’m using … and provides a picture of what devices are using that energy.
There are a bunch of home energy monitors on the market that you can use, but Sense caught my attention because of the machine learning aspect of the product. I had a chance to sit down and talk with Mike Phillips, the CEO of Sense, about how this type of system works and how it came about.
Matt Ferrell:
What is it that got you to start the company? What was your motivation behind it?
Mike Phillips:
It started from this broad notion of people want to be able to save money by saving energy in the house, but they just don’t have visibility into where it’s going. So that’s what we set out to attack.
Matt Ferrell:
There’s other systems like this out there. But the thing that drew me to Sense was the simplicity of the system, because most of these systems I was looking at were like, you have to put all these connections on every single line on your electrical box, but here was Sense, just put two clamps around your mains and the machine learning will take care of the rest. Machine learning seemed to be the secret sauce of Sense. Was that your basic approach that pulled this all together?
Mike Phillips:
Yeah. I also started by putting on these systems in my house that had the clamps on every single circuit and it just was too cumbersome. And even then, knowing power by circuit wasn’t really what I wanted. I wanted to know what the different devices were doing. So we started wondering, well, if we measure the power in a detailed enough way, can we figure it out just from the power signals? We’re making use of the fact that the different things in your home use power slightly differently. So if we measure well enough, can we tell the difference? And sure enough, it’s not a new idea, it’s called load disaggregation and there’ve been tons of research done on this, but what we found was it’s just much harder than most people thought it was.
Things worked in the lab, but not in the real world, which is how speech recognition was 30 years ago. I don’t know if you know or remember what that was like. Systems that would work in a lab but not the real world. And this is just as hard. And we realized, okay, first of all, we got to get really ambitious about the signals that we collect. So we have very high resolution data and then just the machine learning problem on top is a tough, tough problem, but we’ve been deep into it.
Matt Ferrell:
I’m not an expert. So my understanding is base level, but I know there’s basically two approaches you can take. There’s the, I think you even described it on your website, is there’s the supervised and the unsupervised paths towards teaching something, supervised being, showing a machine here’s a picture of a dog, a picture of a cat, and then you show them a whole bunch of pictures and you tell them when they got it right or wrong versus unsupervised where they’re figuring it out for themselves. Is that approach for unsupervised a little bit more challenging approach, but with more upside, I’m curious, why one approach over the other?
Mike Phillips:
Yeah. A great point. Look, it’s the case that supervised machine learning is much easier. So if you can get a nice label data that tells you this is a refrigerator and this is a toaster or whatnot, that’s great. The problem was we can’t do that here. In speech recognition or image processing, speech recognition, you can play audio to people, they can listen to it and say, “Oh, the person said this,” or you can show people pictures of cats and dogs, that was a cat, that was a dog. Here for the most part, you can’t do that because you show people electrical signals that they don’t even know. And a lot of our users want to help and it’s great to get their help, but they don’t know when the condensate pump of their furnace is running. They’re just not the so-called ground truth that we can do. So we had to go down the unsupervised learning approach.
Matt Ferrell:
As an end user, when it detects a device in my home and it says, “We think it’s a washing machine.” And then it says, “But it could be one of these other things,” I’m supposed to pick which one it is. Is that the training of the system or is there more to it than that? Is there something that you guys are doing at Sense to really shepherd it along?
Mike Phillips:
Yeah. There’s a lot that goes on behind the scenes and that one particular thing, where if we find something that we believe to be unique and we don’t know what it is, we can say, “We found something, it looks like maybe a motor,” and then 30% of people call this a blender, 20% people call it a coffee grinder. You go, “Oh, it must be my coffee grinder.” You type in coffee grinder, it’s good. It says that in your app. And then that feeds back to us.
So that is super helpful. But some of the other things behind the scenes that comes before that is we can’t even show you something in the app until we believe it to be unique, because if we just did this with your… We found your hallway light, a 60 watt incandescent in your hallway, and said, “Great, we found it,” but if you also happen to have a 60 watt light in your bathroom, we would get them confused. So we have a severely hard part of what we do, which is a uniqueness test, to say, “Have we found something and we modeled it correctly and do we believe it’s unique?” We have to do all that first before we say, “Oh, we found something, help us know what it is.”
Matt Ferrell:
Why do I sometimes see duplicates that show up? It will detect my refrigerator, and then six months later it detects my refrigerator. I’ve had that happen on occasion and I think I have understanding as to why that happens on occasion, but I was curious if you had any insight as to why that might happen.
Mike Phillips:
By the way, one thing I didn’t mention on the machine learning side that I think is important to point out, in addition to having to be unsupervised, in addition to having this uniqueness tests that we have to do, and wanting to do things in real time, you got to realize the signals are all on top of each other. So it’s like doing speech recognition with 30 people all talking at the same time. It’s not just one more layer of why this is hard. Now back to your question about why do you sometimes get duplicates. I mentioned that we have this uniqueness test to decide, is this a unique thing? Well, sometimes we get that wrong. And sometimes we think that what we see now is a different device than we saw before. This is actually quite active internally right now, we’re working on that. We know it’s a problem. So you should see better performance around that here coming up pretty soon.
Matt Ferrell:
That actually leads right into one of the things I was really excited to talk to you about, which was a few months ago, I read in one of your blog posts on your website about the challenges of detecting EVs as an EV owner, I am super interested in knowing how much energy is going into my car, how much it’s costing me. I was curious if you could expand on what you just mentioned as well as that as to the challenges around actually identifying devices, especially things complicated, like an EV.
Mike Phillips:
So first of all, we are super excited about EVs also, both in terms of be able to detect them and also be able to start to do things like load shifting. We can talk about that later. But it starts with be able to detect them and look, everyone thinks, and we do too, they should be super obvious. There are these great, big things in your home. We actually made life a bit harder for ourselves, but for a good reason. And that is, we wanted to be able to show you things real time in the app. There’s a big benefit in the app to be able to turn on your microwave, seeing turn it on right now and go, “Oh, that’s my microwave.” Same with the coffee grinder and so on. So we put a lot of work into making the app real time.
And a lot of what we do is focus on how do we make that real time? The reason that makes things difficult now for EVs is when they turn on, they actually don’t turn on in a big, clear way, right away. They’re very complicated devices inside the cars where the charger is, and they start by, they do a first little test to make sure everything’s connected. And then they ramp up in a little slow way and then they ramp up to some intermediate thing. So there’s this complicated dynamic of how these things ramp up and down when they start and even harder for us is, as the battery starts to get full, they start to gradually ramp down their charged rates. Sorry for all the excuses here, but what it means is trying to show what your EV is doing in real time is a tough, tough problem for us. We’re not giving up. We’ll increasingly be able to do these things, but still not perfectly, but it’s a major focus of our team to better and better do EVs over time.
Matt Ferrell:
Where is it happening? Where does the model actually live? Is it on the device, in my home or is it actually being processed out on a server?
Mike Phillips:
Yeah, a great question. At least that one has an easy answer. A lot of your questions have tough answers, that one’s an easy one to answer, which is what we call the runtime modeling. So the thing that all day long is saying, “Did my toaster just turn on?” That’s happening in the device, because we don’t want to have to wait and we don’t want to send up data up to the server for doing that.
But the thing that says, “Do you have a toaster?” And, “What does your toaster look like?” because it turns out to do this well enough. We have to not just have a generic model of toasters, which we do, but we also have to know specifically what your toaster looks like. In order to get that real time we see right now your toaster is turning on, we need a specific model for your toaster in your house. Sorry. That model learning happens server side. So what that means is we send enough data up to both drive the application, but also to then learn what your toaster looks like. And then we push those models back to the device to say, “Okay, keep looking for that toaster.”
Matt Ferrell:
One of the things I mentioned in the beginning is my motivation for it was to understand where my energy was going. I’m curious if what you see the average Sense user is saving on energy based on what they’re doing. If that makes sense.
Mike Phillips:
Look, we think even in energy savings, there’s a lot that we can do over time automatically to really point out where your potential savings are. And that was our first notion that we would just identify for you where there’s savings opportunity and automatically tell you. But the first thing we found was that most of the savings are not well-defined categories, like utilities think it’s around light bulbs and refrigerators. Well, that’s done. They’re all pretty good now. And meanwhile, there’s this big chunk that we just call energy hogs, which is random stuff that if you knew about it you’d go, “Oh, I don’t need to do that. I can fix that thing.” It’s a really broad range. It ranges from you leaving the roof coils on that melt your ice dams and you leave them all on all summer.
Or you didn’t realize a dehumidifier uses a third of your power or there’s something wrong with the setting of your heat pump system or … the list just goes on and on. And because it’s so broad, we realized, well, we can’t automate all that. So let’s start by giving consumers just visibility. And then hopefully they’ll be like you and actively use the app and the functionality to go track these things down. So that’s been the focus just based on the relationships being so broad. And what we’re finding is if you’re really active doing that, you can get 15% or so savings in most homes, but only about half of the people do that. So if you average across everyone, it’s half that.
Matt Ferrell:
I liked your examples because that actually happened to me. I had a dehumidifier that was on a smart system in my garage and something had happened to the smart outlet and it wasn’t doing what it was supposed to be doing, the dehumidifier was running 24 hours a day and I noticed in Sense that there was this huge uptick in energy use and it helped me track down exactly what had happened.
Mike Phillips:
And it really is that, that if you have no visibility… Look, if water was running in your basement, you’d know. It’d be clucking. If electricity is leaking and it doesn’t leak in the same way, there’s puddle on the floor in the basement, but it leaks in these other ways. You have no way to know. You just have no visibility. So just giving people this direct real time visibility where, “Oh, what’s this?” They get savings.
Matt Ferrell:
I don’t think you’ll necessarily be able to answer this, but it’s because it’s probably about future plans, but have you guys considered deeper integrations into other devices? We’re talking about load balancing. So let’s say my solar system is over producing electricity. It can automatically turn on my wall charger, the wall connector to start my car charging. Have you guys considered something like that?
Mike Phillips:
Yeah, absolutely. And we already have a few things that are live today that don’t go as far as what you just said, but we’ve already done integrations with if this and that, we’ve done integrations with TP-link plugs, with MIMO plugs. And so we are already using those where you can… We use that to get both a direct measurement of what’s going on in the application and in the app, you can control it. So if you see in the Sense app that you left your garage lights on, and if it’s on the Waymo plug, you can click there to turn off. But the next step of starting to automate these things very much is where we’re heading.
So in fact, your case is exactly… So in my house I have solar panels. I have two EVs and I have a very early rigged up version of Sense where it’s actually controlling my EVs to only charge when solar is over producing.
Matt Ferrell:
That’s awesome.
Mike Phillips:
If not, it goes and finds the marginal carbon intensity. Now we haven’t productized that yet just due to priorities. That kind of thing is very much what we’re excited about. We’re experimenting and absolutely is the future.
Matt Ferrell:
It’s one of the reasons I was excited about the Sense when I got it, was it felt like something that will definitely grow and evolve longterm. So I felt like I was investing a little bit into the future of the product when I bought it.
Mike Phillips:
A lot of these things that we want to be able to do, you were just talking about, we couldn’t really do until we start to get data and we couldn’t start to get data until we got these things out there. So there is this a virtuous loop now where we’ve got ton of data flowing in, ton of engaged consumers like yourselves and just more to be done. And now we’re limited by how fast we can get this stuff out.
Matt Ferrell:
So where do you see, not just Sense, but in the broader picture of this type of system, where do you see this going in five or 10 years? Where’s the end game for you on this?
Mike Phillips:
Well, look, we really do view that your home itself needs to become smarter in ways that it hasn’t before. Most people, when they talk about smart home, they’re talking about automated lights or their entertainment system automating, which are all fine. I’m not trying to knock those, but that’s to me just fundamentally different than your home itself or the core systems in your home becoming smart for efficiency, for reliability and safety. Things like that I think is what’s coming and has real, real benefit to consumers from a saving money, from a making their homes more reliable and more healthy.
So I think that’s where the opportunity is. It’s like what happened in cars. Cars, 50 years ago were independent mechanical systems and carburetors and points that you had to go and tune up every couple of weekends. And since then they became incredibly instrumented and automated resulting in huge efficiency gains and huge reliability gains. And the same should be true of homes. Homes are harder because they are built out of independent systems versus built by one as a system itself. But it doesn’t mean that the same principles don’t apply that by making those things smart, we can make these big gains around efficiency, reliability, and convenience.
Matt Ferrell:
Whenever I’d make smart home videos, one of the things I’m always telling people is, it’s more than making your lights flash green and red by talking to your voice assistant. Smart homes have so much more value to our everyday lives than just that.
Mike Phillips:
And I think one that’s been under explored even by us also, is even things like health of the occupants. Things like air quality, do you have a Radon fan and is it running at the right rate? Do you have a ventilation system, and is it running appropriately? These things will start to be… There’s no reason your home can’t know things like the indoor air quality and adjust as needed to make sure you’re healthy. And those are big, big benefits to people, obviously. So one of the other things that we’re excited about and working on that’s going to result in even more doing what we can do today, is just getting more details about what’s actually in your home, because right now what we’re doing on the machine learning side, on the unsupervised side is we get these data signals and a little bit of hints from the consumer, but then we have to figure out everything about what’s going on in the home.
If we start to work with home builders, for example, where they already know what dishwasher is there, what furnace is there, what hot water heater is there and knowing half time, what systems are there, how they’re supposed to be working, and now our job is to make sure that they are working as expected. There’s a lot more value that we can provide versus our current handle every single house out there case.
Matt Ferrell:
So it’d be more curated because you’d know exactly what’s going on ahead of time.
Mike Phillips:
That’s right. So that kind of curated, we think also becomes really important for things like performance monitoring over time. If you have an air conditioner, maybe it was super efficient when you got it installed. Is it still efficient three years later? Well, if we already know we’ve been watching it the whole time, we can say, “Yeah, it’s working fine,” or, “No, it looks like it’s been degrading, maybe the refrigerants leaked out. Maybe you should have it looked at.”
Matt Ferrell:
Well, thank you so much for taking the time to talk to me today.
Mike Phillips:
You’re welcome. It was great to meet you.
I’d like to thank Mike Phillips for taking the time to talk to me. It’s companies like Sense that are helping to raise awareness of not only where our energy use is going, but building the foundation of AI systems that can eventually let us know when devices are breaking down and help keep our homes safe and healthy. With knowledge comes power … yeah, bad dad joke.
Comments