When millions of smart meters rolled out across the country at the turn of the last decade, many people hoped it would create the backbone of a digital grid.
Today, you’ll find few who think meters lived up to expectations. One survey found only 3% of advanced meters supported by the 2009 stimulus bill brought customer savings.
Mike Phillips, the CEO of Sense, is still bullish on the role of advanced meters for grid intelligence and bill savings. But as utilities start a new wave of rollouts to replace old technology, he worries they aren’t investing in the right architecture.
“Most people think of meters just as data collection devices. You have to start to change that mindset, and once you start to think of this as a distributed platform — not just a data collection device — this entire world of making use of machine learning at the edge starts to get opened up,” said Philips.
This week: a conversation with Mike Phillips on what AMI 2.0 should look like. Past deployments of smart meters didn’t bring the intelligence promised. How do we avoid the same outcome?
The Greenhouse Gas Reduction Fund will provide $27 billion for clean energy projects nationwide, potentially mobilizing up to $150 billion in public and private capital. Join Latitude Media and Banyan Infrastructure on July 18th for an in-depth discussion on how we can deploy these billions with the highest impact. Register for free here.
Make sure to listen to our new podcast, Political Climate — an insider’s view on the most pressing policy questions in energy and climate. Tune in every other Friday for the latest takes from hosts Julia Pyper, Emily Domenech, and Brandon Hurlbut. Available on Apple, Spotify, or wherever you get your podcasts.
Transcript
Speaker 1: Hey Chat, are you there?
Chatbot: Hey. Yes, I’m here. What’s up?
Stephen Lacey: Back in May, OpenAI released GPT-4.0 And it was a big leap forward for voice and video interaction with a chatbot.
Speaker 1: Hey, let’s have some fun.
Stephen Lacey: There are a bunch of demos from OpenAI like this one where the chatbot is mimicking human emotion in ways we really haven’t seen outside science fiction.
Speaker 1: I’d like you to be super sarcastic. Everything you say from now on is just going to be dripping in sarcasm. How does that sound?
Chatbot: Oh, that sounds just amazing. Being sarcastic all the time isn’t exhausting or anything. I’m so excited for this. Note the sarcasm. Let’s get this party started or whatever.
Stephen Lacey: Generative artificial intelligence has come a long way in just the last two years. Built of course on many decades of research before.
Mike Phillips: A lot of the stuff that is being leveraged today in the AI world comes out of language models. And we were doing very early days of speech recognition and language models, back 20, 25 years ago,
Stephen Lacey: From the late eighties into the mid-nineties, Mike Phillips was a research scientist at MIT. He worked on speech recognition and natural language processing. Mike founded multiple companies in the speech recognition space, one of which built the voice assistant for the Samsung Galaxy phone.
Mike Phillips: In the early days of these voice assistants, smartphones were just coming out, we realized that the phones were getting connected to the data networks and they were going to start to be more like the internet and then therefore become people’s personal information communication entertainment device. And we also realized, well, you shouldn’t have to type on a little keypad, you should be able to just talk to these things.
Stephen Lacey: After selling both of those companies, Mike turned to energy and he asked what platform can unlock savings and decarbonization that is analogous to the smartphone. He co-founded Sense, which created an energy monitor installed in an electrical panel to give real-time information on every device in a home.
Mike Phillips: Look, if smart devices were fully enabled and every device out in your home told us what it was doing, we’d be happy to use that. But it’s far from being the case, and we realized that if we could measure the power in a detailed enough way, could we figure out what’s going on just from the power signatures. So you can start to see where the similarities come in that we used to do speech recognition on audio signals, we are now doing electrical device identification based on power signals.
Stephen Lacey: So the company was started in 2013, and then you had years of time deploying devices in the field, and then you discovered that suddenly you had a lot of visibility beyond the home. What kind of view of the grid did disaggregation uncover?
Mike Phillips: The first thing we did was collect some signals and realize, “Oh look, we can’t do this unless we have high resolution data.” And our initial view is, we had previously written software for smartphones, we’d write software for smart meters and we’d be all set. Quickly found out though that the existing smart meters just did not have the data we needed. So that led us down a long path that many of you know that we started to build these little orange boxes that go inside electrical panels and collect data at super high resolution, up to a million samples a second in the little orange box. So that high resolution data was the key to unlocking what happens in the home. And now to your point, we realize sometime afterwards that that same technical capability, so high resolution data, edge computing, real-time networking, that we use to interact with consumers on what’s happening in their home, we can look the other direction and we can see what the grid is doing in real time from the edge.
Stephen Lacey: Are you saying that the first class of smart meters weren’t actually as smart as we think they are?
Mike Phillips: Hey, you said it not me.
Stephen Lacey: Sense is now focused on putting its machine learning technology into new generations of smart meters. Mike is bullish on the role of advanced meters for grid intelligence, but as utilities start a new wave of roll-outs, he worries they aren’t investing in the right architecture.
Mike Phillips: Most people think of meters just as data collection devices. You have to start to change that mindset. And once you start to think of this as a distributed platform, not just a data collection device, this entire world of making use of the data at the edge and AI machine learning at the edge, starts to get opened up.
Stephen Lacey: This is The Carbon Copy. I’m Stephen Lacey. This week a conversation with Sense CEO Mike Phillips, on what AMI 2.0 could and should look like. Past deployments of smart meters didn’t bring the intelligence promised. How do we avoid the same outcome?
So as you said, the first generation of smart meters were just seen as data collection devices, and most people who look at the rollout of AMI 1.0 would say it did not live up to expectations. Ultimately, what do you think the limitations of the first rollout of smart meters was? Was it just that the devices themselves weren’t powerful enough, so it was the wrong infrastructure? Was it just as a data collection device, the utilities didn’t know what to do with that data and they had an inability to use it properly? Give us your broad view on what went wrong with AMI 1.0, which I think most people think wasn’t that successful.
Mike Phillips: Yeah look, I think it was basically the wrong architecture. Look, it made sense back in 2008 or whenever that happened, but it hasn’t transitioned since then. So, and by wrong architecture, look back to my telecom example. Imagine if the way Google Maps on your phone worked is your phone would collect 15 minute interval data of your location, send it up in batches to your telecom provider, who would then make it available through what… gray button instead of green button or something like that. And then applications like Google Maps could get that data a day later and then do something with it. Well, what would Google Maps on your phone be? It would be like a static map and maybe a monthly historical report of your traffic on your route to work and maybe compare your neighbors how you drive compared to your neighbors. But would you use that app? You might use it every now and then, but you would not engage in it in the way you do with Google Maps.
And what I’ve just described exactly matches AMI 1.0 architecture, low resolution data sent up to the service provider, made available later on. And look, there’s some things you can do with that, but you just can’t have a real time consumer facing app and then you can’t see the grid in real time from the edge either.
Stephen Lacey: So we’re at a point now where we’re rolling out a lot of new meters. We’ve had a decade to develop the architecture. What is your sense now for the current architecture that utilities are reinstalling?
Mike Phillips: We’re crossing the threshold now. Look, there are meters available today on the market that do all the stuff we want to be able to do and unlocks all this potential. And by unlocking potential, let me circle back to something that you all are quite into, and everyone’s talking about AI for the grid. Well look, we know AI is driven mainly by machine learning based approaches these days, and that’s mainly driven by data. And if you don’t have the right data, you’re kind of stuck. And there’s a lot of talk in the industry about grid edge intelligence, but people are mainly talking about taking data from the edge, this 15 minute interval data, processing it in the cloud. And there’s some things you can do with that, I’m not denying that, but to fully unlock the potential for AI for the grid, we need the right data.
And this is what we learned long ago at Sense. To have a real time consumer experience – by real time, I mean you turn on your microwave and it shows up in the app a second later – to have that kind of experience, you can only do it with high resolution data. What we’ve learned since then is that same high resolution data let’s us see the grid from the edge, let’s us see transformers arcing, let’s us see vegetation hanging power lines in real time. That only happens through high resolution data. So sorry to keep going on about this, but the number one step is to get the right data in the meters themselves and be able to process it there.
Stephen Lacey: Well, I know you’ve been somewhat disappointed with how some utilities are reinvesting in their metering networks. Is that just because they’re essentially investing in a technology that is not much better than the first generation?
Mike Phillips: Yeah, we are certainly worrying about this, that there’s a big opportunity now that, like I say, there are meters available today that can provide a lot of headroom for what happens in the future. And there’s still decisions being made for a previous generations of meters. So we’re trying to help utilities and help others. In fact, we’re just publishing up on our website an AMI buyer’s guide. And look, we don’t have the full picture of all the things that a utility needs to consider for AMI deployments, but we do know a lot about how to deploy data intelligence AI at the edge of the grid. And so we’re putting together the point of view of all the things you need; high resolution data, enough computation, and the ability to have real-time networking are the three things you need, but you got to pay attention to the details and get it right.
Stephen Lacey: So give us a sense for what the generation of meters that you’re working on can give us in terms of data versus the traditional meters.
Mike Phillips: The most basic thing is to get high resolution data of voltage and current in the grid. Let’s us see what’s happening in the home, let’s just see the grid. And by that, sorry to go into some of the techie details, but this means continuous sampling of these voltage and current waveforms of at least… there’s meters on the market that are doing that at 15,000 times per second. So that’s 50 million times more data than AMI 1.0. And again, for those of you who get scared we don’t even know how to deal with AMI 1.0 data. We don’t mean you send all that up to the network, you’re processing at the edge. So with 50 million times more data, we see the first a hundred harmonics of signals at the edge, and that is what we need to see what’s happening inside the home, see what’s happening in the grid.
So those meters are available today. Utilities can go get rate cases approved and be able to deploy them, and it’s happening in pockets right now. And look, I know there’s been a lot of talk, and we are involved in this too, about what comes after that also. So we are starting to work with meter makers to go all the way up to a megahertz of sampling. So there’s progression and things are happening in this world, but there’s meters on the market today that are sufficient for doing what we do, and we want to make sure that utilities are fully equipped with the information about that so they don’t go buy the previous generation.
Stephen Lacey: So GPU chip maker Nvidia is working with a company like Utilidata to add a lot more processing power into meters. Do we need a lot more processing power to process this amount of data that you’re talking about?
Mike Phillips: Yeah, so we’re paying a lot of attention to that too. And look, I will never argue against more processing power. This is the lessons we’ve learned, more data, more processing is what’s been driving these revolutions in AI. We’ve actually been more concerned about the access to the data. We know what we can do with more data. Since we’ve had this kind of high resolution data all the way up to a megahertz for years now, we know what we see in that data and we know how to use it. On the computation side, sure, we want more computations. GPUs in meters, when they become practical we will absolutely make use of. I’ve been a little less concerned about that because for computation, you can always be very efficient about your algorithms and you can make trade-offs from the algorithms. The reason we’re more concerned about the data is if you don’t have the data, you don’t have the data. There’s just nothing you can do if you don’t have the data.
Stephen Lacey: So let’s apply this to actual use cases out in the power sector. What were some of the successful use cases in AMI 1.0, and what do you think the new use cases for this next generation of smart meters are?
Mike Phillips: So let’s start with the consumer side. On the consumer side with AMI 1.0, you could provide a next day view of how power was used in your home yesterday. So this was used in these various portals that utilities provide, that give people some insights into how their bill’s tracking, which is a big leap compared to waiting until the end of the month before you have any idea what’s going on in your home. But look, frankly, not a lot of people use those portals. I think there’s pretty well-defined metrics that these haven’t got a lot of use.
Once you get to this new architecture, applications like Sense become possible where we can provide consumers with a real-time, detailed view of what’s happening in their home. And this becomes super relevant for energy efficiency. People can see what’s going on in their home and track down energy hogs, we call them, in the home. It also is having a big impact on people’s participation in demand flexibility, load flexibility. How do you deploy time of use rates or demand charges if you can’t let the users see what’s happening in real time in their homes.
And then we’re also using this high resolution detailed view of what’s happening in homes for helping with electrification, find the homes that are best candidates for heat pumps and so on.
Stephen Lacey: Okay, so that’s the consumer side. What about the grid side?
Mike Phillips: Well, this is actually a surprise for us. We did not actually realize that the edge of the grid is largely unmonitored. And again, the technical capabilities in these new meters, high resolution data, edge computing, real-time networking, that we use to see inside the home, we can see the other direction, to the extent that when things happen on the grid, if it shows up as power fluctuations in the home, we see at the edge. And when you just look at power once a second, which is what was done even in the internals of AMI 1.0, you don’t see that much. You see big level shifts whether conservation voltage reduction is going too far and stuff, you could do that with AMI 1.0. As you get to high resolution data, even tiny little arc flashes that are happening on the secondary side, so transformers that are arcing, problems in the meter socket and stuff like this, we see all that in detail and things like vegetation hitting primary lines. We see the entire all the way up through the distribution system and all the way into the transmission systems. When things happen, we see it at the edge.
Stephen Lacey: So what are the consequences of getting these meter roll-outs wrong? These roll-outs tend to happen around 15 year cycles. So the first cycle was between 2008, 2011. And so now we have a bunch of utilities that need to reinvest as we’ve said. What are the consequences for operations’ wasted potential if we get it wrong?
Mike Phillips: Yeah, it’s fairly obvious that the energy transition is going to happen over the lifespan of the meters. And we don’t know all the things that have to happen. We don’t know what the uptake of EVs is going to be, we don’t know what the generation mix is going to be, we don’t know the effect of data centers on mode. A lot of worries about those things. And look, by having the right infrastructure at the edge such that this can now be software, this gives us the flexibility to adapt over time. We know the things we need to do today; efficiency, demand flexibility and so on, and vegetation management and others on the grid. So we know the things we need to do today, but we can’t predict all the use cases. And again, as you guys are covering the benefits of AI in the grid. But we do know that by having the right infrastructure and the right data, that we can do this as upgradable software, gives us the flexibility to upgrade this over time. So the lifespan of these meters are going to be during that time of the energy transition, so we really need to get the right infrastructure out there now.
Stephen Lacey: And I know that you want to be careful about calling out other specific companies, but there are a lot of utilities who listen to this show, people who may actually be involved in metering programs. If they did it wrong in your view, what would be the technological choice that they made that was wrong or not sufficient?
Mike Phillips: There are just these three main topics that I’ve been talking about, which is getting access to the right data, so high resolution waveform data. It’s well known how to do that. It’s mature technology, we know how to do that. Having enough computation so that you can run software in the meters, have it upgradable and so on, and have the ability to get onto real-time networks of various sorts. So those are the three main technical capabilities. There also then has to be layered on top, policies and software to protect consumer privacy and so on.
So these are the elements of what’s needed. And I think everyone’s talking about these things. So this is not a foreign notion and people are talking about AMI 2.0. My concern is when you get to the details, paying attention to really the details around the computation available, around the data rates and so on. And I know these seem like dry technical topics, but they’re what matter because if you don’t have the right technical capabilities, you’re kind of stuck.
Stephen Lacey: So you’ve been involved in machine learning, natural language processing for decades now. What are the biggest advancements that you think are going to push the energy transition forward?
Mike Phillips: So let me tell you a little story. Again, we came out of this natural language world. We used to train language models on a million words of text. They could predict the next word based on the last two or three words. That was fine for, “What’s the weather in Boston tomorrow?” But what’s happened since then is like GPT is trained on a million times more data. When you add a million times more data, amazing things happen. Now these things can predict the next word based on the last 30,000 words. And now you have context, meaning, emotion, domain knowledge, all built into that thing.
So the lesson from that world, is the power of data and the techniques that are layered on top of these to find amazing patterns within this data. So that’s what we are leveraging here and be able to make use of the huge amount of data we can now process at the edge. Again, we can’t just collect it, we have to process at the edge, combined with the latest techniques. And we’re getting increasingly plugged into all the latest techniques that have come from this space.
Stephen Lacey: So the field of machine learning is advancing rapidly. How are you making sure that you’re up on the latest machine learning techniques?
Mike Phillips: This particular problem of how to decide on devices in the home just from power signals, has got some very particular details that we’ve been deep into, but we also are seeing the parallels of what’s happened in audio processing and so on. So we actually have someone coming from… that was leading these efforts at Google, is going to be joining us shortly as an advisor for the company. So he’s really helping us take all the latest techniques of what’s happening in the big tech companies and apply it to this particular problem.
Stephen Lacey: And you also have a new generation of meter that’s coming out as well, right?
Mike Phillips: Yeah. Look, as I said earlier, the meters on the market today are great. They support… Like I say, we can process up to a hundred harmonics of the power signal, so it supports the use cases that we see out there today. But technology is always advancing. So we are working with some of the meter makers to have the next generations of meters that have even higher sampling rates than that, going all the way up to megahertz. And that’s unlocking some more use cases around arc fault detection. When there’s an arcing on a transformer or something like that, it shows up in these high frequencies. So if you start to worry about even a broader set of use cases, continue to push the envelope on data and as you mentioned, more and more computation.
So it’s a tricky topic. This happens in every technology field. We don’t think you should wait because the latest generation is great. So if you’re making a meter decision now and you need to replace your meters, there are meters on the market that do what we need to do. But the next step is coming here shortly.
Stephen Lacey: Mike Phillips, co-founder and CEO of Sense. Thank you so much.
Mike Phillips: Thank you, Stephen.
Stephen Lacey: And that’s going to do it for the show. If you want to learn more about AI and smart meters, we’re covering that subject in particular and AI generally at latitudemedia.com. You can subscribe to our newsletter there. You’ll get all our editorial coverage in your inbox in the middle of the week, and then a roundup of our podcasts at the end of the week. We also have a great new game that people are responding to in our regular newsletter, so check that out.
You’ll find transcripts of these episodes on our website as well. And of course, if you’re over there listening on Spotify or Apple, go ahead and give us a rating and review. This show is produced by me, Sean Marquand is our technical director, Anne Bailey is our senior editor. Sean composed our theme song. And Latitude Media is supported by Prelude Ventures. If you want to learn more about Prelude’s investment strategy in climate tech, go to Preludeventures.com.
Thank you so much for being here. I’m Stephen Lacey. This is The Carbon Copy.
