CEATI 2026 | AI-Powered LifecycleOptimization for Modern Utilities

So our journey at Dominion with AI started several years ago, um, and we really saw it as a, a very, uh, dovetailing technology along with the expansion of our drone program.

Um, we recognized early on that the amount of video and imagery that was gonna be coming back from the ex- the expansion of that program was gonna greatly increase the amount of data that we were getting in, and it was gonna very quickly exceed the capability of humans to process it all in a meaningful way and, and make good decisions with it.

Uh, so we, we looked to, at that time, to find a partner in the AI space who we could work and learn with.

Uh, at the time, a lot of the solutions out there were very early on, focusing heavily on, "Hey, we can just automate all of your inspections.

We'll tell you where all your defects are.

No humans needed.

Go for it." Um, as a utility person who's gonna get fussed at by leadership if we miss something and then a line locks out, we weren't comfortable with that.

So we looked for something and some opportunities where we could start small and grow from there and build that partnership in a way that builds trust, not only from the utility to the, the company, but from the company to better understand what we as a utility would need.

So that way, as they saw things out in the industry, they could bring them to us and say, "Hey, maybe this is something that we've seen elsewhere that would be useful for you." It, it was really something that we saw as a long-term partnership, not just a single product or subscription service.

Uh, and so with that, we were using a lot of our existing imagery while that drone program came online, uh, and that came in the form of helicopter corridor imagery that I'm sure a lot of our, you know, utility peers have, have similar out there.

Uh, and we started out, one of the things we started with was, um, was a very basic classification problem of...

Is anybody out here, who has polymer insulators on their transmission system? Who has a maintenance plan to replace those before they start falling out of the sky?

[laughing]

Who knows exactly where every single polymer on your system is? 'Cause I sure as heck don't.

Well, I didn't at the time.

Uh, and so that was kind of our starting program, is we had a eight-year-old spreadsheet that was done by a predecessor of mine.

It had never been touched in that whole time, and so we needed something that could give us some better information as we saw more and more transition in the employees, um, that were running these programs to get them up to speed faster without having to do a ton of research.

Um, and that's an example of, you know, what we see here is a, a perfect example of that human-in-the-loop training that we went through to start training some of these early models on these are polymers, these are porcelain, these are glass, uh, et cetera, so we could start building that up.

So like I said, we started small.

We started with those insulators.

Um, and with that, we ended up with a, a pretty good result.

We had a eighty-six percent F1 score overall on being able to classify polymer, porcelain, glass, or mixed usage, where you've got maybe a multi-circuit structure that has polymer on one side and glass on the other.

Um, and so from there, we kind of started branching out.

We were having r- a lot of success in that classification space, and we were wanting to test the boundaries of what's, what's possible.

What can we do? How can we build out these datasets that we, we think we have, but they're probably stale? Um, they're not something that is well-maintained, um, and they're something that are a lot of times legacy data that's out there, uh, and how can we make that more accurate? And so the next one, what we went with was, uh, was bird perch deterrents, um, to try and reduce the, you know, equipment that we install out there on structures to reduce or mitigate the risk of, of bird streamers causing operations.

So our journey at Dominion with AI started several years ago, um, and we really saw it as a, a very, uh, dovetailing technology along with the expansion of our drone program.

Um, we recognized early on that the amount of video and imagery that was gonna be coming back from the ex- the expansion of that program was gonna greatly increase the amount of data that we were getting in, and it was gonna very quickly exceed the capability of humans to process it all in a meaningful way and, and make good decisions with it.

Uh, so we, we looked to, at that time, to find a partner in the AI space who we could work and learn with.

Uh, at the time, a lot of the solutions out there were very early on, focusing heavily on, "Hey, we can just automate all of your inspections.

We'll tell you where all your defects are.

No humans needed.

Go for it." Um, as a utility person who's gonna get fussed at by leadership if we miss something and then a line locks out, we weren't comfortable with that.

So we looked for something and some opportunities where we could start small and grow from there and build that partnership in a way that builds trust, not only from the utility to the, the company, but from the company to better understand what we as a utility would need.

So that way, as they saw things out in the industry, they could bring them to us and say, "Hey, maybe this is something that we've seen elsewhere that would be useful for you." It, it was really something that we saw as a long-term partnership, not just a single product or subscription service.

Uh, and so with that, we were using a lot of our existing imagery while that drone program came online, uh, and that came in the form of helicopter corridor imagery that I'm sure a lot of our, you know, utility peers have, have similar out there.

Uh, and we started out, one of the things we started with was, um, was a very basic classification problem of...

Is anybody out here, who has polymer insulators on their transmission system? Who has a maintenance plan to replace those before they start falling out of the sky?

[laughing]

Who knows exactly where every single polymer on your system is? 'Cause I sure as heck don't.

Well, I didn't at the time.

Uh, and so that was kind of our starting program, is we had a eight-year-old spreadsheet that was done by a predecessor of mine.

It had never been touched in that whole time, and so we needed something that could give us some better information as we saw more and more transition in the employees, um, that were running these programs to get them up to speed faster without having to do a ton of research.

Um, and that's an example of, you know, what we see here is a, a perfect example of that human-in-the-loop training that we went through to start training some of these early models on these are polymers, these are porcelain, these are glass, uh, et cetera, so we could start building that up.

So like I said, we started small.

We started with those insulators.

Um, and with that, we ended up with a, a pretty good result.

We had a eighty-six percent F1 score overall on being able to classify polymer, porcelain, glass, or mixed usage, where you've got maybe a multi-circuit structure that has polymer on one side and glass on the other.

Um, and so from there, we kind of started branching out.

We were having r- a lot of success in that classification space, and we were wanting to test the boundaries of what's, what's possible.

What can we do? How can we build out these datasets that we, we think we have, but they're probably stale? Um, they're not something that is well-maintained, um, and they're something that are a lot of times legacy data that's out there, uh, and how can we make that more accurate? And so the next one, what we went with was, uh, was bird perch deterrents, um, to try and reduce the, you know, equipment that we install out there on structures to reduce or mitigate the risk of, of bird streamers causing operations.