How does AI reshape the energy grid?
Powering the world requires an enormous amount of field infrastructure. In the U.S. alone, this includes roughly 180 million poles strung together by more than 5 million miles of conductor cable, a sweeping and interconnected grid that has expanded and evolved over time. Understanding the state of this grid is essential in order to modernize it effectively, but it’s a remarkably difficult and expensive thing to do. Now AI is changing that. To learn how, we welcome Josh Wepman, Leidos Vice President and CTO for Commercial Energy Solutions. Wepman’s team has created an AI-powered solution called Infrastructure Insight that gives utility companies a faster, cheaper, and better way to take inventory of the energy grid.
Q: First, why is it so important to modernize the energy grid?
Wepman: To make it cleaner, more reliable, and more affordable. We need to build a 21st century energy delivery and management system, but the only way to do that is to get our arms around what we’ve already built over the last hundred years to figure out the right way to change it. We can use AI to do that. We can use the cloud to do that. We can use processing and machine learning to do that. But as long as the inputs are garbage, the outputs won’t be useful, and it will reduce progress. What we need is a more effective way to understand the grid as it is today so we can tackle the most exciting challenges of the future: decarbonizing the planet, democratizing energy production, and letting everybody have the relationship with energy they’d like to have.
, CTO, Commercial EnergyWe all want clean energy. We all want wind, solar, and distributed energy, and we all want cheap and green and reliable. But this is fundamentally impossible without understanding the current state of the grid.
Q: What makes it so difficult for utility companies to account for their own field infrastructure, and why is it so important for this to change?
Wepman: Power companies have massive spatial infrastructure out in the field. They’ve got lines, poles, and other equipment spanning huge geographies. As you think about the evolution of the energy marketplace, you can think of it in two dimensions. If you think of an x-axis as a shift from analog to digital, all of those assets across all of this infrastructure are moving from analog assets to digitally empowered assets. They have microprocessors. They’ve got communications assets to tell us about their state and performance. We’re getting data rich, and we know much more now than we’ve ever known before.
If you think about the y-axis, we’re moving from a procedural operating model to a more analytical and data driven model. We’ve always run the power grid based on a whole bunch of assumptions about what’s happening out there. We run calculations and we try to estimate as best we can, but it’s always been very procedural, not very data-driven. What to build, where to build it, how big should it be, what risk does it represent over the next 20 years; it all depended on a narrow and generally certain view of the future. Now that we’ve gone across that x-axis from low data to high data, we’re beginning to make more progress migrating from procedural grid operations to data-driven analytical modes.
One of the things that’s really important about making that leap from procedural to analytical is better input data. The key challenge that our customers face is that an ‘approximate’ understanding of their field infrastructure has always been good enough for their procedural analyses. They’ve always worked in rough approximations. But now that they’ve got all this measurement data, actually knowing what is out in the field, where it is, and how it relates to each other is more important than ever before.
Q: How does AI now help us do that?
Wepman: This is what’s called a conflation program. Conflation means I’m reconciling what’s in my digital system with what’s really out in the field. What we recognized was that if we flew LIDAR missions to find out where the macro infrastructure are, we can use machine learning and feature extraction to say ‘these are poles and these are lines.’ And we’re talking about millions of poles. So for the first time ever we actually know where all these poles are in a consistent, latitude-longitude scheme. And for a great deal of these locations there are images from Google Street View, so now we can query Google Street View, and instead of trying to train machines to look absolutely everywhere for the small assets—the electrically connected devices that live on each individual pole and are important to actually delivering quality power to your home—we now know precisely where to look. That reduces the problem space dramatically.
We were able to take our data science and engineering prowess in this company, our electrical engineering design prowess, and our mission software prowess, and put all those things together. We took these images, used engineers to identify poles, cross arms, transformers, fuses, reclosers, and street lights. And we literally brought our engineers together and labeled thousands of example images to establish ground truth. And we had our data science and engineering team use TensorFlow to build training images to show what all of these assets are. We trained and we tested, and we trained and we tested to validate. And what we found was that we became very, very good at teaching machine vision and convolutional neural networks to take in millions of images of poles and produce inventories of those poles without ever sending a person out into the field.
Q: When it came to working with AI tools to develop this solution, what was your critical success factor?
Wepman: Intellectual curiosity. We said ‘philosophically, this ought to work. We ought to be able to train the images. We ought to be able to train a classifier.’ Our lead developer spent the better part of a weekend just being curious. What he found was there are a thousand ways to do this, but boy do a lot of people rave about TensorFlow and how easy it is to get started. So he implemented it, wrote a little bit of Python glue code, and tried it out. He labeled a few images, processed them, and found that it worked really well. A shockingly low number of inputs produced surprisingly good outputs. If you’re a developer, if you understand how to read code, and you’re interested in trying things out, the friction of getting started and becoming dangerous is shockingly low. There’s nothing about what we did that is fundamentally not solvable by intellectually curious people, which means we have to continue to push and evolve.
Q: Looking forward, what excites you most about the AI solution you’ve created?
Wepman: The opportunity space is vast. If data science and engineering is going to solve the problems of the future, we’re going to need digital twins of physical realities, which can be obtained through measures like this. We applied this AI solution to electrical poles, but there’s no reason this couldn’t be applied to damn near anything.
