Southwest Arkansas Electric Cooperative’s drone keeps an extremely detailed eye on its grid.
Crews get unprecedented visibility on system conditions and potential trouble spots as the drone takes hundreds of photos and hours of video while journeying over some 5,000 miles of distribution infrastructure and 125 miles of transmission line.
“You get a perspective from the drone you just can’t get from the ground,” says Dion Cooper, Southwest Arkansas’s vice president of information technology, communications, and cooperative services. “You might see pole-top deterioration and cracked insulators or crossarms that you wouldn’t see from the ground.”
But as the
Texarkana-based cooperative’s drone use has rapidly expanded, looking through all those videos and photos has become an increasing challenge, Cooper says.
Southwest Arkansas isn’t alone. About 25% of electric co-ops in the United States are now generating “tons of data” with drones, says David Pinney, NRECA’s software engineering manager. “We consistently hear that reviewing all this footage is a real pain point.”
To ease the pain, Southwest Arkansas is working with NRECA on a solution that uses machine learning to train a computer algorithm to sort through drone video and photos and identify potential issues.
Image recognition is just one of several applications NRECA is developing as part of the Grid Resilience and Intelligence Project (GRIP), an ambitious U.S. Department of Energy (DOE) effort that seeks to harness the power of artificial intelligence, or AI, and machine learning to transform the grid.
GRIP’s goal is to develop a wide-ranging suite of tools that will help co-ops and other electric utilities handle power fluctuations, avoid failures, resist damage, and recover more quickly from disruptions like major storms, cyberattacks, or solar flares. Other GRIP applications are taking advantage of AI and machine learning to improve the integration and control of distributed energy resources like solar and wind power.
NRECA was one of DOE’s principal partners in the program, receiving a $900,000 grant over three years to develop software. The effort began with a general investigation of machine-learning techniques that could be applied to utilities and then was narrowed down to areas where NRECA saw opportunities to provide a more immediate benefit to co-ops, focusing on specific areas of innovation, Pinney says.
Advancing App Development
Several of NRECA’s GRIP applications are available at its
Open Modeling Framework website:
Energy storage control. Pinney says battery storage is becoming a more affordable and realistic option for cooperatives, particularly in a power supply agreement when it’s important to manage peak demand. But it’s important to be able to accurately forecast peak load to determine when to use stored energy. Machine learning, a form of AI that uses algorithms capable of sifting through data and learning through experience, enables NRECA to use a co-op’s existing load data to train an application to forecast when it would be optimal to shave peak load with stored energy. “The end result is an application that, using your historic data, can predict three or four days ahead whether it would be good idea to dispatch your battery storage,” Pinney says.
Load disaggregation. “Something like 85% of all co-op meters are smart meters, and that’s a tremendous source of information about what’s going on on the grid,” Pinney says. That wealth of data also provides an opportunity if it can be used to determine individual sources of load on the other side of the meter, for rate design and other offerings targeted to help co-ops and their members save money. The algorithm NRECA developed looks at the correlation between several factors, including load duration, load shape, and temperature, and then uses machine learning to determine what type of appliance is drawing power. “Ultimately, it gets within a couple of hundred watts for each of the constituent loads within these combined load shapes,” Pinney says. “We’ve been happy with the accuracy and are working with a couple of co-ops to validate the results with more field data.”
Solar generation detection. This application gives cooperatives and other utilities the ability to detect solar installations on their systems that may not have been disclosed to the co-op.
Power phase identification. If the phase of a consumer’s meter is unknown, it can make balancing load on the system more difficult. NRECA’s solution is a model that compares meter-reading voltages to substation voltages over time and uses machine learning to correctly predict the meter’s phase. NRECA is working with several cooperatives to deploy the model in real-world conditions.
As co-ops collect more image data from drone surveys, they have an opportunity to use computer-vision techniques to speed up the image analysis. Pinney says his team’s ongoing image-recognition work is particularly challenging, even for an AI application that learns as it goes, because to work properly, the system has to recognize and classify objects that might not look exactly alike. For example, single-arm, side-arm, and double-arm crossarms can be hard for a computer system to differentiate. The system also needs to be able to recognize objects from different angles, perspectives, and distances.
Ultimately, to be useful to a power utility, the algorithm has to be able to flag images that show something might not be right with a piece of equipment. Training an application, Pinney says, can take thousands of images. Southwest Arkansas and
Wake Electric Membership Corporation, in Wake Forest, North Carolina, have been supplying drone imagery to NRECA as it develops its image-recognition app.
“Our initial attempts have focused on identifying poles and pole attachments in video,” Pinney says. “Ultimately, with higher quality models, we hope to identify not just what’s on the poles but whether the equipment is in good condition.”
Industries of all kinds are investing heavily in image-recognition software and accelerating advances in the field. The automobile industry, for example, has already had success in creating sophisticated computer-vision models that are helping to make vehicles nearly self-driving, Pinney says. He expects the research for utility image-recognition models to progress rapidly.
“I think we’re only a year or two away from having something useful day to day,” he says.
Such an app could save countless man-hours.
“If we could fly those lines, capture that video or those images, run them through a machine-learning program, and have that program flag the trouble spots—then you’re not looking at everything,” Southwest Arkansas’ Cooper says. “You’re looking at the trouble spots, and you’re saving a lot of time, and you’re able to really focus on them.”
Beyond Line Of Sight
One factor that limits drone data collection, Pinney says, is a Federal Aviation Administration rule that a drone must remain within the pilot’s sight unless the user receives a waiver, which comes with prohibitive requirements like the use of ground-based radar. But he predicts that as drones become increasingly vital tools for utility reliability, pressure will grow to relax the line-of-sight restrictions.
“These techniques already can be useful in reviewing the large amount of data collected from manual flights,” Pinney says. “But as beyond-line-of-sight and automated flights become more common, the techniques will become even more valuable as the volume of data collected increases dramatically.”
The transformation of the electric utility industry toward smart technologies that improve resilience and efficiency is ongoing, Pinney says, but it will take widespread use of AI and machine-learning apps to fully realize the value of these new systems and devices.
“The technology works. We’re going to see more and more of it,” he says. “There’s so many useful application areas that it’s just a question of how long it’s going to take us to integrate these techniques into our business.”