The mission of this group is to bring together utility professionals in the power industry who are in the thick of the digital utility transformation. 


Utilities explore the promise, power of AI

image credit: © Wrightstudio |
Matthew Abbitt's picture
VP of Product Management ARCOS

Matthew has a strong background in helping software companies grow and thrive through product management practices and strategy. Previously working in the startup world, Matthew has led product...

  • Member since 2021
  • 3 items added with 2,934 views
  • Oct 27, 2021

This item is part of the Advances in Utility Digitalization - October 2021 SPECIAL ISSUE, click here for more

More work and less workers along with ever-rising demand by energy consumers has utilities increasingly interested in the advantages of automation and artificial intelligence, or AI. 

According to the U.S. Energy Information Administration, utilities’ spending on building and maintaining distribution and transmission systems continues to rise. In 2019, utilities spent $31.4 billion—more than half of all distribution spending—to “replace, modernize, and expand existing infrastructure,” says EIA. The Wall Street Journal reported on Oct. 14, that “Workers are quitting at or near the highest rates on record in sectors like manufacturing . . . and utilities.” The utility industry also has a slightly older workforce than other industries, so retirements are a factor. Data from the Bureau of Labor Statistics for 2020 show the median age of all workers across industries is 42.5 years, 44.8 years for the utility industry. 

Your access to Member Features is limited.

Modeling solutions 

Utilities’ interest in getting more work has managers examining an array of processes—some more complex than others—that have always worked but could benefit from greater efficiencies. One way to do that is by automating manual processes that rely on keystrokes, whiteboards and pen and paper. Automation performs repetitive, monotonous tasks to move things through a cycle. But as complexity grows, automation isn’t always enough. That’s where AI comes in. AI taps computer systems to mimic the way humans learn. AI also employs machines to decide whether to move something forward without having the problem already programmed. 

Recently, my company created two machine-learning models for the utility industry. One model focused on detecting images to train machines to spot problems with utility assets like poles, transformers, and cutouts. We began with the premise that not everyone in the field is a line worker, but other workers are still on site and could assist with inspections if they had a mobile app to capture photos of assets and parse them to look for problems.  

In our first model, a worker could look at a pole and take a photo of a transformer. Next, the mobile app would electronically deliver the picture to a workforce management system then upload and save the image to a work management solution that runs the photo through AI. As AI takes over, the machine might see corrosion on the transformer, which would kick off a corrective action for the asset. With a system like this, the utility could catch problems before they cause an outage and gain a greater confidence about rolling a truck to make a repair. As AI digests more and more photos, the machine grows smarter about picking out problems and the necessity for triggering a truck-roll.  

What’s required to make AI smart is uploading enough imagery—a set of training photos—to teach the machine what and when to tag something as broken or problematic. With our model, utilities could set conservative measures by which AI operates. For instance, the machine wouldn’t activate a repair crew unless the AI system reached a threshold of 95 percent certainty or greater about a problem it analyzed with an asset. 

Some utilities may soon adopt what our models show is possible. But here’s something else for utilities to consider for the future. Imagine a world where customers get their electric bill, and they’re encouraged to take a photo of their electric meter to earn a discount. Collecting these photos could save the utility a truck-roll by using AI to determine the meter’s physical condition (i.e., information beyond the usage reading) and take corrective action. Here’s another example. Let’s say a utility’s service territory is prone to wildfires. By asking customers to take photos of the pole bringing service into their home, AI could analyze if there’s a need for a vegetation crew. If, for instance, a tree limb stretches across or near a primary, AI would spot that and trigger a truck-roll. A system like this could immediately put “eyes” on assets that would otherwise require hundreds or thousands of utility workers to manually sift through and analyze photos. 

Going beyond automation 

Our second model involved the call out of line workers for storm restoration. Think of having to activate several hundred line workers on, say, Christmas Eve. As things work today, a utility storm manager would begin the process by launching an automated callout. While that’s far faster than a manual process, the manager still can’t know how many line workers are likely to accept the work. With the automated callout, the manager chooses a roster of line workers from the callout system, saves it, launches the call and within a few minutes sees how many accept. Let’s say the manager gets 150 line workers on that first round of calls. The utility would then choose another roster and launch a second, or maybe a third, callout.  

Using our AI model, we took a utility’s callout data, and the machine scanned the rosters and compared line workers’ previous acceptance rates. The AI could tell which line workers were more likely than others to accept a call on a weekend, holiday or during different times of the night or year. With that data, AI built a list with the line workers most likely to accept, while accounting for work agreements. This approach would save storm managers or dispatchers valuable time they’d otherwise use to re-launch callouts. Another way to apply AI would be for shift scheduling. Utility manages could give an AI-powered system its team's shifts, and the machine could create the shifts based on required skills, seniority, past availability to work certain hours and days.  

Our own survey data of the hundreds of utilities we work with has shown executives and front-line managers have an interest in AI tools like the ones described here. At the end of the day, the paper processes (e.g., damage assessment and callout) that many utilities have already automated could be augmented by AI. That, in turn, would allow utilities to do more work with fewer resources, while better serving consumers of energy.

Matt Chester's picture
Matt Chester on Oct 27, 2021

 With our model, utilities could set conservative measures by which AI operates. For instance, the machine wouldn’t activate a repair crew unless the AI system reached a threshold of 95 percent certainty or greater about a problem it analyzed with an asset. 

Is the existing suite of potential 'problems' with the equipment well defined enough that you can be confident little would be missed? Or is there still an amount of human spot checking required to see if there are any unexpected issues that weren't a part of the training set? 

Matthew Abbitt's picture
Matthew Abbitt on Oct 27, 2021

The initial data set is important. Think of it like this. When looking at a photo for a piece of equipment, how would a human classify it? The ones that can be easily classified by humans are the classifications you should start with. This will provide the AI model with greater accuracy. Then as accuracy levels rise for the common problems, you can start introducing other asset or equipment problems to the AI model. As those the accuracy levels of those rise, you can introduce more and more. This is how the existing suite of potential problems with equipment become defined with confidence. 

Paul Korzeniowski's picture
Paul Korzeniowski on Nov 24, 2021

Good point. The model should become better over time as the user collects more information. The challenge initially is collecting enough information so modeling can occur. The process never reaches 100% certainty but the probability numbers should increase with more data and more tuning of the algorithm. 

Matthew Abbitt's picture
Thank Matthew for the Post!
Energy Central contributors share their experience and insights for the benefit of other Members (like you). Please show them your appreciation by leaving a comment, 'liking' this post, or following this Member.
More posts from this member

Get Published - Build a Following

The Energy Central Power Industry Network is based on one core idea - power industry professionals helping each other and advancing the industry by sharing and learning from each other.

If you have an experience or insight to share or have learned something from a conference or seminar, your peers and colleagues on Energy Central want to hear about it. It's also easy to share a link to an article you've liked or an industry resource that you think would be helpful.

                 Learn more about posting on Energy Central »