Powder River Energy Corporation has one of the nation’s more industry-heavy loads. More than 80 percent of its power goes to 13 coal mines and extensive oil and gas development in Wyoming’s rugged northeastern region.
The volatile nature of those industries, whose power use can shift significantly based on the commodities markets and other external circumstances, makes accurate load forecasting both challenging and a priority for planning.
As he considered ways to meet that challenge, Quentin Rogers, Powder River’s vice president of engineering and technical services, began experimenting with using neural networks, which are part of artificial intelligence (AI) or machine learning, to improve load forecasting. Those experiments are “still in their infancy,” he says, and the co-op is not using neural networks for now, partly because some federal and state agencies are unfamiliar with the approach.
But as he looks down the road, Rogers believes AI/machine learning is likely to be an important part of load forecasting and other electric utility operations.
“I think it’s something exciting that a lot of co-ops will be involved in in the next 10 years,” he says.
Industry analysts and other technology experts agree with that assessment. An old joke about AI is that it’s been “five years away” for the past 30 years. But in truth, it has steadily become more common in an array of U.S. businesses.
“Artificial intelligence is all around us. We just don’t often realize it,” says Jim Spiers, NRECA senior vice president for business and technology strategies.
From in-home devices like Amazon’s Echo, which learns consumer interests and desires as it responds to queries, to business software that fine-tunes delivery networks by monitoring and adjusting to real-time results, AI is changing the way complicated systems are managed.
Yet, despite its growing importance, a degree of mystery still surrounds AI. Even experts don’t agree on a precise definition. But in essence, it involves developing computer systems that can learn and mimic human decision-making by sifting through large amounts of data and recognizing patterns.
For example, when using neural networks—a set of algorithms modeled loosely after the human brain—for load forecasting, “you train them with known data, and then it becomes a kind of black box, where you put in your information, and the output on the back end should provide an accurate prediction of what’s going to happen,” Rogers explains.
Spiers notes that the use of AI is simply the latest step in a long history of automation replacing manual calculation.
“We’ve moved from counting things by hand through a whole series of steps where we’ve automated data collection and analysis,” he says. “You now have machines making decisions, but it’s based on a series of rules. … The theory is that it’s using precisely the same logic that a human would use, just without human intervention.”
‘The potential … is huge’
AI is already being used in the utility industry, says Mark McGranaghan, vice president, integrated grid for the Electric Power Research Institute (EPRI). For example, General Electric has an initiative to use AI for advanced analytics on gas turbine generators. Another company is using intelligent software to improve the sun-tracking capabilities of large-scale solar arrays, resulting in up to 6 percent in energy gains.
McGranaghan says electric power utilities have yet to make widespread use of AI’s potential, but hopes a recent EPRI initiative to facilitate data sharing between utilities and AI vendors will help move the industry’s use of the technology forward.
“We are living right now in a world of engineering-based models,” he says. “But the potential to take the data and let AI learn from it and see what it can do in conjunction with engineering-based models, or in place of engineering-based models, is huge.”
NRECA is a partner in the Department of Energy’s Grid Resilience and Intelligence Project (GRIP), which is advancing the use of AI.
“The GRIP project exists to create a software platform and a set of intelligent applications to improve the resilience of the grid in the United States,” says David Pinney, analytics program manager in NRECA’s Business and Technology Strategies group.
Pinney, who is NRECA’s GRIP project manager, says load forecasting is one of the key areas where AI/machine learning can play an important role. Tools are already on the market that provide the neural net algorithms that can be used to develop different intelligent models that help with forecasting.
Rogers sees a related benefit in using a machine-learning algorithm to automate a level of demand-side management. AI could use shorter-term load forecasting to see a peak approaching “and give signals to members to where they could potentially back off load,” he says.
Anomaly detection is another area where Pinney sees promise.
“We’ve built some initial models here, and they’re very similar to load-forecasting models,” Pinney says. “The application is somewhat different. After you detect anomalies in load data, typically what you’re looking for are malfunctioning equipment or nontechnical losses, i.e. theft, or you might find load behavior that’s really strange, and you might be able to tell the member that you’re peaking at an odd time, that sort of thing.”
As AI algorithms learn, he adds, they could be used for predictive maintenance—helping co-ops get to equipment before it fails—and in root-cause analysis of outages. The system would use data on the age of assets, their maintenance history and reliability, and other factors to point toward lines, transformers, and other system components that should be considered for repair or replacement.
Load and generation disaggregation also presents an opportunity for AI/machine learning.
“You’re basically given a meter reading over time, and you want to know what it breaks down into—is it their HVAC, their water heater, their lights—what’s in that reading,” Pinney says. “Machine learning could help answer those questions.”
For load disaggregation, an AI model could be trained on meter signals that have known components, and then that model could be applied to a non-disaggregated meter signal to determine component use.
“For generation disaggregation, the technique is similar, except the components there are types of generation, such as solar, gas, wind, etc.,” Pinney says.
Spiers notes that emerging distributed energy management systems, or DERMS, provide a logical place to apply artificial intelligence. Rooftop solar and other distributed generation, storage, fluctuating demand, and other factors are interacting at a level of complexity that an AI system can handle more effectively than human operators.
“It’s now this multiple two-way flow of electrons, data, and money,” he says. “You’ve got all these things happening, and you’ve got to have a tool that will manage everything, that will get them to operate together in an optimal fashion. We are still a ways off from such a system, but research and development are pursuing the promise.”
The key in this area and others, McGranaghan says, is building out the large data sets that AI algorithms need to draw on if they are to teach themselves how to run an operation more efficiently.
“The technology itself for the artificial intelligence is there, and it’s open, and there’s lots of companies that are ready to apply it,” he says.
Within five years’
Both Spiers and McGranaghan see untapped potential for AI in anticipating the needs of consumer-members. The use of AI to analyze consumer data should take privacy concerns into consideration, but “the customer side is so ripe for services,” McGranaghan says. “And this is an area where change will happen quickly because there are so many customers that the return on innovation is very fast.”
Artificial intelligence could allow cooperatives to reinforce the value of their consumer-centric model, Spiers says.
“Cooperatives have a significant amount of data about consumer-members. But I think what we’ll end up doing, for some of our consumers—not all—is buying data from others,” he continues. “There are consumer behaviors in other parts of their lives that help us to know that they might really like a particular product that helps them save energy or control their energy consumption, costs, and quality of life, for example.
“You’ll be able to mine other sorts of data in areas totally unrelated to the energy space to identify those consumers that might be good participants in co-op programs that also have huge value to the co-op and its consumer-members.”
On the national level, much of the focus on AI has been connected to cybersecurity and system resilience, using the capabilities of intelligent systems to identify potential threats or weaknesses that could disrupt the grid. But Spiers notes that the systems also hold the potential of increasing efficiency of the grid.
Utility use of AI/machine learning may be in its early stages, but McGranaghan believes the old joke about its widespread adoption forever being just five years away will soon be outdated.
“I think we’re on the verge of actually using AI for a lot of things,” he says. “In areas like DERMS and distributed resources, those are just starting to really be widely applied anyway. We’ll be using AI as the penetration gets larger—those things will kind of mature together.
“I’m saying it’s within five years.”