An AI power play: Fueling the next wave of innovation in the energy sector

How Vistra Corp. is partnering with McKinsey to improve efficiency and reduce emissions by using AI

Tatum, Texas might not seem like the most obvious place for a revolution in artificial intelligence (AI), but in October of 2020, that’s exactly what happened. That was when Wayne Brown, the operations manager at the Vistra-owned Martin Lake Power Plant, built and deployed a heat rate optimizer (HRO).

Sidebar

The “heat rate” is a measure of the thermal efficiency of the plant—the amount of fuel required for each unit of electricity produced. To reach the optimal heat rate, plant operators continuously monitor and tune hundreds of variables, or “set points” on things like steam temperatures, pressures, oxygen levels, and fan speeds.

It’s a lot for any operator to get right 100 percent of the time—so Vistra thought AI could help.

Partnering for innovation

With this goal in mind, Wayne and his group worked together with a McKinsey team that included data scientists and machine learning engineers from QuantumBlack AI by McKinsey, to build a multilayered neural-network model—essentially an algorithm powered by AI that learns about the effects of complex nonlinear relationships.

This model went through two years’ worth of data at the plant and learned which combination of external factors—such as temperature and humidity—and internal decisions, like set points that operators control, would optimize the algorithm and attain the best heat-rate efficiency at any point in time.

We strive to provide individuals with disabilities equal access to our website. If you would like information about this content we will be happy to work with you. Please email us at: McKinsey_Website_Accessibility@mckinsey.com

Vistra team members provided continuous guidance about the intricacies of how the plant worked, and identified critical data sources from sensors, which helped McKinsey engineers refine the model, adding and removing variables to see how those changes affected the heat rate.

Through this training process, and by introducing better data, the models “learned” to make ever more accurate predictions. When the models were accurate to 99 percent or higher and run through a rigorous set of real-world tests, a McKinsey team of machine learning engineers converted them into an AI-powered engine. This generated recommendations every 30 minutes for operators to improve the plant’s heat-rate efficiency. At a meeting with all of Vistra’s leaders to review the HRO, Lloyd Hughes, a seasoned operations manager at the company’s Odessa plant, said, “There are things that took me 20 years to learn about these power plants. This model learned them in an afternoon.”

There are things that took me 20 years to learn about these power plants. This model learned them in an afternoon.

Lloyd Hughes, operations manager

With this kind of power at their fingertips, Wayne and his team could make better, more informed decisions. Acting on the HRO recommendations helped Martin Lake run more than two percent more efficiently after just three months in operation, resulting in $4.5 million per year in savings and 340,000 tons of carbon abated. This carbon reduction was the equivalent of taking 66,000 cars off the road. 1 If that doesn’t sound like a lot, consider this: companies that build gas-fueled power plants invest millions of dollars in research and development over four to five years to achieve a one-percent improvement in power-generation efficiency. Vistra hit that improvement level in only one-twentieth the amount of time using the data and equipment it already had. 2

Vistra has since rolled the HRO out to another 67 power-generation units across 26 plants, for an average one-percent improvement in efficiency, and more than $23 million in savings. Along with the other AI initiatives, these efforts have helped Vistra abate about 1.6 million tons of carbon per year, which is ten percent of its remaining 2030 carbon-reduction commitment. That’s equivalent to offsetting about 50 percent of what a 500-megawatt coal plant emits.

What happened at Martin Lake has happened at dozens of Vistra’s other power plants, with more than 400 AI models (and counting) deployed across the company’s fleet to help operators make even better decisions. It also reflects a core trait of Vistra’s AI transformation, which is that it isn’t a story of one massive hit, but rather the story of dozens of meaningful improvements snowballing to deliver significant value in terms of accelerating sustainable and inclusive growth. It’s also the story of how an organization architected an approach to rapidly scale every successful AI solution across the entire business. And it’s a story of how a continuous improvement culture, combined with a powerful AI modeling capability, helped leaders and plant operators do their jobs better than ever before.

With more than $60 million captured in about one year of work and another $40 million in progress, Vistra is well on its way to delivering against a roadmap of $250-$300 million in identified EBITDA and more than two million tons of carbon abatement per year. The AI-driven advances at Vistra have heralded a generational shift in the power sector in terms of improvements in efficiency, reliability, safety, and sustainability. If the one-percent improvement in efficiency the HRO delivered across the fleet was carried across all coal- and gas-fired plants in the US electric-power generation industry, 15 million tons of carbon would be abated annually—the equivalent of decommissioning more than two large coal plants or planting about 37 million trees. 3 That means less fuel needed to deliver power to the hospitals, schools, and businesses that rely on it. AI has the potential to bring similar levels of improvement to renewables as well, making them a more cost-effective and attractive energy option.

We strive to provide individuals with disabilities equal access to our website. If you would like information about this content we will be happy to work with you. Please email us at: McKinsey_Website_Accessibility@mckinsey.com

Turning points on the AI journey

Healthy skepticism and a culture of favoring action over words at Vistra meant that the biggest hurdle in the AI journey wasn’t the technology: it was the people. Vistra leadership and operations managers needed to know what AI could do and be convinced it could really work.

It was this ingrained culture of continuous improvement—alongside a highly competitive market and a commitment to sustainability—that convinced Vistra’s leadership they needed to give AI a chance.

Seeing real possibilities

The first question was a relatively simple one: “How can AI help improve the way Vistra generates power?”

Answers to that question bubbled over when 50 of Vistra’s top leaders came to a McKinsey-hosted workshop. In multiple sessions, experts explained how AI worked, walked through detailed case studies showing how other companies used analytics and AI to generate value and gave live demonstrations of technologies, including digitized workflows and machine learning. Leaders in analytics from various sectors—including Amazon, Falkonry, Element Analytics, and QuantumBlack, AI by McKinsey, as well as G2VP from the venture-capital world—provided insights and examples of how AI works.

“I saw an example of how a metallurgical plant was using AI to help its operators optimize set points and it clicked for me,” remembers Patrick “Cade” Hay, the plant manager for Vistra’s Lamar power plant. “I saw how I could translate that into helping me run my own plants more efficiently. This was my lightbulb moment.”

Company leaders and plant managers poured over process flow sheets and engineering diagrams to determine pain points as well as opportunities. This exercise allowed them to focus first on finding where the value was, then secondly on what technologies were needed to deliver it. Many of the operations opportunities were around yield and energy optimization and predictive maintenance, which according to our research, were the top AI use cases for manufacturing industries.

By the end of the session, Vistra had developed a strategy to develop a series of AI solutions that could capture $250-$300 million in potential EBITDA while helping the company achieve its 2030 emissions-reductions goals.

We strive to provide individuals with disabilities equal access to our website. If you would like information about this content we will be happy to work with you. Please email us at: McKinsey_Website_Accessibility@mckinsey.com

Codeveloping the AI

While the analysis looked promising, proving it in the field was what mattered. “If our plant managers aren’t bought in, then things don’t happen,” explained Barry Boswell, Vistra’s executive vice president of power-generation operations. “So, we said, ‘Let’s pick a leader who is knowledgeable and skeptical, because if we can win them over, we can get everyone.’”

They picked Cade. Not only did he run a top-performing plant in terms of profitability, reliability, and heat rate, he had a reputation for telling it like it is—Barry trusted Cade to tell him whether or not the value potential in AI was real. When he approached Cade about testing out a proof of concept to optimize duct burners, he was intrigued but predictably skeptical. Cade saw the potential in AI but was interested in finding out if it could actually help in the field.

Duct burners essentially work like afterburners in jet planes; they provide a surge of energy when needed. Operators use them as supplements to hit energy targets, which are known as their “dispatch point.” The issue is that powering duct burners uses more fuel than regular methods, so it’s more expensive, generates more carbon emissions, and increases the wear and tear on equipment.

McKinsey subject matter experts, data scientists, and analytics translators from QuantumBlack worked closely with a team from Vistra comprised of power generation and process experts as well as front-line operators to understand how the plant works, what data was available from the sensors already in place, and what variables could be directed—like the fact that the number of cooling fans running could be controlled, while the ambient temperature couldn’t.

Denese Ray, operations shift supervisor for Vistra’s Coleto Creek power plant; Doug Richter, the shift supervisor at the same plant; and Muro Kaku, McKinsey analyst and data scientist, discuss how they worked together.

Denese: “Every week, I’d bring in a different supervisor to join us for our check-in calls with McKinsey. In this way, Muro and the McKinsey team saw what we were dealing with. We really needed all our supervisors and their feedback to build the model so we could get the most from it.”

Muro: “That really mattered. For example, when we first trialed the HRO at the Coleto Creek plant, the recommendations were frequently rejected by the operators. We didn’t know why. When we spoke with the operators, they told us about specific rejection reasons, such as the tool recommended increasing superheat temperature by 30 degrees in a 15-minute interval to maximize the heat rate, but it takes the plant longer to heat up, so they rejected the recommendation. We then added a constraint to the tool to keep the increase and decrease to five degrees in a 15-minute interval, which operators could achieve.”

Doug: “During load changes [increase or decrease in power output], the parameters are changing constantly, so recommendations to increase air flow at that time were counterproductive. We worked with McKinsey to have the recommendations stop for 30 minutes while we were changing load until the plant was stable again.”

Denese: “We were all pretty skeptical of the tool at first, but when we got to play with the heat rate optimizer and see how it worked, and how well McKinsey worked with us in the plant, we understood how it could help.”

As the teams developed the models, plant operators reviewed recommendations to see what made sense, what other variables needed to be tested, and what kinds of recommendations the operators would find most helpful. By analyzing the effect of various inputs and set points on the plant—such as pressure and humidity, the angle of blades in the gas turbine, usage of inlet cooling technologies, and the age and performance of various components like filters and condensers—and running it through the model, the analysis was clear: overall duct-burner usage could be reduced by approximately 30 percent, which would result in the equivalent of $175,000 of yearly savings on fuel costs and wear and tear on the system, in addition to an abatement of about 4,700 tons of carbon per year.

“We worked closely with the team from McKinsey to develop AI models that reflected the realities of how power plants operate,” said Cade, “and then when we saw the recommendations coming out of the AI tool, I saw how much real value there was.”

Going for scale and adoption from the beginning

Vistra’s leadership realized from the beginning that the only way to achieve their efficiency and carbon-abatement aspirations was to scale every solution. “We manage the Vistra fleet as one. If a plant is doing something that works, we want every plant to do it,” says Barry. “That’s what we’re built to do.”

That realization led Vistra to invest in a five-part system to scale and sustain AI solutions.

***

QuantumBlack’s Ayush Talwar, an expert associate partner, provides more details on the AI model development:

  • What kind of models did you use?

    We used a range of models to fit the specific needs of the plant and the solution. These ranged from Bayesian regression models to deep-learning models.

  • How did you choose which model to use?

    We wanted to find the right balance between model performance, meaning its accuracy; explainability, which describes relationships to the operators; the source richness, which refers to the amount of data available; potential actionability, or how many of the most important features in the model people can actually control; and maintainability, which will reduce need for more technical skills to maintain them.

  • How would you compare the models you used to older ones?

    Our models were more accurate than previous models, and importantly, more actionable. We used a variety of metrics to measure model performance such as the mean absolute percentage error (MAPE) and root mean squared error (RMSE). Our models typically achieved MAPEs of less than 1.5 percent.

  • How did you build up your models?

    Each neural network model was made up of several layers with each layer containing batch normalization, dropout, and activation. We created cross validation and out-of-sample testing sets using time-based splits to prevent overfitting models. We added features based primarily on domain expertise.

  • How did you turn models into tools operators could use?

    We used machine learning models to make accurate predictions of outcomes. We then wrapped meta-heuristic optimization algorithms around those models to generate recommendations that helped operators make decisions about how to run the plant better. The models were then embedded within the existing production workflows and deployed live in the operator room.

Vistra’s story is far from finished. As Cade put it, “We’re just looking at the tip of the iceberg.” Vistra’s roadmap for 2022 and 2023 includes bringing AI to its rapidly growing renewables fleet of solar and batteries to optimize yield and reliability, among other initiatives.

To help sustain this ambition, Vistra is building up its talent bench. In addition to hiring a small team of data scientists and engineers, Rachit has partnered with the University of Texas at Dallas to offer basic, intermediate, and advanced courses in AI and analytics for Vistra employees. Some 70 people have already completed courses, including those reskilling from statistics to machine learning. Vistra has also built relationships with local colleges and universities to develop internship programs and work with students in capstone projects to identify top technical talent.

“We can’t sit around and just do what we did yesterday to be ready for tomorrow,” Barry says. “We’ve seen enough to know what’s possible.”

The authors would like to thank Wayne Brown, Rachit Gupta, Patrick “Cade” Hay, Lloyd Hughes, Denese Ray, and Doug Richter from Vistra Corp., and Richard Bates, Dan Hurley, Pablo Illuzzi, Nephi Johnson, Muro Kaku, Jay Kim, George Lederman, Abhay Prasanna, Noel Ramirez, and Ayush Talwar from McKinsey.