Faced with the dual challenge of meeting both growing energy needs and ambitious carbon abatement goals, power grids have a friend in Patrick “Cade” Hay.
Hay is the operations manager at the Lamar Power Plant, a 1076-megawatt (MW) natural gas power station in Texas. He also was the first Plant Manager in Vistra Corp’s fleet to explore whether AI could be used to run their plants more efficiently. He saw a demonstration of how AI was being used by a metals company to improve recovery and realized how that could translate to his own power plant. Cade and his team partnered with McKinsey and decided to first apply AI to optimize Heat Rate (a measure of the thermal efficiency of the plant, namely the amount of fuel required to produce each unit of electricity).
The first insight to come out of the AI model was that the Lamar team could reduce the amount of time they run their duct burners. Duct burners essentially work like afterburners in jet planes: they provide a surge of energy when needed and operators use them as supplements to hit energy targets. The issue is that powering duct burners uses more fuel than regular methods, so it’s more expensive, generates more carbon emissions, and increases wear and tear on equipment. The AI models, once tested and validated, made recommendations that resulted in operators meeting their energy targets while being more efficient and prolonging asset life. As a result, operators were able to cut duct burner usage by approximately 30%—providing roughly $175,000 of annual fuel savings—while reducing the site’s carbon footprint.
Lloyd Hughes, a technical expert in power plant operations at another plant Vistra plant, has a reputation for always looking for opportunities to improve. When the analytics team worked with Lloyd to build a digital twin for his power plant in Odessa, Texas, he reflected, “There are things that took me 20 years to learn about these power plants. This model learned them in an afternoon.”
The quest for zero carbon emissions
The work at the Lamar and Odessa power plants are just two examples of a broader commitment by Vistra to harness AI to improve efficiency, reduce greenhouse gases, and supply more reliable and predictable power. This commitment led them to turn the heat rate insights into a solution they called the Heat Rate Optimizer (HRO) and scale it across all their fleet.
They rolled out the HRO to more than 68 power-generation units across 26 plants, racking up $23 million in savings. They also got more efficient along the way. It took 10-12 weeks to build the first HRO. Now, rolling the HRO out to each new plant only takes two to three weeks.
With the insights from the duct burner analysis and HRO solution clearly demonstrating the power of AI, Vistra hoped to build on that success. To date, more than 400 AI models have been deployed across the company’s fleet of power plants, leading to substantial environmental and operational impact. AI-assisted optimization efforts are successfully helping the fleet avoid greenhouse gas emissions.
Such reduction is an important part of the company’s journey to a 60% reduction in CO2e emissions by 2030—compared to a 2010 baseline—and achieving net zero carbon emissions by 2050.
The AI-driven advances at Vistra thus mark a shift in the power sector in terms of better efficiency, reliability, safety, and sustainability. If the 1% efficiency improvement from HRO in Vistra’s fleet was carried across all coal- and gas-fired power plants in the U.S., 15 million tons of carbon would be abated annually. That’s the equivalent of decommissioning more than two very large coal plants or planting about 37 million trees.
Scale every solution
From the beginning, Vistra’s leadership realized that achieving their efficiency and carbon-abatement aspirations required scaling every solution, and that has proved pivotal.
To do this, subject matter experts, data scientists, and analytics translators from McKinsey worked closely with a team of power generation and process experts at Vistra, as well as front-line operators. Together, they devised an approach to building plant-specific AI applications:
- First, they used a range of models to fit the customized needs of each plant and the required solution. These ranged from Bayesian regression models to deep-learning models. Each neural network model was made up of several layers, with each layer containing batch normalization, dropout, and activation. They created cross validation and out-of-sample testing sets using time-based splits to prevent overfitting models;
- They continually worked to find the right balance between model performance, explainability, actionability of insights, and maintainability;
- They extensively tested and validated models both offline (cold tests) and online (hot tests) with operators and power plant experts to make sure the models had learned the intricacies of the plants and were generating the right recommendations for operators to improve plant performance.
- Dedicated teams refactored proven models so they could easily be used and adapted to other plants. The models were then embedded within the existing production workflows and deployed live in the operator room.
- To sustain this effort, they built specific capabilities, including hiring and training talent, and setting up Machine Learning Operations (MLOps) infrastructure.
Vistra managers also made sure many voices contributed to the project along the way. “Every week, I’d bring in a different supervisor,” said Denese Ray, an operations shift supervisor. “We needed their feedback to build the model so we could get the most from it.”
When a solution has proven its value at a pilot site and is approved for scaling, a team of software and machine learning engineers immediately takes over to refactor, modularize, and containerize the code. That way there is a single software ‘core’ package for each deployment that can be updated and improved. A product owner manages the overall process and takes ownership of use and adoption.
Creating the next generation of visionaries
Vistra is just at the beginning of its journey with AI. The company has developed a strategy to create a series of AI and digital solutions that could capture several hundreds of millions of dollars in EBITDA improvements while improving reliability and contributing to its long-term emissions-reductions goals. Vistra plans to bring the power of AI to its other businesses, as well.
In the past year alone, Vistra has taken the lessons learned from using AI to improve the operations of its conventional power plants and applied them to its burgeoning zero-carbon fleet: namely, battery storage and solar. For their battery storage units, they were able to use analytics to find optimal temperatures for running their batteries. They’ve also been able to improve their dispatching strategies for when to charge and discharge in order to optimize the value of the battery over its lifecycle.
These new analytics tools helped create a bridge and conversation between their operations teams and commercial teams. Together, they’re able to see how the batteries perform each week and talk about what opportunities the models saw, enabling Vistra to run them in a way that found the optimal tradeoff between producing power and prolonging battery life. Many operators didn’t think there would be an opportunity to optimize renewables and storage fleets–Vistra sees it differently.
This story was produced by WIRED Brand Lab for McKinsey Digital.
