Take wind farms, for instance. Some already rely on software that continuously adjusts turbine rotation to crank out as much energy as possible at any given moment. That’s where Colin Parris and his team at GE Research come in. They have been developing technology that makes AI a bit more human by giving it a sense of its own capabilities — in a way, by programming it with a little humility.
Their “humble” AI software has helped many renewable energy companies save money by wringing more kilowatt-hours out of their turbines. But Parris and his colleagues wondered whether they could develop the humanlike characteristics of AI a little further. They thought they’d see whether they could make it “curious.”
What keeps Parris’ software “humble” is that it knows its limitations. But the software’s decisions about how to adjust a turbine to current atmospheric conditions are only as good as the digital model of the turbine that it relies on. For example, when an unexpected wind blows that the humble AI cannot model, it will take a step back and switch the turbine to a safe mode. While ensuring reliability, that can sacrifice peak performance.
To maximize efficiency, Parris and his team wanted to raise the quality of the models that the software relied on to make judgments. That’s where curiosity comes in.
Curious AI kicks in when the confidence level in a digital model of the turbine falls below a certain range. Rather than accept the limitations of the model, it asks why certainty is below a given threshold. Say it’s a snowy winter morning in Alaska with stiff winds. Perhaps the model is missing information regarding two or three key variables — extremely low temperatures or unusually heavy snow, for instance — that would boost its accuracy. Curious AI will push for more data on those missing variables, helping engineers understand how to improve the software.
“This is just part of a model that points out how to get better,” explains Parris. “It’s like a person saying, ‘I know my weakness is punctuation in English — let me go practice that.’ ”
With the models’ shortcomings laid bare, GE engineers can figure out how to improve the models. One of the most effective ways to sharpen predictive skills is to connect to high-powered simulators, such as those operated by Illinois-based Argonne National Laboratory, which is part of the U.S. Department of Energy. Parris and his team joined with Argonne to improve the accuracy of GE’s software.
To a researcher like Parris, gaining access to the national labs is a bit like winning a golden ticket to Willy Wonka’s chocolate factory. Its 17 labs house some of the world’s best supercomputers and throngs of experienced physicists, making it possible to design simulators powerful enough to calculate almost as many curveballs as Mother Nature can throw. The more simulations GE Research can run, the smarter the models become.
From there, GE Research can test its refurbished models by running them at a 40-plus-turbine wind farm in the Southwestern U.S. The next step will be for chief engineers to observe the models over the next year. If the engineers see improved performance, Parris and his team will expand the Curious AI model to other wind farms around the world.
The early signs are promising. In sample sets, new humble AI models bolstered energy production by a few tenths of a percentage. That might not sound like much, but those decimals represent hundreds of thousands of dollars in savings. Those upgrades allow wind farms to gain extra power without sacrificing production, making for a more reliable supply of renewable energy.
“You don’t need a weatherman to know which way the wind blows,” Bob Dylan sang, and Parris’ work might bear him out. Turns out all you need is artificial intelligence and some mighty powerful simulators.