Researchers at the Department of Energy’s (DOE) Argonne National Laboratory are midway through a $1 million, three-year project to explore how artificial intelligence could lower the operating costs of nuclear power.
The premise of Argonne’s project to develop smart, computerized systems for nuclear plants is that their costs will have to come down if they are to play a role in the U.S. clean energy economy by providing large amounts of clean electricity.
Nuclear plants are expensive, in part, because they demand constant monitoring and maintenance to ensure consistent power flow and safety, according to the Argonne researchers. The expense of running nuclear plants has made it difficult for them to stay open, they said.
“Operation and maintenance costs are quite relevant for nuclear units, which currently require large site crews and extensive upkeep,” Roberto Ponciroli, a principal nuclear engineer at Argonne, said in a statement. “We think that autonomous operation can help to improve their profitability and also benefit the deployment of advanced reactor concepts.”
The project, funded by the DOE Office of Nuclear Energy’s Nuclear Energy Enabling Technologies program, aims to create a computer architecture that could detect problems early and recommend appropriate actions to human operators.
“In a world where decisions are made according to data, it’s important to know that you can trust your data,” Ponciroli said. “Sensors, like any other component, can degrade. Knowing that your sensors are functioning is crucial.”
A typical nuclear plant has hundreds of sensors. The job of inspecting each sensor — and the performance of system components such as valves, pumps, heat exchangers — currently rests with staff who walk the plant floor. Instead, Argonne is exploring the potential for algorithms that could verify data by learning how a normal sensor functions and looking for anomalies. Once a sensor’s data is validated, an artificial intelligence system would then interpret signals from the sensor and recommend specific actions.
The technology could save the nuclear industry more than $500 million a year, Ponciroli and his colleagues estimate.
An artificial intelligence method called reinforcement learning replicates judgments humans make all the time by teaching the system to make decisions by evaluating potential outcomes.
At a nuclear plant, computers could detect problems and flag them to plant operators as early as possible, helping optimize controls and also avert more expensive repairs, as well as cutting back on maintenance on equipment that doesn’t need it, the Argonne researchers said.
In partnership with industry partners Argonne engineers have built a computer simulation, or “digital twin,” of an advanced nuclear reactor that can also be adapted to existing nuclear plants.
The Argonne team is validating its artificial intelligence concept on the simulated reactor and has completed systems to control and diagnose its virtual parts. The remainder of the project will focus on the system’s decision-making ability — what it does with the diagnostic data.
“The lower-level tasks that people do now can be handed off to algorithms,” Richard Vilim, an Argonne senior nuclear engineer, said in a statement. “We’re trying to elevate humans to a higher degree of situational awareness so that they are observers making decisions.”