Research reveals hidden gifts of the ‘black box’ for modeling grid behavior

To create a more resilient electric grid that meets the nation’s increasing power demands, utilities are incorporating a wider array of energy sources. But this shift requires the ability to predict how the grid will react to fluctuations in the flow of electricity from new sources of power.

To plan ahead and avoid disruption to the power supply, utilities use models to anticipate when and where to direct a given amount of electricity. A model is a series of calculations—in this case, estimated electricity supply and demand.

Researchers at the Department of Energy’s Oak Ridge National Laboratory have developed a dynamic modeling method that uses machine learning to provide accurate simulations of grid behavior while maintaining what is called a “black box” approach. This technique does not require details about the proprietary technology inside the equipment—in this case, a type of power electronics called an inverter.

Engineers incorporated the new modeling capability into an open-source software tool and demonstrated its success with different scenarios and inverter brands. The work is published in the journal 2024 IEEE Energy Conversion Congress and Exposition (ECCE).

“Normally, it’s hard to get modeling accuracy without understanding the structure and control parameters of internal systems, proprietary information that companies may not want to share,” said Sunil Subedi, who led members of ORNL’s Grid Modeling and Controls group on the project.

“And while that level of detail improves accuracy, it also adds to the computational load and makes analysis burdensome.” It often requires the use of high-performance computing, which is energy-intensive and time-consuming, he said.

The ORNL model uses a deep learning algorithm to address these challenges. Researchers trained the model using test cases that reflect changes in power flow and sudden shifts in voltage. They then ran a simulation based on a specific vendor’s equipment, repeating the process with data from another vendor to compare results for consistency.

The team found that their black box model—the first of its kind to work with free open-source software—produced results with an average error rate below 5% over a range of operating conditions. This exceeds industry standards for grid system planning and operation, design testing and field deployment. The model also runs 10 to 20 times faster than more energy-intensive conventional methods, Subedi said.

“The machine learning approach lets you get what you need by representing a system with just data, which is fascinating,” Subedi said. “The technology strikes a balance between accuracy and flexibility, overcoming the limitations of previous approaches and providing utilities and manufacturers with new capabilities.”

The method allows producers of power electronics to more easily evaluate how new controls and protection designs would function in full power distribution systems. This insight could shorten product development timelines to help new technologies reach the grid faster. The modeling capability can also build utility confidence in diversifying energy sources to enhance the overall power resilience and reliability.