Restoring Power to the Grid

This model can also accurately approximate alternating current power flow, which is computationally more complex than direct current and is a more accurate representation of the grid during severe disruptions such as black start conditions, said Richard Garrett, a Sandia computer scientist who led this portion of the project.

Recently the team began validating the restoration schedules generated by their code against ones created by industry-standard optimization software to ensure that Sandia’s model recommends plans that can be realistically executed, Kevin said.

Modeling Operator Decision-Making
The fulcrum of the overall model lies with the operator decision-making code, Kevin said. This algorithm takes the results from the optimization code and enacts it on a third code, which is a physics-based simulation of the grid and how it dynamically responds to the operator’s actions. Walt Beyeler led the development of the third portion.

The operator decision-making model is based on a research-based cognitive model created by scientists at Carnegie Mellon University, said Casey Doyle, a systems analyst who is leading this portion of the model. This well-established model was adapted for power restoration by coding in expert knowledge about the subtasks necessary to complete main tasks such as the steps required to start a generator and then connect it to the nearest substation. They also added in safeguards so the cognitive model wouldn’t freeze if the grid behaved unexpectedly, Casey said.

“We’re trying to create a cognitively sound agent that can read in a restoration plan that’s created by the optimization model and then try to implement it on a simulation of the power system,” Casey said. “It goes through step-by-step and reads what the schedule says has to happen and it tries to implement it. It does all the decision processes in between, including making sure that the frequencies match before lines are connected.”

The operator model interacts with the model of the grid through a simulated console and is limited to the knowledge presented by the console, rather than presuming the grid operator knows everything, which is typically assumed in power-restoration models.

In fact, the operator model can assess whether the network model’s behavior matches up with what it is expecting based on the results of the optimization algorithm, Kevin said. The simulated console can also allow the team to swap in actual feeds of information from the grid for the network dynamic model, if a partner provides the information, he added.

“Black starts are really rare, extreme events, but when one happens it’s really bad,” Casey said. “Even in partial blackouts, like what happened in Texas in 2021, people died because they didn’t have power, they didn’t have heat. If you have a complete blackout, it’s likely that it would be caused by a hurricane or earthquake and operators are trying to restore power to whole communities. Delays in power restoration could cause even more damage or loss of life. It’s hugely impactful to understand how to bring the power back as quickly as possible.”