AI model could add more reliability to power grids

The new model from the University of Virginia is based on multi-fidelity graph neural networks (GNNs), a type of AI designed to improve power flow analysis, which is the process of ensuring electricity is distributed safely and efficiently across the grid.

The multi-fidelity approach allows the AI model to leverage large quantities of lower-quality data (low-fidelity) while still benefiting from smaller amounts of highly accurate data (high-fidelity). This dual-layered approach is said to enable faster model training while increasing the overall accuracy and reliability of the system.

By applying GNNs, the model can adapt to various grid configurations and is claimed to be robust to changes, such as power line failures. It helps address the longstanding ‘optimal power flow’ problem, determining how much power should be generated from different sources.

MORE FROM ENERGY & ENVIRONMENT

As renewable energy sources introduce uncertainty in power generation and distributed generation systems, along with electrification increasing the uncertainty in demand, traditional grid management methods struggle to effectively handle these real-time variations. The new AI model integrates both detailed and simplified simulations to optimise solutions within seconds, improving grid performance, even under unpredictable conditions.

“With renewable energy and electric vehicles changing the landscape, we need smarter solutions to manage the grid,” said Negin Alemazkoor, assistant professor of civil and environmental engineering and lead researcher on the project. “Our model helps make quick, reliable decisions, even when unexpected changes happen.”

Key Benefits are scalability, as it requires less computational power for training, making it applicable to large, complex power systems; higher accuracy, as it leverages abundant low-fidelity simulations for more reliable power flow predictions; and improved generalisability, as the model is robust to changes in grid topology, such as line failures, a feature that is not offered by conventional machine leaning models.

A paper detailing the work is published in Electric Power Systems Research.