Revealing drivers and risks for power grid frequency stability with explainable AI

Machine learning can predict power grid stability better than traditional models: The researchers successfully used AI to predict critical frequency stability indicators (RoCoF and Nadir) across three major European power systems, demonstrating that modern algorithms can capture complex grid dynamics that conventional human-built models cannot.

Explainable AI reveals what actually matters for grid stability: By using interpretable AI methods, the team identified that load ramps (sudden changes in electricity demand), generation ramps, electricity prices, and forecast errors are the key factors driving grid stability—insights that would remain hidden in standard "black-box" AI models.

Black-box AI poses security risks for critical infrastructure: The paper highlights that the opacity of many machine learning models creates "severe security risks" for power grids, suggesting that using unexplainable AI for such critical systems could leave vulnerabilities undetected and make it impossible to verify safe operation.

Explainable AI enables scientific understanding, not just predictions: Rather than simply accepting AI predictions as correct, XAI methods allow grid operators and researchers to understand why the system makes certain predictions, enabling better decision-making and more trustworthy automation of power grid management.

Forecast errors and market prices influence grid stability: The finding that electricity price signals and prediction errors affect frequency stability suggests that improving forecasting accuracy and market design could be practical ways to enhance grid resilience.