Item

Abstract

Assessing and mitigating risks in future power grids requires comprehensive analysis of their dynamic behaviour. Probabilistic stability analyses, which evaluate large ensembles of disturbances, are well-suited for this purpose and became mandatory for many grid operators. However, the computational costs of simulations impose strict limits on the number of configurations that can be evaluated. This study demonstrates how machine learning (ML) can address this challenge by enabling efficient prioritization of scenarios for detailed analysis in probabilistic dynamic stability assessments. We apply fault-ride-through probability—a practical metric measuring the likelihood of all grid components remaining within operational bounds after a fault—to show how ML can bridge the gap to real-world applications. A new dataset comprising thousands of dynamic simulations of synthetic power grids is generated to train ML models. Results reveal that ML models not only accurately predict fault-ride-through probabilities but also effectively rank the criticality of buses, identifying components most likely to destabilize the system and requiring further analysis. Importantly, the models generalize well to the IEEE-96 Test System, underscoring their robustness and scalability. This work highlights the transformative potential of ML in enabling efficient, scalable probabilistic stability studies, paving the way for integration into contingency screening for real-world grid operations.