Interpretable power system stability assessment using Kolmogorov-Arnold Networks

Machine Learning (ML) models have shown promise in addressing complex challenges in power systems, where traditional simulation-based methods encounter limitations. Despite the high accuracy demonstrated by these models, most of them are black-box in nature, which hinders the necessary trust to foster widespread adoption in safety-critical power systems. Existing Interpretable ML (IML) approaches primarily focus on making models transparent, but may sacrifice predictive accuracy. This paper proposes the use of Kolmogorov Arnold Networks (KANs) to develop ML solutions that are accurate and inherently transparent for power systems applications. By leveraging decomposition theory and extracting symbolic-analytical expressions that map inputs to outputs, KANs enable unique insights, sensitivity analysis, and visualisation of power system stability boundaries. We compare KANs against neural networks, in the modified IEEE 9-bus, 39-bus and the Texas 2000-bus systems. Validation results demonstrate that KANs consistently achieve significantly higher accuracy with fewer trainable parameters, by at least an order of magnitude compared to neural networks, while also providing transparent functional representations of relationships between features and model outputs—a capability lacking in existing ML and IML.