Explainable Machine Learning: A SHAP value-based approach to locational frequency stability

The increasing complexity of modern power systems due to the high integration of Converter Interfaced Generation (CIG) challenges the effectiveness of current analytical frequency approaches, leading to instability risks and a diminished understanding of local frequency dynamics. To address this, we propose a Machine Learning (ML)-based technique, through Artificial Neural Network (ANN) to capture the frequency characteristics of the system at a local level and SHapley Additive exPlanations (SHAP), an additive feature attribution method, to enhance the understanding of the frequency dynamics. The proposed method further leverages these insights to inform system optimisation models for secure generation dispatch. Validation results from time-domain simulations conducted on a modified version of the IEEE 39-bus network indicate that the proposed method can accurately identify important system variables shaping the local and global frequency stability boundaries, and simple rules can be derived to guide the system optimisation for enhanced system security. • In power systems with renewables, frequency dynamics exhibit stronger local trends. • The proposed method uses machine learning to capture locational frequency dynamics. • Explainable machine learning uncovers critical parameters influencing stability . enabling targeted, actionable interventions to enhance stability both in terms of metrics of interest and locations. • Explanations enable targeted, actionable interventions to enhance stability. • Actionable insights can enhance stability in over 90% of our test cases.