Interpretability of Power System Transient Stability Assessment by Deep Learning

In the context of increasing randomness and complexity in power systems, conventional methods for transient stability assessment, which rely on mechanism modeling and simulation, are insufficient for current needs. Data-driven AI methodologies, characterized by their ability to facilitate intelligent state identification and control strategy generation through offline training and online applications, have garnered significant attention. To address these challenges, we employ Random Forests to build a data-driven model for power system transient assessment and classification. We then establish an interpretable analysis framework using SHAP (Shapley Additive explanations) within the Machine Learning model. This framework quantifies the marginal contributions of input features to model outputs, enhancing interpretability. By analyzing features from both global and local interpretability perspectives, we achieve comprehensive insights: global interpretability ranks feature importance, while local interpretability elucidates the contributions of each feature to the predictions of individual samples. This dual-perspective analysis facilitates timely control strategy formulation by dispatchers. Experimental results demonstrate that SHAP effectively clarifies the contributions of features to classification outcomes, providing a clear and interpretable understanding of model behavior.