Data-driven Method and Interpretability Analysis for Transient Power Angle Stability Assessment

Abstract

With the large-scale grid connection of Renewable energy sources as well as modern power electronics devices, the accuracy and rapidity of transient stability evaluation in power grids are increasingly stringent.The data-driven power system transient stability assessment can achieve online real-time prediction by learning the temporal characteristics before and after faults, but its application is limited by its inherent black-box nature. In this paper, an extreme gradient boosting (XGBoost) model-based transient stability assessment method is proposed for power systems, which can optimize the prediction accuracy while ensuring immediate response. To improve the interpretability of the evaluation results, the Yellowbrick algorithm is used to interpret the evaluation model prediction results from the perspective of feature importance, and the correctness of the interpretation results is verified by making predictive attribution analysis on a single sample based on Local Interpretable Model-agnostic Explanations (LIME) algorithm, which provides a reliable basis for online assessment of grid transient stability.