Abstract: Machine learning algorithms are widely applied in extracting key featuresoftransient stability in power systems, but their poor interpretability has become a key constraint. Therefore, this paper proposes a key feature extraction and interpretability analysis method based on the improved modified light gradient boosting machine model optimized by Optuna. This method introduces the Optuna automatic hyperparameter tuning algorithm to automatically and optimally adjust the hyperparameters of the model, thereby enhancing the performance and speed of model optimization. A hybrid attribution analysis framework based on SHAP and LiME is proposed. The global feature importance ranking is obtained through SHAP values, and LiME is used to conduct local analysis on individual samples for verification. The test results on the IEEE 10-machine 39-node system show that the proposed method has higher evaluation accuracy and faster evaluation speed, and can provide reasonable and effective explanations for the extracted key features.

Key words: modified light gradient boosting machine, transient stability assessment, interpretability, feature selection