<![CDATA[Power grids today operate under unpredictable and rapidly changing conditions, making reliable stability prediction increasingly important. This study evaluates two hybrid learning frameworks that integrate deep feature transformation with ensemble classification. In the first framework, an autoencoder (AE) is used for feature encoding before classification with extreme gradient boosting (XGBoost), while the second applies a TabTransformer (TT) followed by the same classifier. For comparison, conventional ensemble models, including random forest and standalone LightGBM, are also assessed. The models are tested on a large public dataset using stratified cross-validation and standard performance metrics. Results show that the AE-XGBoost hybrid achieves the highest performance, with a test accuracy of 97.73% and an F1-score of 0.98 for both stable and unstable states. LightGBM also performs strongly, offering consistent accuracy (95.8%) and good interpretability. In contrast, TT-XGBoost, despite its architectural novelty, achieves lower accuracy (89.4%) and struggles with unstable states. These findings highlight that model effectiveness depends not only on architectural complexity but also on the synergy between feature transformation and classification. The results provide practical insights for building dependable, confidence-aware predictive systems to support smart grid decision-making.]]>
Copyrights © 2026
