Ensemble deep learning framework with wrapper-based feature selection and sequential encoders for PMU-based event classification

The precise classification of power system events utilising Phasor Measurement Unit (PMU) data is essential for improving grid reliability and security. However, obstacles like high-dimensional data and class imbalance impede efficient event classification. The proposed work introduces a novel methodology named Wrapper-based Feature Selection with Sequential Encoders (W-FUSE) integrated with an ensemble Deep Learning (DL) model for event classification. Publicly accessible PMU datasets from MSU-ORNL were employed, and an appropriate dataset was chosen for its equitable event distribution. Data preparation was carried out to enhance the quality of model inputs, and the class imbalance was mitigated using the Synthetic Minority Over-sampling Technique (SMOTE), and relevant features were selected through the W-FUSE method. A hybrid Transformer model was initially developed, which achieved a test accuracy of 86.80 %. Subsequently, performance was significantly improved by proposing a novel cascading DL architecture integrating a CNN-GRU-Transformer model, leading to an enhanced test accuracy of 92.67 %. The performance was further enhanced through an ensemble model integrating the CNN-GRU-Transformer cascaded model with XGBoost, resulting in a test accuracy of 97 %. Comprehensive evaluations, including confusion matrices, classification reports, and SHAP-based interpretability analyses, confirmed the effectiveness of the proposed model. These results demonstrate that the proposed W-FUSE framework, combined with the XGBoost-based ensemble CNN-GRU-Transformer cascaded model, provides a robust and interpretable solution for accurately classifying power system events using PMU data. • Establishment of a benchmark Deep Learning (DL) model for cyber-physical event classification in power system networks using PMU data. • Proposed novel W-FUSE feature selection, which identifies the most predictive features from datasets and class imbalance removal using SMOTE. • Hybrid Transformer DL with multi-head attention, feedforward transformer blocks, and residual connections classifies events accurately. • A cascaded CNN-GRU-Transformer model is developed for deep feature learning and sequential attention, then ensembled with XGBoost. • Application of Explainable AI Techniques (SHAP Analysis) to ensure model interpretability and transparency.