Akademik Veri Yönetim Sistemi
IET Renewable Power Generation, cilt.20, sa.1, 2026 (SCI-Expanded, Scopus)
The increasing penetration of renewable energy sources and converter-based distributed generation has significantly intensified power quality disturbance (PQD) challenges in modern distribution systems. Accurate and reliable multi-class classification of disturbances is therefore essential to ensure grid stability and protect sensitive equipment. In this study, an explainable deep learning framework is proposed for multi-class PQD classification using electrical structured descriptors (ESD), which provide physically interpretable representations of electrical signals. Three representative architectures—MLP, GRU and BiLSTM—are systematically evaluated under a unified preprocessing and stratified cross-validation scheme. Model performance is assessed using imbalance-aware metrics, particularly macro-averaged measures. Experimental results demonstrate near-saturated classification performance across all models. The GRU model achieves the best overall accuracy (0.9979) and macro-F1 score (0.9980), while all models reach an identical macro-AUC of 0.9998, indicating excellent class separability. These findings suggest that increasing architectural complexity does not necessarily yield significant performance gains when structured and physically meaningful features are employed. To enhance transparency, explainability analyses based on SHAP and LIME are integrated into the framework. The results reveal that RMS voltage, peak voltage and total harmonic distortion (THD) are the most influential features, aligning with established power system knowledge. The proposed framework provides a balanced solution in terms of accuracy, computational efficiency and interpretability, making it suitable for real-time PQ monitoring in renewable-integrated smart grid environments.