An explainable hybrid deep learning approach for power quality disturbance classification

Abstract

This study presents a deep learning-based framework for the accurate classification of power quality (PQ) disturbances using time-series and environmental data. Four architectures: artificial neural network (ANN), convolutional neural network (CNN), long short-term memory (LSTM), and a hybrid model were implemented and evaluated on a dataset containing key electrical and environmental features such as voltage sag, harmonics, temperature pollution level. Comprehensive preprocessing, including normalization, correlation assessment of skewness and kurtosis analysis, ensured statistical reliability. Among the evaluated models, the hybrid architecture achieved the best performance, with an accuracy of 98.91% and an F1-score of 98.90%, outperforming all standalone approaches. Model interpretability was enhanced using local interpretable model-agnostic explanations (LIME), which identified feature contributions for individual predictions. A comparison with eight recent studies demonstrated competitive or superior performance in both accuracy and explainability. The integration of high-performing hybrid modeling with interpretable AI makes the proposed system well-suited for real-time PQ monitoring in smart grid environments. Future work may incorporate unsupervised and transfer learning methods to improve adaptability across varying grid conditions and data scarcity scenarios.