Comparative Analysis of CNN and Hybrid CNN-LSTM Models for PQD Classification with Explainability Insights

Power Quality Disturbances (PQDs) must be accurately detected and classified to preserve stability, efficiency, and power reliability in modern power systems as nonlinear loads and renewable energy sources become more common. Despite the excellent classification accuracy of Deep Learning (DL) models, their computational complexity and implementation feasibility are often overlooked. This work uses time-frequency representations obtained from the Fourier Synchrosqueezing Transform (FSST) to assess several DL architectures for PQD classification. To assess performance and efficiency, the standalone Convolutional Neural Networks (CNNs) and CNN–recurrent models are analyzed. With DenseNet-based architectures achieving the best performance, the analyzed models exhibit high, closely spaced classification accuracies, typically ranging from 98% to 99.4%. The study underlines the importance of evaluating both computational performance and efficiency, which also shows that such slight accuracy gains often come at the expense of significantly greater computational complexity. In addition to classification accuracy, this study also highlights model explainability. Gradient-weighted Class Activation Mapping (Grad-CAM) is used to show discriminative regions in FSST representations, which sheds light on the DL models’ decision-making process.