Hybrid Continuous Wavelet Transform and GoogLeNet Framework for Accurate Classification of Power Quality Disturbances

Abstract

Accurate classification of power quality disturbances (PQDs) is critical for maintaining grid stability amidst the increasing integration of renewable energy sources. However, traditional feature extraction methods and standard Convolutional Neural Networks (CNNs) struggle with non-stationary signals due to fixed-size convolutional kernels that cannot simultaneously capture features at multiple temporal and spectral scales. To address this limitation, this paper proposes a hybrid framework integrating Continuous Wavelet Transform (CWT) with the GoogLeNet (Inception v1) architecture. The method converts one-dimensional voltage waveforms into two-dimensional time-frequency scalograms, which are then processed by GoogLeNet's Inception modules—featuring parallel 1 × 1, 3 × 3, and 5 × 5 convolutional pathways—to extract multi-scale features simultaneously. Extensive experimental validation on a balanced dataset of 2,100 simulated samples across seven disturbance types demonstrates robust performance, achieving a mean classification accuracy of 90.95% ± 1.60% over 10 independent trials, with best-case performance at 93.29%. Notably, frequency-domain disturbances (Harmonics and Oscillatory Transients) attain perfect classification (100%, σ = 0%) across all trials. These results demonstrate that the proposed CWT–GoogLeNet framework effectively addresses the multi-scale feature extraction challenge, demonstrating reliable statistical performance for automated power quality monitoring in modern smart grid applications.