Ghost-ResDSCNet:Lightweight Residual Depthwise Separable Convolution with Spatial Pyramid Pooling for Power Quality Disturbance Classification

The large-scale integration of renewable energy and power-electronic devices has led to increasingly complex power qualitydisturbances (PQDs), characterized by multi-type superposition, transient behavior, and high-frequency components. Suchconditions challenge grid stability and equipment safety, while traditional feature-engineering-based methods often lackrobustness under strong noise and compound disturbances. This paper presents Ghost-ResDSCNet, a lightweight multi-scaledeep learning network for PQD classification. Voltage signals are transformed into time–frequency spectrograms via short-time Fourier transform (STFT). A residual-enhanced depthwise separable convolution (ResDSC) backbone extracts salientfeatures, with a Ghost module generating low-cost redundant feature maps to reduce parameters and computation. A spatialpyramid pooling (SPP) module further aggregates multi-scale contextual information, enhancing classification performancefor disturbances of varying durations. On a 21-class synthetic PQD dataset with signal-to-noise ratios (SNRs) of 20 dB,30 dB, and 40 dB, Ghost-ResDSCNet achieves accuracies of 98.80%, 99.00%, 98.81%, and 94.11%, using only 0.29 Mparameters—outperforming lightweight baselines such as MobileNetV2 and EfficientNet. Grad-CAM++ visualization confirms itsfocus on physically relevant time–frequency regions, ensuring interpretability. The results demonstrate that Ghost-ResDSCNetoffers an effective balance of accuracy, noise robustness, and efficiency, making it suitable for real-time PQD monitoring insmart grid edge computing scenarios.