Power Signal Processing and Feature Extraction Algorithms Based on Time-Frequency Analysis

Feature extraction in power signal processing plays a crucial role in accurately identifying and classifying various power quality disturbances. Power signals are often non-stationary and complex, containing both transient and steady-state components, which necessitates the extraction of meaningful features that capture their underlying characteristics. In this process, features are derived from multiple domains—time, frequency, and time-frequency—to ensure a holistic representation of the signal behavior. Time-domain features such as mean, standard deviation, skewness, kurtosis, root mean square (RMS), and entropy help in capturing statistical variations and signal energy fluctuations. Frequency-domain features like Total Harmonic Distortion (THD), spectral centroid, and spectral entropy provide insights into harmonic content and frequency distribution, which are critical for detecting distortions and resonances in the power system. This paper proposes an efficient and intelligent framework for power signal classification using a Stacked Whale Optimization-based Machine Learning (SWO-ML) model. The approach combines robust feature extraction from time, frequency, and time-frequency domains with advanced optimization and classification techniques to enhance power quality assessment. A total of 13 features were extracted, including statistical, spectral, and wavelet-based parameters, from different signal conditions such as normal, fault, transient, harmonic distortion, and load switching. The SWO algorithm was employed to select the most informative 18 features out of the initial pool, significantly reducing dimensionality while maintaining high discriminative performance. The proposed SVM + SWO model achieved a classification accuracy of 96.8%, precision of 96.2%, recall of 95.9%, and an F1-score of 96.0%, outperforming baseline models such as SVM without optimization (90.2%), SVM + PSO (93.1%), and SVM + GA (92.4%). In addition, the training time was reduced to 1.85 seconds, showcasing the computational efficiency of the system. Performance evaluation over 100 training epochs showed stable learning with final validation accuracy reaching 98.2% and a minimal loss of 0.05. The results confirm that the SWO-ML framework is highly effective for intelligent, real-time classification of power signals, offering promising applications in power system monitoring, smart grid stability, and fault diagnosis.