Detection and Classification of Power Quality Disturbances with Feature Extraction and Machine Learning

The detection and classification (D&C) of power quality disturbances (PQDs) are critical for ensuring the reliability, efficiency, and safety of modern electrical power systems. Prompt identification of these disturbances facilitates the swift implementation of protective measures, thereby preventing equipment damage and minimizing system downtime. This study presents a comprehensive framework for PQD D&C using an extensive 21-class dataset that incorporates both single and multiple disturbances, such as voltage sags, swells, harmonics, and interruptions, synthesized under a wide range of noise conditions (20–60 dB). The proposed methodology employs a dense feature space of 30 parameters,, including novel applications of periodograms and modified periodograms, extracted from windowed time-series signals. These features were utilized to train and optimize four machine learning algorithms within a rigorous hyperparameter search space. Based on 10-fold cross-validation, the Extra Trees (ET) algorithm achieved the superior performance with an accuracy of 96.97%, followed by Decision Trees (DT) at 92.83%, Multi-Layer Perceptron (MLP) at 83.31%, and Logistic Regression (LR) at 66.45%. The results demonstrate that the proposed ET-based approach offers high accuracy and robustness in noisy environments while maintaining a low computational load for real-time monitoring applications.