A New Comparative Approach Based on Features of Subcomponents and Machine Learning Algorithms to Detect and Classify Power Quality Disturbances

• Published in: Electric power components and systems Vol. 52; no. 8; pp. 1269 - 1292
• Main Authors: Akkaya, Sıtkı, Yüksek, Emre, Akgün, Hasan Metehan
• Format: Journal Article
• Language: English
• Published: Philadelphia Taylor & Francis 08.05.2024; Taylor & Francis Ltd
• Subjects: Algorithms; chi-squared; Classification; cross validation; Datasets; feature selection; flicker; harmonics; hyperparameter optimization; interruption; Machine learning; machine learning algorithms; notch; Parameters; power quality disturbances; sag; spike; swell; time-series signal; transient
• ISSN: 1532-5008; 1532-5016
• DOI: 10.1080/15325008.2023.2260375

Current measurement systems based on the IEEE-1159 standard have some limitations and robustness problems under noisy and fast-changing conditions. Besides, applying different methods for each Power Quality Disturbance (PQD) to every window is required but time-consuming and not feasible. Therefore, different kinds of two-stage methods, Detection and Classification (D&C), have been improved in many studies. Then, the required measurement can be performed to define disturbance. For this purpose, a new approach based on features of subcomponents with Machine Learning Algorithms (MLAs) to detect and classify PQDs is proposed. 21-class dataset including single and multiple PQDs under different noisy conditions was prepared randomly. Of this dataset, determined features were extracted and some of these were selected. Then, selected features were trained and tested with some MLAs in a workstation. Results obtained from comparative MLAs and the other classification methods show that the best MLA with related features is Random Forest with 96.97% while LightGBM, k-Nearest Neighbors, and XGBoost 96.85%, 96.73%, and 92.82% accuracy, respectively. Because the selected features, optimized parameters, and the related MLA were obtained by investigating for features provided from the PQDs in the whole parameter space, this approach brings the advantages of high accuracy, low D&C complexity, and computing load.