Exploring the potential of a machine learning-based methodology for fault classification in inverter-based resource interconnection lines

• Published in: Electric power systems research 2023-10, Vol.223, p.109532, Article 109532
• Main Authors: Davi, Moisés J.B.B., Oleskovicz, Mário, Lopes, Felipe V.
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Fault diagnosis; Full-converter; Inverter-based generators; Signal processing
• ISSN: 0378-7796; 1873-2046
• DOI: 10.1016/j.epsr.2023.109532

Considering the increasing penetration of Inverter-Based Resources (IBRs) in electrical power systems, and the observed impacts on protection and fault diagnosis functions due to their atypical operational characteristics, this paper proposes a fault classification methodology that presents high efficiency for IBR interconnection lines. The fault classification methodology, which combines Discrete Wavelet Transform (DWT) with Machine Learning (ML) tree or rule-based methods, is presented and evaluated. The generalization ability of this methodology as well as the influence of the number and representativeness of training instances on the percentage of correct fault classifications are scrutinized. The proposed methodology was also evaluated for systems with different voltage levels, noisy signals, and considering variations in fault types, locations, inception angles, and resistances, in addition to different grid short circuit levels and IBR controls. Finally, comparisons are made with state-of-the-art fault classification methods to prove the impact of IBRs on these methods and to highlight the superiority of the proposed methodology. For the case studies, two systems with different voltage levels and typical connection topologies of IBRs to the grid are modeled in the PSCAD software, and several contingency scenarios are simulated. •ML methods are promising for fault classification in IBR interconnection lines.•Wavelet Transform enables the feature extraction of attributes weakly impacted by IBR.•More important than the number of training instances is their representativity.