Search

• Data-driven time-domain protection method based on Support Vector Machine. • Machine learning-based protection scheme implemented in hardware. • Energization events, internal and external faults real-time classification. • Proposed method was scarcely affected by CT saturation. • External-to-internal evolving fault identification. This paper presents a power transformer differential protection scheme based on support vector machines (SVM) combined with high-frequency features extracted with the real-time boundary stationary wavelet transform (RT-BSWT). SVM models are derived with synthetic data, considering a wide variety of events, such as inter-turn faults, external faults during CT saturation, and evolving external-to-internal faults. A comparative performance assessment is carried out considering accuracy and other reliability indices, as well as operating time, and good results were achieved. The simplicity of the presented SVM-based relay, without hard-to-derive parameters, built on the classical differential protection framework, highlights potential aspects towards real-life implementation. [ABSTRACT FROM AUTHOR]