A model-free method to detect the risk and locate the sources of sub-synchronous oscillations in a large-scale renewable power system

• A model-free SSO locating strategy is proposed only based on the training data sets collected from practical power system. • A semi-supervised learning method is applied to enhance the representation ability of NN-SVDD for locating SSO sources. • Case studies carried out in this paper demonstrate the efficiency and effectiveness of DSLS. Sub-synchronous oscillations (SSOs) are serious threat to the stability of power systems integrated with renewable generation, and locating SSO sources is important to counteract SSOs. However, model-based methods are inefficient in this task when the parameters of renewable generation are unknown. Measurement-driven machine learning-based approaches have emerged as alternatives, but face the challenge of unbalanced training data in power systems, where non-oscillation datasets outnumber oscillation datasets. Although the transfer learning (TL) algorithm helps mitigate the issue, it requires certain power system parameters to reconstruct the source domain. Subsequently, incorrect parameters limit the applicability of TL. To solve the problem, this study proposes a novel model-free method to detect SSO risk and locate the source. The proposed method applies the deep learning support vector data description and label spreading approaches. The training process of the proposed method only requires datasets measured from the power system with a few SSO source labels. Thus, the method is truly model-free and requires no parameters of the power system. Case studies are presented in this paper to demonstrate and validate the performance of the proposed method.