A Novel Transfer Learning Method for Sub/Super-Synchronous Oscillation Mode Identification in Power Systems

Abstract

With the increasing integration of renewable energy in power systems, accurate and rapid mode identification of the sub/super-synchronous oscillation (sub/super-SO) is essential for ensuring grid security. This article presents a deep-learning (DL)-based method for early prediction of the dominant frequency in sub/super-SO events. To address the challenge of limited sub/super-SO samples for DL training, we propose a domain-adversarial neural network-based transfer learning framework, which leverages easily obtainable forced oscillation data to capture sub/super-SO features. Our method introduces several enhancements over previous approaches in both model structure and optimization. First, an encoder-decoder module with a mask component is designed to reconstruct missing data from field sub/super-SO events, utilizing a specialized reconstruction loss function. Second, the proposed dual-predictor configuration applies supervised learning in both source and target domains, imposing stronger optimization constraints. Furthermore, the proposed method exhibits generality and robustness when the topologies of target and source wind farms differ, highlighting its potential for practical applications in power systems.