Spatiotemporal Adversarial Domain Generalization for Locating Subsynchronous Oscillation Sources Under Unseen Conditions in Large-Scale Renewable Power Systems

Abstract

Subsynchronous oscillations (SSOs) in renewable power systems have emerged as a major challenge, jeopardizing the stability and safety of power system operations. Thus, it is essential to accurately and timely locate SSO sources. Artificial intelligence (AI)-based methods for locating SSO sources have become increasingly popular, existing AI-based methods usually fail in practical applications due to unavailable or insufficient real-world SSO data for model training, and significant distribution gaps in samples under different operational conditions. They also fail to fully utilize the temporal characteristics of oscillations and the spatial topology of the system. Moreover, these methods only focus on locating either negative-damping-SSO or forced-SSO sources. To overcome these limitations, we introduce a novel strategy termed Spatiotemporal-Adversarial-Domain-Generalization (STADG) to locate oscillation sources in both SSO scenarios of real power systems. This method allows the model to train on multi-source domains (simplified-simulation power systems) with sufficient labeled samples, and to be directly applied to an unseen test target domain (real power system) under unknow operating conditions. The proposed approach employs a graph-attention network and a long-short-term-memory network to fully leverage spatial and temporal features of SSOs. Extensive experiments on the modified IEEE-39 and WECC-179 bus systems confirm the effectiveness of the proposed approach.