A Deep Imbalanced Learning Framework for Transient Stability Assessment of Power System

Abstract

Maintaining transient stability is a core requirement for ensuring safe operation of power systems. Hence, quick and accurate assessment of the transient stability of power systems is particularly critical. As the deployment of wide area measurement systems (WAMS) expands, transient stability assessment (TSA) based on machine learning with data of phasors measurement units (PMUs) also develops rapidly. However, unstable samples of the power system are rarely seen in practice which affects greatly the effectiveness of transient instability recognition. To address this problem, we propose a deep imbalanced learning-based TSA framework. First, an improved denoising autoencoder (DAE) is constructed to map the training set to hidden space for dimension reduction. Then, adaptive synthetic sampling (ADASYN) is further used to synthesize unstable samples in hidden space to balance the proportion of different classes. Third, the synthesized data are decoded into the original space to enhance the training set. Finally, an ensemble cost-sensitive classifier based on a stacked denoising autoencoder (SDAE) is designed to extract different feature patterns, and the SDAEs are merged with a fusion layer to classify the status of the power system. The simulation results of two benchmark power systems indicate that the proposed method can effectively improve the recognition accuracy of unstable cases by combining nonlinear data synthesis with ensemble cost-sensitive learning methods. Compared with other imbalanced learning methods, the proposed framework enjoys superiority both in accuracy and G-mean.