Deep Learning Based Feature Reduction for Power System Transient Stability Assessment

Abstract

A novel feature reduction method based on deep learning is proposed for power system transient stability assessment (TSA) in this paper. First of all, an original feature set including three parts of features is set up. Secondly, the stacked autoencoder (SAE) is used to extract the multi-level features from raw inputs by its reconstruction weights. Then, a weight analysis method based on mutual information is put forward to optimize the number of hidden nodes in SAE, which can remove the irrelevant and redundant features in each layer of SAE at the same time. Finally, test results demonstrate that the proposed method can reduce the training burden of SAE effectively, and the assessment of transient stability is more accurate by applying the features obtained by the proposed method.