A hierarchical power system transient stability assessment method considering sample imbalance

Abstract

In order to make full use of the dynamic information contained in the electrical quantity response trajectories, improve the reliability of critical sample prediction results, and correct the bias problem caused by sample imbalance to model prediction, a transient stability assessment (TSA) method based on bidirectional long short-term memory (BiLSTM) network is proposed. The method takes the dynamic trajectories of the underlying measurement data as the inputs, abstracts the features step by step from the multivariate time series through the multilayered neural network, and then establishes the nonlinear mapping relationship between the input feature and the stability category. In this paper, BiLSTM is improved by introducing truncation threshold and penalty coefficient into the loss function to give higher weights to hard samples and unstable samples, thus optimizing the gradient descent direction. Furthermore, this method enables sustainable hierarchical prediction and effectively reduces uncertain samples. The experimental results on the New England 39-bus system integrated with wind farm show that the proposed method significantly reduces the missing alarm rate of unstable samples and the false alarm rate of stable samples, and improves the credibility of the prediction results of critical samples.

Keywords

• Transient stability assessment
• Deep learning
• Power systems
• BiLSTM
• Sample imbalance
• Loss function