Multi-task transient stability assessment of power system based on graph neural network with interpretable attribution analysis

Abstract

Transient power angle stability analysis and voltage stability analysis are the key basis of power system security and stability control. The diversified operation modes of power system put forward higher requirements for real-time and topology generalization of transient stability assessment. Based on Convolution Neural Network (CNN) and Graph Attention Networks (GAT), this paper uses the wide-area transient response features to evaluate the transient angle stability and voltage stability of power system. Firstly, CNN is used to encode the temporal variation characteristics of transient features of electrical components in hidden layer space. Secondly, the encoded hidden layer features are mapped into graph data samples according to the spatial topological correlation of electrical components, so as to train and build a multi-task transient stability assessment model based on GAT. The model can adaptively capture the interaction relationship between encoded hidden layer features of topologically correlated components, and realizes real-time evaluation of transient angle stability and transient voltage stability by perceiving the temporal–spatial correlation of transient features. Finally, Shapley additive explanation (SHAP) method is used to explain the decision-making mechanism of the model and extract the key features that affect the transient stability, so as to enhance the ability of power grid operators to perceive the transient stability situation. The results of IEEE-39 system show that the transient stability assessment model has good accuracy and topology generalization, and the results of interpretable analysis are in line with the conclusions of traditional mechanism cognition, which helps to improve the credibility of the model results.