Power System Dominant Instability Mode Identification Based on Improved Multi-Head Attention and CNN

Ensuring the regular operation of the power grid and maintaining energy supply relies on the transient stability of the power system. Formulating emergency control strategies requires identifying the system's dominant instability mode. This paper introduces a method for rapid and accurate identification of the dominant instability mode, employing an improved multi-head attention mechanism and convolutional neural network (CNN). Constructing a characteristic matrix reflecting the power system's operation state is the first step, followed by building a deep neural network that combines the improved multi-head attention mechanism and CNN. CNN can extract the feature information of transient electrical quantities with the most significant correlation with the dominant instability mode. Combined with the multi-head attention mechanism, it focuses on the critical information in different positions and different characterization subspaces of the input data, learns the global features in electrical quantities, and introduces the deep residual network structure to capture long-term dependencies. The weighted cross entropy loss function enhances the training effectiveness with unbalanced samples. Subsequently, the proposed method undergoes analysis using a modified IEEE-39 node test system example. The obtained results confirm the accuracy, effectiveness, and interpretability of the method in identifying the dominant instability mode of the power system.