Exploration of Graph Multi-Attention Neural Network Approaches for Assessing the Transient Stability in Power Systems

Artificial intelligence (AI) techniques are now widely used in power system stability diagnosis. Current AI methods for stability assessment mainly rely on large datasets containing fault scenarios. However, traditional graph neural network methods are often implemented in a way that ignores the intrinsic correlations within nodes and the interactions between geographically separated nodes. To address these limitations, researchers have proposed the graph multi-attention neural network (GMANN). The framework first captures the interdependencies among various state variables associated with a single node through a self-attention mechanism. Subsequently, the long-range associative attention mechanism is introduced to model the interaction dynamics between geographically distant nodes in the grid structure. By combining the long-range correlation attention with the neighborhood correlation attention inherent in graph attention networks, a composite global attention mechanism is formed to effectively reflect the overall importance of different segments of the power infrastructure. In addition, a channel attention mechanism is introduced to assess the applicability of different pooling strategies in graph convolutional networks, thereby assisting in assessing the relevance of different convolutional operations. Finally, global and channel attention are embedded into the underlying graph attention neural network through a multi-source attention fusion strategy. The performance validation of GMANN uses a simulated fault scenario based on the New England 10-machine, 39-node system, with an accuracy of 97.34%. The results show that the GMANN performs better compared to other techniques, especially in terms of significantly lower false-negative rates for key stability metrics, thus highlighting its enhanced reliability.