Physics-Guided Cross-Attention Framework for Real-Time Voltage Stability Assessment

Abstract

Driven by the increasing integration of renewable energy, modern power grids frequently operate near their physical stability boundaries, rendering online Voltage Stability Assessment (VSA) critical for secure grid operation. Although purely data-driven deep learning models facilitate rapid margin prediction, they often exhibit limited robustness and generalization capability under unseen operating conditions. To address these limitations, this paper proposes a Physics-Guided Cross-Attention (PGCA) framework for real-time VSA. Unlike conventional physics-informed approaches that incorporate physical constraints through loss-function regularization, the proposed method utilizes physics-guided embeddings derived from Jacobian-based stability indicators. Through a customized cross-attention mechanism, the framework integrates high-resolution temporal phasor measurements with physics-derived structural information, enabling effective interaction between temporal features and system stability characteristics. Extensive experiments on the IEEE 118-bus system and the real-world 1,807-bus Taiwan Power System (TPS) demonstrate enhanced predictive performance, strong scalability, and improved robustness under unseen operating conditions.