The ongoing transition of power systems from conventional centralized architectures to high-penetration renewable energy integration introduces two prominent challenges: inaccuracies in state estimation due to minute-level topological changes and increased security risks arising from intermittent energy fluctuations. To address the mismatch between fixed-topology assumptions and real-time grid conditions, this study proposes a hybrid modeling framework based on a spatial-temporal graph neural network (STGNN) integrated with dynamic weighted attention (DWA) to capture spatio-temporal feature coupling under highly variable conditions. Compared with traditional graph convolutional networks (GCN), the proposed approach achieves a 52% improvement in prediction accuracy. Furthermore, a dual-driven constrained propagation mechanism incorporating physics-informed neural networks (PINN) is constructed to ensure strict compliance with all security constraints. Finally, a spatio-temporally decoupled parallel training architecture is designed, reducing power fluctuation prediction error from 18.7 kW to 9.2 kW. Experimental analysis on an Institute of Electrical and Electronics Engineers (IEEE) test system with 45% renewable energy penetration demonstrates that the proposed method achieves state estimation delays within 50 ms under complex scenarios and reduces the N-1 contingency response time from 120 ms to 5.8 ms. This research provides a real-time, robust analytical tool for enhancing grid resilience under short-term load demands, offering technical support for power dispatch decision-making systems.
Ma et al. (Wed,) studied this question.