Ensuring stable power grid operation requires analyzing large datasets within physical models. The current mainstream Supervisory Control and Data Acquisition (SCADA) systems face challenges due to low data update frequencies, hindering their ability to capture rapidly evolving dynamic processes. To mitigate the delayed response issues of renewable energy sources in high-penetration scenarios, this research enhances the timeliness of topology identification by integrating spatiotemporal features with power system information. The findings can be directly applied to online early warning systems within dispatch centers, reducing the risk of failures caused by network topology errors. This study establishes a Multi-Head Self-Attention (MHSA)-based model capturing the nonlinear mapping between data and topology structures. This model mitigates the impact of data noise and missing values on identification accuracy. Furthermore, we developed a Graph Convolutional Network (GCN) algorithm that incorporates the Hausdorff Distance (HD) to quantitatively assess the potential impact of topology changes on system stability. By deploying a lightweight attention module onto IoT devices, real-time computation at the edge is achieved, meeting the stringent response requirements of power systems. Performance is validated using a test system incorporating a high proportion of renewable energy sources. Compared to traditional methods, the proposed approach demonstrates significant improvements: identification accuracy increases by 7.6% and 10.3% respectively, robustness improves by 5.2%, predictive capability surges by 41.2%, and the false alarm rate decreases by 32.3%. These results confirm the method’s effectiveness, which provides a novel technical pathway for dynamic topology management in power systems.
Liu et al. (Sat,) studied this question.