Power System Frequency Response Prediction with Spatial-Temporal Graph Convolutional Networks

The increasing penetration of renewable energy has significantly reduced power system inertia, compromising frequency regulation capabilities and raising critical concerns about frequency stability. In this context, accurate prediction of dynamic frequency responses to power disturbances has become essential. This paper proposes a novel physics-data hybrid approach for frequency response prediction based on SpatialTemporal Graph Convolutional Networks (STGCN). First, a dual-layer architecture utilizes Chebyshev graph convolutions to capture spatial relationships between power system buses while employing temporal convolutions to extract dynamic frequency variation features. Second, a hybrid feature input framework integrates steady-state operating parameters with disturbance information, embedding physical characteristics into the graph structure to effectively represent disturbance locations and magnitudes. Validation on IEEE 39-bus system with wind power integration demonstrates the superior performance of the proposed approach. Compared to existing methods, the developed model achieves lower prediction errors for both frequency nadir values and the occurrence times, while exhibiting enhanced accuracy, robustness, and generalization capabilities across diverse operating conditions and disturbance scenarios.