Interpretable Transient Frequency Stability Prediction Based on Graph Convolutional Learning

To optimize China’s energy structure and improve energy utilization efficiency, an increasing number of large-capacity HVDC along with high-proportion renewable energy sources are being integrated into the power systems. This makes the power systems gradually gain a low-inertia nature. Low-inertia power systems exhibit weak inertia support, high uncertainty in power output, and poor frequency regulation capabilities, leading to high risk of encountering transient frequency stability issues after large disturbances. Therefore, it is of paramount importance to predict the transient frequency stability of a specific system in a reliable and efficient manner. To cope with the limitations of existing data-driven transient frequency stability prediction (TFSP) methods that simply learn from temporal information without sufficient consideration other valuable data like spatial information, this paper proposes an interpretable TFSP approach based on a Graph Convolutional Network via Initial residual and Identity mapping (GCNII). By incorporating the power network spatial information with temporal responsive data acquired from individual buses into the GCNII model, the proposed approach enables the model to comprehensively capture and learn the complex mapping relationships between system spatial-temporal dynamics and transient frequency stability status. Furthermore, the interpretable SHapley Additive exPlanations (SHAP) method is used to interpret the stability prediction results for better understanding. Finally, numerical tests on a realistic regional power grid in China demonstrate the efficacy of proposed method as well as its desirable interpretability.