An End‐To‐End Large Language Model and Graph Neural Network Framework for Unified Load‐Shedding‐Based Power System Security and Stability Assessment

ABSTRACT

With the increasing penetration of renewable energy and power electronic devices, power system security and stability assessment is challenged by heterogeneous data sources and fragmented evaluation criteria. To address this issue, this paper proposes a unified security–stability assessment framework by integrating a large language model (LLM) and a graph neural network (GNN). A unified security–stability score is defined based on the expected weighted load‐shedding ratio, which considers N  – 1 contingencies, equipment failure probabilities, renewable‐induced weak‐grid effects quantified by short‐circuit ratio (SCR) constraints, and load priority. The designed LLM is used to standardize non‐standard and heterogeneous operational data, achieving a data standardization accuracy of 97.49% while significantly reducing workflow design time. On this basis, a dual‐head GraphSAGE model is developed to predict node‐level and edge‐level security scores. Case studies show that the proposed SCR‐based and weighted scoring method provides a more conservative and realistic assessment of renewable‐related risks. Moreover, the dual head GNN achieves high prediction accuracy for node and edge scores, respectively, outperforming single head architectures. The results demonstrate that the proposed framework enables fast, accurate, and interpretable security–stability assessment for renewable‐rich power systems.