Adaptive State Estimation for High‐Renewable Power Systems Via Physics‐Guided Graph Neural Networks

To tackle the strong stochasticity and non‐Gaussian measurement noise introduced by high‐penetration renewable integration, the unobservability caused by insufficient Phasor Measurement Unit (PMU) coverage, and the heavy online computational burden of traditional Time‐Synchronized State Estimation (TSSE) as well as the sensitivity of existing learning‐based models to topology changes and real‐time measurement loss, this paper proposes a physics‐guided Graph Neural Network (GNN)‐based topology‐adaptive state estimation method. The approach embeds power‐system physical constraints and statistical priors into the geometric learning pipeline, enabling high‐speed and robust estimation without iterative computations or matrix inversions. First, a Gaussian Mixture Model (GMM) is employed to characterize missing features arising from limited PMU deployment, and the expected activations are computed in the first GNN layer to intrinsically address unobservability while enhancing adaptability to renewable‐induced randomness. Second, a topology‐adaptive architecture combining a Graph Convolutional Network (GCN) and a Multi‐Head Graph Attention Network (MH‐GAT) is constructed, and an upper bound on the state‐estimation error with respect to topology perturbations is derived, allowing the estimator to accommodate line outages and PMU failures without retraining. Finally, simulations on the IEEE 118‐bus system demonstrate that, the proposed method achieves higher accuracy and lower online latency under complex operating conditions. © 2026 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.