Graph Neural Networks boost current transformer estimation by 38%

Current transformers (CTs) are critical for power system protection and measurement, but transient core saturation can severely distort the secondary current and degrade accuracy. Traditional dynamic state estimation methods rely on numerical discretisation and iterative solvers, but their initialization is not informed by the physical dependency structure of the estimation problem, limiting robustness under noisy conditions.

To address this, the paper presents a physics-informed enhancement for CT dynamic state estimation using COMTRADE measurements generated in WinIGS-T. After benchmarking four discretisation schemes and three iterative solvers—identifying Gauss-Newton with Quadratic discretisation as the strongest baseline—the authors construct a Graph Neural Network from the Jacobian sparsity pattern to generate warm-start initial state estimates. This approach improves estimator conditioning, achieving average gains of 25% in initialization distance and 38% in initial weighted objective value across all tested SNR levels. The results demonstrate that embedding physical structure into initialization significantly improves CT saturation correction, supporting more reliable measurement and protection in modern power grids.