Uncertainty-aware temperature field reconstruction from sparse samples and sensing via latent-space enhancement for GIS digital twin applications

Efficient and reliable full-state temperature estimation is critical for condition monitoring and digital twin applications of enclosed high-voltage equipment, yet it remains challenging due to sparse measurements and the high computational cost of multiphysics simulations. This paper proposes an uncertainty-aware workflow for full-state temperature estimation of a 550 kV gas-insulated switchgear (GIS) disconnector under contact degradation scenarios, based on a proper orthogonal decomposition (POD) reduced-order representation. A probabilistic forward surrogate enables posterior-driven data augmentation, while a lightweight inverse surrogate is trained to infer POD coefficients from only three temperature measurements on the enclosure surface, followed by full-state reconstruction with uncertainty quantification. Numerical results on the evaluated cases show favorable reconstruction accuracy and noise tolerance, with lower errors than gappy POD under the reported settings despite using fewer sensors. These results indicate the potential of the proposed workflow for digital twin-oriented temperature-field estimation in GIS.