Robust multi-physics field reconstruction via sparse enclosure temperature measurements for GIL Digital Twin application

Fast and reliable field reconstruction from sparse sensing is essential for real-time Digital Twin (DT) models, yet it has been overshadowed by the emphasis on fast-forward simulations within the high-voltage apparatus community, despite both approaches leveraging the low-rank nature of data through Reduced-Order Modeling (ROM). This challenge is further intensified in enclosed equipment like Gas-insulated metal-enclosed transmission lines (GIL), where harsh internal environments and limited sensor accessibility render the inverse problem severely ill-conditioned. This work leverages the Shallow Recurrent Decoder (SHRED) architecture for robust multi-physics field reconstruction in a 550 kV GIL, using only three temperature measurements from the outer surface of the enclosure. As a test case, the temperature and velocity fields during a heating process are reconstructed from noise-polluted measurements, exploiting the temporal dynamics learned by the Recurrent Neural Network (RNN) and the spatial mapping conducted by the Shallow Decoder Network (SDN) embedded in SHRED. Experimental results show that SHRED consistently outperforms traditional Gappy POD methods and SDN across varying noise levels, achieving lower reconstruction errors. Its lightweight design and high reconstruction accuracy make it particularly suitable for real-time DT applications in high-voltage apparatus. • Clarified field reconstruction and fast simulation under a unified ROM framework. • End-to-end lightweight and robust field reconstruction from sparse sensor. • Validated reconstruction accuracy with high-fidelity simulations and experiments. • Outperformed existing methods with lower error under polluted sensor data.