.png){kind=icon}
.png){kind=icon}
Return
Quantifying insulation aging risks in power transformers: A spatially selective Bayesian digital twin framework
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ress.2026.113323.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Safety-critical energy infrastructures require high-fidelity digital twins (DTs) for reliable dynamic risk assessment. However, conventional state estimation faces trade-offs between computational efficiency and uncertainty calibration accuracy under degraded sensor data. To address this challenge, we propose a spatially selective Bayesian inference framework for real-time uncertainty quantification (UQ) in inverse heat transfer problems (IHTPs). By dynamically decoupling the physical domain into localized region-of-interest (ROI) subgraphs via graph neural networks (GNNs), the framework eliminates redundant far-field computations and isolates measurement anomalies. Validated on a million-node converter transformer, it compresses MCMC latency from 3631.44 s (restricted to a single serialized chain under the full mesh baseline) down to 217.23 s for a 6-chain parallel 5000-step execution, while maintaining a minimal memory footprint of 332.50 MB. Under extreme overloads and high measurement noise, the framework achieves stable parameter tracking and well-calibrated posterior uncertainty estimation. Furthermore, resolved uncertainty profiles are rapidly propagated for full-field thermal reconstruction and solid insulation aging prognostics. The results quantify heavy-tail degradation risks, providing grid operators with rigorous probabilistic safety margins to support proactive Condition-based Maintenance Plus (CBM+) and dynamic load rating (DLR) decisions.
Keywords:
Digital twin
Uncertainty quantification
Bayesian inference
Graph neural networks
Risk assessment
Inverse heat transfer problem
Journal
R
IF:
11
Papers:
813
Citations:
0
Organization
No organization information available