A Comparative Analysis of Model and Measurement Uncertainties in Neural Network-Based Transformer Health Estimation

The reliability of AI-based condition assessment in power systems is often limited by the presence of multiple uncertainty sources, stemming from both measurement processes and model behavior. This work presents a comprehensive framework for quantifying and comparing these uncertainties in transformer diagnostics. A Neural Network (NN) model is developed and employed to evaluate how measurement uncertainty-arising from typical sensor and laboratory errors in dissolved gas and oil analyses-and model uncertainty-caused by variations in initialization, data partitioning, and training duration-affect the stability of diagnostic outputs. The Health Index (HI) of power transformers is adopted as a case study to demonstrate the approach. The analysis reveals that both uncertainty types can propagate to the final HI estimation and its qualitative health classes, with potential class transitions driven by stochastic effects rather than real condition changes. Results obtained from a publicly available transformer dataset confirm that the proposed framework enables systematic, interpretable, and reproducible assessment of uncertainty in AI-driven diagnostics, offering valuable insights for the reliable deployment of machine learning models in power asset management.