Physics-Informed Neural Networks for modelling insulation paper degradation in Power Transformers

Power transformer’s insulation is an integral part of the health and performance of this power component. This paper uses Physics-Informed Neural Networks (PINNs) for predicting the lifetime and health indicator of the power transformer’s insulation material, which is expressed as the Degree of Polymerization (DP) of the polymeric material (in this case kraft paper). PINNs are a promising deep learning technique for solving scientific computing problems and are designed to incorporate prior knowledge of physical or chemical systems and to respect any symmetries, invariances, and conservation laws. The dynamics of the degradation process is modeled using ordinary differential equations. One major challenge in analyzing kraft paper degradation is estimating the unknown model parameters (e.g. rate constants) and thus predicting model dynamics. For this work, we aim to solve the data-driven discovery of the degradation process, infer the hidden kinetic parameters and predict the degree of polymerization. The final discussion also addresses the advantages and limitations of PINNs for solving this type of problems.