Application and Performance Evaluation of a Physics-Informed Neural Network for Transformer Insulation Paper Degradation Prediction

Long-term operation of oil-immersed transformers degrades the internal insulation paper, making stable insulation performance difficult to maintain. This study evaluates a forward physics-informed neural network (PINN) for predicting the degree of polymerization (DP) of transformer insulation paper using synthetic data generated from the analytical solution of the Emsley model with relative Gaussian noise. PINN, ANN and LSTM share training observations, normalization and evaluation grids, while only PINN receives the governing equation, its true parameters and a noise-free initial condition. Temperature is fixed at 98 °C. With three observations, PINN achieves a median RMSE approximately 1/11.5 that of ANN and 1/54.0 that of LSTM, both trained with fifty observations. Interpolation and extrapolation errors are reported separately, and sensitivity to physical parameter misspecification is evaluated. These comparisons quantify the benefit of exact physical prior information under controlled synthetic conditions; they do not establish the necessity of a neural solver when the analytical solution is known or demonstrate performance on field measurements.