Self-Adaptive Physics-Informed Neural Network for Power Transformer Insulation Paper Degradation Based on Emsley's Model

This paper presents a predictive model based on a physics-informed neural network (PINN) framework to estimate the degree of polymerization (DP) in power transformer insulation paper. The PINN algorithm integrates two fundamental objectives: aligning predicted DP values with actual measurements through data-driven optimization and maintaining consistency with the mathematical model governing DP degradation, known as Emsley's model. Based on Emsley's model, the relationships between the model output (predicted DP value) and the three input variables (historical DP value, its corresponding time, and the prediction time) are formulated as a set of partial differential equations and incorporated into the proposed PINN framework. To enhance prediction accuracy, the training process is adaptive, dynamically adjusting the weighting factor of each objective within the loss function, resulting in a self-adaptive PINN (SAPINN). Additionally, this paper introduces a regional collocation point generation strategy in SAPINN to improve data mapping in periods during which industrial data is not available. The proposed SAPINN is trained and tested on industrial data collected from 83 transformers, each with an average of seven data points recorded annually over seven years. The obtained results indicate that the proposed SAPINN model achieved an average mean absolute error of 15.7, compared to 17.4 and 18.8 obtained using the standard PINN and a basic neural network, respectively. Furthermore, the accuracy of the proposed model is compared to that of multiple machine learning models, PINN with the basic Emsley's model, and the theoretical model, where the proposed SAPINN demonstrates greater reliability and robustness.