Explainable Deep Neural Networks for Evaluation Status of Transformer Paper Insulation With Multiple Chemical Indicators

Abstract

Chemical indicators are essential for evaluating transformer paper insulation aging. However, relying solely on a single chemical indicator is insufficient for comprehensive and reliable evaluation. To address this problem, the study proposes a method using furfural, methanol, and ethanol to evaluate the aging status of transformer paper insulation. First, molecular dynamics (MDs) simulations identify the key pathways and reaction types during paper insulation pyrolysis. Oil-paper insulation samples are prepared under various aging factors using an accelerated thermal aging test platform. Experiments reveal the change law in multiple chemical indicators content due to insulation degradation. A deep neural network (DNN) is then used to evaluate the aging status, achieving the value of MAPE is 8.1%, validating the model’s effectiveness. Finally, Shapley additive explanation (SHAP) is applied to interpret the contribution of multiple chemical indicators, providing a reasonable explanation for their impact. Research shows that the DNN-SHAP model effectively evaluates transformer paper insulation aging, offering theoretical support for accurate characterization.