Integrating Weibull analysis and KAN-ODEs for enhanced flashover prediction in contaminated composite insulators

Abstract

This paper introduces a novel machine-learning method for predicting flashover time (FOT) in outdoor polluted polymer insulators. Due to the unpredictable nature of leakage current (LC), a probabilistic approach utilizing the Weibull distribution is employed to assess insulation failure. Analyzing Weibull coefficients establishes a critical condition for flashover prediction, informed by LC behavior and time-frequency indicators derived from Wavelet and Fourier transforms. After estimating the probabilities of insulation failure through the Weibull model, an artificial intelligence (AI) algorithm based on Kolmogorov-Arnold Network Ordinary Differential Equations (KANODEs) has been developed. This algorithm is significantly efficient in the dynamic modeling of the LC behavior and flashover prediction. By combining these two concepts, the flashover of the insulator was predicted using a new machine-learning technique based on the probabilistic neural network (PNN). The probability section is Weibull analysis, and the predictive section is KANODE. Various insulator types and aging conditions were tested to validate the model, considering factors like wetting rate and pollution intensity. The results demonstrate a strong correlation between the predicted and actual outcomes, indicating that the proposed approach can be an effective online monitoring tool to prevent pollution-related flashovers in outdoor insulators.