A novel hybrid mathematical deep learning technique for early warning of flashover in composite insulators.

This paper presents a novel approach based on a hybrid mathematical deep-learning model for flashover prediction of the polluted composite insulators. The proposed prediction method has been designed according to critical pre-flashover conditions, including the dry-band arcing (DBA) and flashover (FO) stages. The feature achievement is developed based on weighted continuous wavelet transform (WCWT) analysis that can discharge counting as a performance index according to real sample test leakage current (LC) data. It is shown that the application of the presented WCWT is superior to that of the typical wavelet analysis. Then, flashover prediction is achieved using the combination of a novel graph attention network (GAT) based on extracted indexes of the discharge peak number and power fitted on a Gaussian function. The developed prediction model can be used as an early-warning tool to prevent outages due to insulator failures. Validation of the presented model for predicting dry-band arcing and flashover is investigated for several parameters, such as the geometry specification of the samples and pollution levels. In addition, a flowchart is represented for online monitoring applications based on the proposed approach.