Inception 1D-convolutional neural network for accurate prediction of electrical insulator leakage current from environmental data during its normal operation using long-term recording

Contamination flashover remains one of the biggest challenges for power grid designers and maintenance engineers. Insulator leakage current contains relevant information about their state so that continuous monitoring is considered the most effective way to prevent contamination flashover. In this work, we attempted to accurately predict insulator leakage current in real time during normal operations based on environmental data using long-term recordings. We first confirmed that the history of environmental data also contained relevant information to predict leakage current by conditional Granger analysis and determined that 20 was the optimal number of previous samples for this purpose. We then compared the performance of typical regression models and convolutional neural network (CNN), when using both current and the last 21 samples as input features. We confirmed that the model with the last 21 samples might perform significantly better. Input features pre-processing by cascaded inception architecture was fundamental to capture the complex dynamic interaction between environmental data and leakage current and significantly improved the model performance. CNN based on inception architecture performed much better, achieving an average R2 of 0.94 +/- 0.03. The proposed model could be used to predict leakage current in both porcelain insulators with or without coatings and silicone composite insulators. Our results pave the way for creating an on-line pre-warning system adapted to individual installations, can anticipate the negative consequences of weather and/or pollution deposits and is useful for designing a strategic high-voltage electrical insulator preventive maintenance plan for preventing contamination flashover and thus increase power grid reliability and resilience.