Abstract

Insulators are essential for the secure and uninterrupted functioning of high-voltage transmission lines. However, since insulators are exposed to the outdoor environment, they are inevitably affected by environmental conditions such as icing. Accumulation of ice on insulator surfaces adversely impacts insulation efficacy and elevates surface leakage currents, resulting in power outages. This research presents a spectrogram-based convolutional neural network (CNN) model for identifying icing conditions on the surfaces of glass, porcelain, and silicone insulators. Insulators are labeled in three classes under laboratory conditions: ice-free, slightly iced (t < 12 mm), and iced (t > 20 mm). High voltage was applied at three distinct levels ranging from 10 to 50 kV, considering the icing conditions of each insulator, and leakage current signals were recorded. The Butterworth and smoothing filters were first applied to the leakage current signals, which were then transformed into spectrogram images using the Fourier transform and used as input for the created CNN architecture. Additionally, spectrogram images were also applied to AlexNet, GoogLeNet, and ResNet-50 architectures. The suggested CNN architecture attained an accuracy of 97.78% to 100% across all operating situations for glass and silicone insulators while demonstrating a classification success rate of 82.22% to 100% for porcelain insulators. Experiments indicate that the accuracy rates of established models in the literature (AlexNet, GoogLeNet, and ResNet-50) diminished to as low as 73%, particularly in porcelain insulator data, thereby validating the developed model’s proficiency in differentiating during icing detection processes and its adaptability to varying conditions.

IPC Classification

Keywords