The recognition of wire icing risk levels based on deep learning

To address the challenges in observing wire icing, this study took the deep residual network ResNet34 as the baseline model and optimized it using normalization, dropout techniques, and data augmentation methods. A novel wire icing risk level identification model based on deep learning was proposed. The results demonstrated that the optimized ResNet34 model (ResNet34+) achieved an average identification accuracy of 93.3% for wire icing risk levels across different regions and wire orientations. Additionally, the identification accuracy was notably higher between 8:00−11:00 and 15:00−17:00. During a coexisting freezing rain and supercooled fog event on Lushan, the model achieved an average wire icing identification accuracy of 89.4% and 90.8% on east-west and north-south oriented wires, respectively, indicating good generalizability of the model. The application of this model provided a novel approach for identifying wire icing risk levels.