ResNet-Based system offers new tool for power grid icing risk assessment

By optimizing the ResNet34 neural network architecture through normalization, dropout, and data augmentation, the team achieved an impressive average identification accuracy of 93.3%. The model, known as ResNet34+, performs consistently across various wire orientations and geographic locations, offering strong generalization and reliable performance even in challenging weather conditions.

Wire icing occurs when snow or ice accumulates on overhead wires, posing serious risks such as cable breakage, tower collapse, and electrical faults. Traditional monitoring relies on manual inspections or costly automated equipment that limits scalability. Previous models, including empirical and linear regression approaches, have struggled to capture the complex, nonlinear nature of wire icing, often failing to deliver reliable results across different climatic conditions. Given the limitations of existing methods, there is an urgent need for more adaptable and efficient technologies. In response to these challenges, researchers are turning to deep learning and image-based classification to enhance risk identification capabilities in a cost-effective and scalable manner.

A study (DOI: 10.48130/emst-0025-0006) published in Emergency Management Science and Technology on 28 May 2025 by Shengjie Niu’s team, Nanjing University of Information Science and Technology, could significantly improve early warning systems and maintenance planning in power grid operations.

To evaluate and enhance the accuracy of wire icing risk level identification, researchers constructed and optimized a deep learning model using image data. The dataset was split into training (6,637 images), validation (948 images), and test (1,898 images) sets, with images sourced from Enshi and Shennongjia. Initially, the performance of two convolutional neural network architectures, VGG16 and ResNet34, was compared. Results showed that ResNet34 outperformed VGG16 in both accuracy and loss metrics due to its residual blocks that maintained gradient stability and enabled deeper network layers. However, the original ResNet34 model showed signs of overfitting, with a noticeable performance gap between training and validation sets. To address this, the researchers introduced normalization, dropout, and data augmentation to produce an improved version-ResNet34+. The optimized model significantly reduced validation loss and improved generalization, achieving recognition accuracies of 94.5%, 91.2%, 89.6%, 96.3%, and 96.7% across five wire icing risk levels. Slightly lower accuracy at light risk levels was due to visual similarity with non-icing cases. Further testing across different regions and wire orientations showed that ResNet34+ maintained robust performance, with accuracy exceeding 83% in most settings. Time-of-day analysis revealed optimal recognition between 08:00–11:00 and 15:00–17:00, owing to favorable lighting. In a separate test involving 230 images during a mixed rain-fog icing event at Lushan, the model achieved average accuracies of 89.4% (east-west) and 90.8% (north-south), confirming its adaptability under varied environmental conditions.Importantly, the model does not measure exact ice thickness but provides actionable risk classifications, making it highly valuable for early warning and operational planning.

The ResNet34+ model offers a novel, cost-effective solution for identifying wire icing risk levels using image data alone. It enables power grid operators to detect and respond to potential icing threats more efficiently, reducing reliance on manual inspections and expensive sensor equipment. The model’s ability to generalize across different locations and orientations means it can be widely deployed without extensive regional recalibration. By integrating this technology into existing monitoring systems, utility companies can enhance grid resilience, prevent costly infrastructure damage, and improve public safety during extreme weather events. Though not a replacement for physical measurement devices, this method greatly complements existing early warning and maintenance systems.

###

References

DOI

10.48130/emst-0025-0006

Original Source URL

https://doi.org/10.48130/emst-0025-0006

Funding information

The authors acknowledge the immense support from the National Natural Science Foundation of China (Grant Nos 42075063; 42075066), the Jiangsu Graduate Scientific Research Innovation Project (Grant No. KYCX23_1316), the China Scholarship Council (CSC) (Grant No. 202309040027) and the Project of China Meteorological Administration Training Center (Grant No. 2024CMATCPY06).

About Emergency Management Science and Technology

Emergency Management Science and Technology (e-ISSN 2832-448X) is an open access journal of Nanjing Tech University and published by Maximum Academic Press. It is a medium for research in the science and technology of emergency management. Emergency Management Science and Technology publishes high-quality original research articles, reviews, case studies, short communications, editorials, letters, and perspectives from a wide variety of sources dealing with all aspects of the science and technology of emergency.