Multivariate Ice-Covering Prediction on Transmission Lines Using CEEMDAN-SC-LSTM and CNN-GRU-Attention Models

Abstract

In the power system, ice-covering on transmission lines in high-altitude areas can easily cause serious problems like increased line weight and wire breakage. To accurately predict the ice-covering and issue timely warnings to ensure the reliable operation, a method for predicting ice-covering in high-altitude areas based on Long Short-Term Memory (LSTM) and deep learning methods is proposed. In the feature extraction stage, a model for ice-covering feature extraction of transmission lines in cold regions based on CEEMDAN-SC-LSTM is proposed, combining spectral clustering, fully adaptive noise set empirical mode decomposition algorithm, and LSTM. This model can utilize the advantages of each algorithm to effectively decompose and cluster ice-covering data, extracting more representative features. A Multivariate Deep Regression (MDR) is built by integrating convolutional neural networks, gated recurrent units, and attention mechanisms, using the extracted features related to ice thickness changes as input. The experiment is based on a dataset collected from a specific main transmission line in a high-altitude mountainous area in western China from October 2024 to March 2025. The results showed that the prediction model had a coefficient of determination of 0.9887, an average absolute error of 0.0130, a root mean square error of 0.0142, and an average absolute percentage error of 1.2558%. Compared with baseline models such as support vector regression and random forest regression; the model can predict ice thickness more accurately. The early warning platform based on this method can effectively deliver ice-covering early warning, providing decision-making basis for power grid operation and maintenance. The warning platform based on this method can effectively deliver ice-covering warnings, providing decision-making basis for power grid operation and maintenance.