Achieving accurate prediction of transmission line icing thickness is crucial for enhancing the power grid's ice prevention capabilities. To address the issue of insufficient accuracy in predicting line icing thickness, caused by the reliance on manual experience for selecting meteorological factors and the significant nonstationarity of the data, a novel prediction model based on feature attention and northern goshawk optimization (NGO)–variational mode decomposition (VMD)–long short-term memory (LSTM) is proposed. First, the feature attention mechanism is employed to quantify the correlations between seven types of meteorological factors such as wind speed and temperature with icing thickness, enabling adaptive screening of key features. Second, the VMD is utilized to decompose the highly nonstationary icing thickness data into a series of components with different frequencies but local stationarity. Simultaneously, the NGO algorithm is applied to optimize the hyperparameters of VMD, ensuring its decomposition performance. Then, a LSTM network is used to predict the different components. To validate the model's accuracy and generalization ability, comparative experiments were conducted using actual data from Henan and Heilongjiang provinces. The results demonstrate that the proposed model improves prediction accuracy by at least 6.03% and 5.62%, respectively, compared to other methods. Finally, to highlight the model's engineering application value, a practical application at the Xinmi transmission line base in Henan Province showed that the reliability accuracy calculated from the predicted icing thickness values of our model exceeds that of a gated recurrent unit model by 64%. This provides a solution for power grid ice prevention with both theoretical value and practical engineering significance.
Zhang et al. (Sun,) studied this question.