Ice accretion prediction of transmission lines based on ICEEMDAN-xLSTM-transformer

This paper proposes a hybrid model for predicting ice thickness that integrates Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN), Extended Long Short-Term Memory (xLSTM), and Transformer. First, ICEEMDAN is utilised to decompose the ice accretion time series data, which is influenced by climate factors such as temperature, humidity, and wind speed, exhibiting uncertainty and nonlinearity. Then, the xLSTM network model is employed to extract features from the multiple Intrinsic Mode Functions (IMF) obtained after ICEEMDAN, enabling xLSTM to learn information across different time scales and establish temporal relationships between the data. Finally, the features extracted by the xLSTM module are fed into the Transformer module, which uses a multi-head attention mechanism to capture data features from different positions and concatenates them. A feedforward network then performs nonlinear transformations, resulting in the Transformer outputting the predicted ice thickness. Experimental results show that the proposed hybrid prediction model achieves better results compared with other methods.