To improve the accuracy of transmission line icing prediction

this paper proposes an icing prediction method that integrates data processing and deep learning. In the data processing stage

the Isolation Forest algorithm is used to remove outliers from the input data

and the maximal information coefficient (MIC) and variational mode decomposition (VMD) are employed to screen and decompose key features

and to construct a high-quality dataset. The core model adopts a serial-parallel architecture combining long short-term memory (LSTM)

Transformer

and bidirectional long short-term memory (BiLSTM). In the serial part

LSTM is first applied to extract data relationships

Transformer is then used to capture global dependencies

and finally BiLSTM enhances bidirectional temporal modeling. In the parallel part

an additional BiLSTM network is introduced to compensate for potential information loss during serial processing. The study shows that

under micro-terrain conditions such as uplifted terrain

high mountain watersheds

and saddle terrain

the proposed model achieves root mean square errors of 0.09 mm

0.07 mm

and 0.13 mm

respectively. Its performance surpasses that of comparative models

meets the requirements of practical engineering applications

and enables accurate prediction of ice accretion growth

which is of significant importance for ensuring the safe operation of transmission lines.