A hybrid CNN-BiLSTM model integrating meteorological and topographic factors for transmission line ice thickness prediction

P P Li, Bing Feng, Zhenyang Fu, Zhou Jian

Transmission-line icing can threaten reliable grid operation, making short-term ice-thickness prediction important for disaster prevention and operational planning. This study develops a late-fusion CNN-BiLSTM regression framework for next-hour ice-thickness prediction on the 500 kV Changmiao 5P01 transmission-line corridor in Huzhou, Zhejiang, China. A 24-h historical meteorological sequence is processed by 1D convolutional layers and a BiLSTM without using observations after the prediction boundary; location-specific static topographic descriptors are concatenated only with the resulting boundary representation before linear regression. The study uses 2,348 continuous hourly records. Inputs include temperature, relative humidity, atmospheric pressure, precipitation, and wind components, together with elevation, slope, and aspect; ice thickness is measured in millimeters. Under the same chronological train/test protocol, the proposed configuration produced lower test errors than the selected standalone CNN, LSTM, and BiLSTM baseline configurations. The results support the predictive usefulness of the terrain descriptors for the studied subtropical hilly corridor, but do not establish causal effects or broader geographic generalization.