Objective Existing transmission line icing thickness prediction methods mostly rely on real-time mechanical monitoring data such as conductor stress and tension, which are difficult to obtain in practical engineering. Without relying on conductor stress monitoring data, a prediction model is developed basd solely on real-time meteorological observation data to explore the nonlinear coupling relationship between meteorological parameters and icing growth and evolution, achieve low-cost and high-precision transmission line icing thickness prediction, and provide support for on-line early warning and risk prevention and control of power grid icing disasters.

Method An Attention-convolutional neural network-gated recurrent unit （Attention-CNN-GRU） hybrid prediction model was constructed by taking conventional meteorological data such as temperature, humidity, wind speed and air pressure as the sole input. Convolutional neural network （CNN） was used to capture short-term fluctuations and mutation detail features of meteorological sequences, gated recurrent unit （GRU） was adopted to describe the long-term time series dependence of icing accumulation and evolution, and the Attention mechanism was applied to adaptively enhance the weight of key ice-inducing meteorological factors. Comparative experiments with multiple time steps were conducted to systematically verify the model performanceunder different time series scales.

Result Verification based on measured transmission line data indicates that the Attention-CNN-GRU model reduces the prediction error by 26.1% and 57.3% respectively compared with the basic GRU and long short-term memory （LSTM） network, and the prediction accuracy and stability are significantly improved. The whole modeling process only adopts real-time meteorological observation data, completely getting rid of the dependence on special sensor equipment for conductor stress monitoring, and effectively cutting down the cost of monitoring deployment, operation and maintenance.

Conclusion The model can achieve satisfactory time-series prediction performance using only real-time meteorological observation data and outperforms traditional single time-series networks. The proposed method features readily available data and strong engineering applicability, enabling support for real-time judgment, hierarchical prevention and control, as well as intelligent decision-making for transmission line icing.