In the "two micro" environment, the transmission lines in the power system are threatened by serious ice, which poses a potential risk to the stability of the power grid operation. In order to improve the efficiency of pre-warning operation and maintenance, an innovative Long Short-term Memory-Multilayer Perceptron (LSTM-MLP) model based on the monitoring time series of ice cover thickness of classical transmission lines was proposed. This paper aims to establish a reasonable and reliable ice thickness prediction method to better capture the wide range fluctuation of ice monitoring data of transmission lines. For this reason, the LSTM-MLP model was used to predict and compare the wire operation and maintenance data of different data capacities. The time series data of the ice-covering amount of the conductor was used to predict the ice-covering thickness. Various ice-controlling factors such as temperature, humidity and wind power were introduced to improve the prediction ability of the model on the fluctuation data. In order to further improve the performance of the model, Grey Wolf algorithm was utilized to optimize the model hyperparameters. The results demonstrate that the optimized multivariable LSTM-MLP model has lower Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and higher coefficient of determination (R2) for predicting ground ice thickness of 12 test datasets, which were 1.0765, 0.7455 and 0.8893 respectively. For the predicted results of 30 test datasets, RMSE, MAE and R2 were 0.8814, 0.5238 and 0.9315, respectively. These results indicated closer proximity to the actual monitoring values than the univariate LSTM-MLP model, thus verifying the high accuracy and reliability of the multivariable LSTM-MLP model. In summary, the multivariable LSTM-MLP model effectively capture the fluctuation of transmission line ice cover data, and provides an innovative solution for accurately predicting non-stationary ice cover thickness.