Power transmission line ice-induced faults often occur in micro-terrain and micro-climate areas, and accurate short-term prediction of ice coating on transmission lines in this region can help guide production and operation. Currently, there are various monitoring data such as tension, fiber optic, image, and micro-climate. However, due to data structure differences and other issues, it is difficult to integrate these monitoring data to predict the future trend of ice coating. Therefore, this paper proposes an ultra-short-term ice coating forecasting method on transmission lines based on two-phase ensemble learning by integrating multi-source monitoring data. For series monitoring data, an ice coating series prediction model based on a long short-term memory network (LSTM) is constructed. For ice-coating image data, an image ice-coating prediction model based on ResNet18 is constructed. Then, the first-phase ensemble prediction model for multi-source monitoring data based on the weighted average method is built. Finally, to improve the robustness and accuracy of prediction, a second-phase ensemble strategy based on stacking ensemble is proposed. The ultra-short-term prediction results are similar to the actual on-site monitoring results, with a root mean square error (RMSE) of 0.37 mm, which proves the effectiveness of the proposed method.