Abstract: In response to the current challenges of weak spatial specificity and significant prediction uncertainty in overhead transmission line ice thickness forecasting, this paper proposes a probabilistic prediction model that integrates mechanism and data, incorporating freezing rain zonation. First, based on the the fifth generation of ECMWF atmospheric reanalysis of the global climate (ERA5) dataset, the light gradient boosting machine (LightGBM) is employed to achieve large-scale identification and zonation of freezing rain severity. Second, for the identified high-risk segments, a mechanism-data fusion probabilistic prediction model for ice thickness is proposed. This model utilizes a long short-term memory (LSTM) network as a temporal feature extractor and a Gaussian Process as a probabilistic output layer. By embedding physical constraints into the loss function, it achieves probabilistic prediction of ice thickness with uncertainty quantification. Finally, an empirical analysis is conducted using the freezing rain event in Zhejiang Province in February 2024 as a case study. The results demonstrate that the proposed model can effectively identify severe icing areas. While improving the prediction accuracy of ice thickness, it also quantifies the uncertainty of the prediction results, thereby providing a scientific basis for risk perception and prevention decision-making in grid ice disaster management.