Spatial Prediction of Daily Maximum Equivalent Ice Thickness on Transmission Lines Based on Multi-Source Data Fusion

Accurate estimation of transmission-line icing is important for regional power-grid risk assessment, particularly in mountainous areas characterized by complex terrain and strong spatial variability. This study develops a multi-source data fusion framework for spatial prediction of daily maximum equivalent ice thickness on transmission lines in Guizhou Province, China. Meteorological variables, topographic factors, remote-sensing information, and geographic coordinates were integrated with operational ice-thickness observations from 160 monitoring stations. Three tree-based ensemble models, namely Random Forest, XGBoost, and LightGBM, were evaluated using a station-based spatial hold-out strategy in which all observations from the test stations were completely excluded from model development. LightGBM achieved the highest R2 (0.8131) and the lowest RMSE (2.5162 mm), whereas XGBoost achieved the lowest MAE (1.0244 mm), indicating comparable spatial generalization performance between the two gradient-boosting models. Stratified error analysis showed that prediction errors increased substantially for extreme icing observations above the 90th percentile, and all three models exhibited negative mean bias errors in this range, indicating systematic underestimation of severe icing. Gain-based feature importance and SHAP analysis showed that temperature-related variables, geographic location, and elevation provided the main predictive information for the fitted model, while their contributions should not be interpreted as direct evidence of physical causality. The optimized LightGBM model was further applied to 1 km gridded environmental predictors to generate a spatially continuous daily icing map, demonstrating pronounced spatial heterogeneity across Guizhou. These results indicate that the proposed framework can support regional-scale estimation of daily maximum equivalent ice thickness at previously unseen monitoring locations, while further improvement is required for extreme icing prediction and transfer to other regions or time periods.