To address the problem of insufficient icing prediction accuracy in remote areas where no online monitoring systems are deployed and only nearby meteorological station data are available

this paper proposes a transmission line icing level prediction method based on mRMR-GSWOA-XGBoost. Considering the cumulative nature of ice accretion growth

along with the high dimensionality and strong nonlinearity of icing prediction data

a multi-source data feature set is first constructed by integrating meteorological data

geographical data

line characteristics

and icing thickness. Through normalization and labeling

heterogeneous information is effectively coupled. The minimum redundancy maximum relevance (mRMR) method is then used to select features that significantly represent the influencing factors of icing

thereby improving the model's generalization ability and resistance to overfitting. Subsequently

an improved global search whale optimization algorithm (GSWOA) is applied to optimize the hyperparameters of the extreme gradient boosting (XGBoost) model

enhancing its capability to capture strong nonlinear relationships. Finally

a multi-classification model is established to predict icing levels over the next 12 hours. Results show that the proposed method achieves an average accuracy of 90.7% and a maximum accuracy of 90.47% in identifying icing levels

demonstrating its effectiveness and reliability. This study provides a useful reference and modeling approach for transmission line icing prediction without online monitoring systems.