Conventional strategies for wind turbine blade icing mitigation incur substantial financial burdens, and icing incidents present ongoing operational hazards for wind farms. To improve the precision of blade icing forecasts, this research introduces the Graph Temporal Attention Network (GTAN) model, a data-driven alternative. This model incorporates a Feature Extractor module specifically designed to amplify the distinctions between varied categories of unprocessed sensor readings. Furthermore, it integrates a Temporal Attention (TA) mechanism to enhance sensitivity towards temporal dynamics. Comparative baseline experiments, utilizing supervisory control and data acquisition (SCADA) data from three wind turbines, demonstrate the superior performance of this method relative to other baseline networks in the domain of multidimensional time-series classification. Moreover, rigorous ablation and robustness analyses affirm the effectiveness of the model’s constituent components and highlight its robustness. Notably, the application of a specialized loss function, custom-tailored for the class imbalance inherent in wind turbine blade icing datasets, yields a marked enhancement in the predictive accuracy for minority classes. In conclusion, this data-driven model exhibits a strong capability for accurate blade icing prediction, thereby offering a pathway to diminish maintenance expenditures within wind farm operations.