Transmission line icing presents a critical challenge to power grid operations, significantly increasing mechanical loads and aerodynamic coefficients. Under extreme meteorological conditions and complex terrain, this phenomenon can lead to catastrophic failures including conductor rupture and tower collapse. Reliance on conventional manual monitoring methods proves inadequate for timely ice risk assessment and preemptive de-icing interventions, potentially resulting in widespread blackouts that severely compromise power system reliability. Ice type classification, as a sophisticated fine-grained visual recognition task, demonstrates substantial limitations in current unimodal image-based approaches: these methods exhibit deficiencies in effectively incorporating essential auxiliary data (particularly meteorological parameters) while remaining vulnerable to complex background interference, ultimately diminishing model accuracy and robustness in practical engineering applications.To address these critical limitations, this study introduces an innovative ice classification framework that synergistically integrates multimodal feature extraction with advanced attention mechanisms. The proposed architecture employs a dual-branch design paradigm: one branch utilizes deep convolutional neural networks for high-dimensional image feature extraction, while the other incorporates multilayer perceptrons for meteorological feature encoding, achieving comprehensive fusion of heterogeneous data sources at the feature level. Furthermore, the implementation of a channel-spatial dual attention mechanism substantially enhances the model's discriminative capacity for critical icing characteristics while effectively suppressing background noise interference.Extensive experimental validation demonstrates the superior performance of our approach, achieving 97.43% classification accuracy on a representative icing scenario dataset - representing a 4.89 percentage point improvement over the EfficientNetV2-S baseline model. Detailed performance metrics reveal significant enhancements across all evaluation criteria, with precision, recall, and F1-score improving by 7.96%, 6.03%, and 7.09% respectively. This research contributes a novel and robust solution for fine-grained ice classification, offering substantial engineering value for power system security. The proposed methodology establishes a fundamental technological framework for next-generation intelligent ice monitoring and early warning systems, enabling three critical capabilities: (1) real-time precise ice type identification, (2) accurate predictive modeling of ice accumulation patterns, and (3) automated risk assessment and classification. These advancements collectively contribute to enhanced grid resilience and disaster mitigation capabilities, representing a significant step forward in power infrastructure protection.