A weather recognition system for transmission lines leveraging multi-source data

Accurate weather recognition along transmission lines is essential for maintaining power system reliability. Since transmission lines are exposed to the external environment, they are highly susceptible to weather conditions, making accurate weather identification particularly important. However, weather images often lack distinctive local textures, making it difficult for conventional convolutional networks to extract discriminative features. To address this challenge, this paper introduces a ConvNeXt-based framework enhanced with two modules: Numeric-Conditioned Feature Modulation (NCFM) and Shallow-Guided Channel and Spatial Attention (SGCSA). The NCFM module integrates auxiliary numerical data—such as temperature, humidity, and wind speed—to adaptively modulate visual representations, enabling the model to capture weather-dependent variations more effectively. The SGCSA module leverages shallow features to guide deeper feature learning through combined channel and spatial attention, thereby improving focus on informative regions. A dedicated transmission line weather dataset was established to train and evaluate the proposed model. Experimental results demonstrate that the framework achieves 96.36% accuracy, surpassing the baseline ConvNeXt by 3.11%. Generalization capability is further validated on an independent test set collected from unseen transmission-line scenarios. Furthermore, A visualization tool integrating Grad-CAM was also developed to support qualitative inspection analysis. Overall, the proposed approach provides an effective solution for transmission line weather recognition by leveraging multi-source data fusion.