GCN-ICD: A Graph Convolutional Network for Icing Cover Detection

We propose a novel icing prediction model consisting of a heterogeneous information embedding layer, a direction-aware information aggregation layer, and a detection layer. In particular, we represent transmission lines as nodes in a graph and establish edges based on their connections. Taking visual images and meteorological data as the content of nodes, we design a large language model (LLM) based embedding layer to fuse the heterogeneous information. Then, the direction-aware information aggregation layer efficiently absorbs and aggregates information from neighbor nodes while considering the edge weights and directions. Finally, with the obtained representations of nodes, the detection layer is able to measure the icing probability of the transmission line. By conducting extensive experiments on a real-world dataset, we demonstrate the superiority of our proposed model compared to baseline methods.