Optimization of Heterogeneous Fault Information Graph for Power Grids Based on Graph Neural Networks

Heterogeneous fault handling in power grids faces significant challenges such as overlapping triple recognition and inefficient multimodal feature fusion. The complexity is further exacerbated by the representational disparity and association difficulty among diverse data types in heterogeneous fault information. To address these issues, this paper proposes a joint optimization framework named Multimodal Hierarchical Contrastive Reasoning with Adaptive Weighted Graph Convolutional Network (MHCR-AWGCN). First, a power grid domain ontology is constructed, defining five entity types and eight relation types, along with standardized entity-relation mapping rules. In the knowledge extraction phase, a multimodal hierarchical contrastive reasoning mechanism is designed, which integrates textual, parameter, and topological modalities. By conducting contrastive learning at the entity, relation, and triple levels, the model enhances the collaborative representation capabilities across modalities. In the knowledge optimization phase, an adaptive weighted heterogeneous graph convolutional network is proposed, which aggregates features from multimodal subgraphs of fault propagation via relation weights constrained by electrical parameters. A joint loss function is further introduced to simultaneously optimize node embeddings and topological structures. We conducted ablation experiments using this model to demonstrate the effectiveness and interpretability of the MHCR-AWGCN model in power grid fault decision support.