A Novel Approach to Analyzing Defects: Enhancing Knowledge Graph Embedding Models for Main Electrical Equipment

The safety of electric power grids can be threatened by defects in main electrical equipment, creating significant risks and pressures for dispatching operations. To analyze defects inmain electrical equipment, we adopt a knowledge graph link prediction approach. We found that using pre-training models, such as BERT, to extract node features and embed initial embeddings significantly improves the effectiveness of knowledge graph embedding models (KGEMs). However, this approach may not always work and could lead to performance degradation. To address this, we propose a transfer learning method that utilizes a small amount of domain-specific electric power corpus to fine-tune the pre-training model. The PCA algorithm is used to reduce the dimensionality of extracted features, thereby lowering the computational cost of KGEMs. Experimental results showthat our model effectively improves link prediction performance in analyzing defects in main electrical equipment.