Abstract:

With the deepening development of new-type power systems, there is an urgent demand for processing unstructured texts in areas such as equipment condition monitoring, intelligent fault diagnosis and dispatch instruction parsing. To address the heterogeneity of entity types such as equipment, operations and faults, along with the complex semantic relations in power texts, a semantic enhanced graph convolutional network method for entity extraction in power systems is proposed. First, data augmentation strategies such as synonym replacement, random insertion, swapping, and deletion are applied to reduce the expression differences in manually recorded texts. Then, a robustly optimized bidirectional encoder representation from transformers pretraining approach (RoBERTa) with a dynamic masking mechanism and whole-word masking is used for fine-grained semantic encoding of domain-specific terms. Next, a semantic relation graph is constructed, including nodes such as equipment, operations, and system states. Finally, a relation extraction framework based on large language models is introduced. This framework incorporates a part-of-speech filtering matrix to enhance graph convolution feature aggregation. Experiments show that the proposed model achieves score of 88.79% on dispatch logs, outperforming existing models by 2.3%~3.1%. The model can accurately identify power operation and status entities such as "closing the tie switch" and "grounding disappears", providing strong support for updating dispatch knowledge bases and enabling intelligent decision-making under fault conditions.

Key words: entity extraction of power system, semantic enhancement, graph convolutional neural network, entity extraction