Abstract: As a key step in building a power knowledge graph, knowledge extraction can accurately extract entities and relationships from massive unstructured power texts. However, the traditional pipeline method has the problems of backward transmission of error information, separation of entity recognition, and relationship extraction tasks, and is easy to generate redundant information, which results in low extraction accuracy, incomplete extraction of information, and ultimately impairs the accurate construction of the knowledge graph. To solve the above problems, this paper proposes a joint extraction method of overlapping entity relationships for the construction of the power knowledge graph. Through the improved sequence labeling scheme, the joint extraction is carried out, the exclusive pre-training model (the PowerRobertsa model) in the power field is constructed, and the confrontation training is increased, which improves the accuracy of the model extraction of power knowledge and the ability to predict unfamiliar information. Finally, by taking the actual substation patrol data as an example, the experimental analysis and the visual construction of the distribution Knowledge graph are carried out. The results show that the joint extraction method proposed in this paper can be adopted to improve the accuracy of knowledge extraction, which reaches 91.67%, and can effectively support the advanced application of distribution network intelligent information retrieval and decision-making assistance.