With the development of smart grid technology, how to effectively deal with the large amount of text information generated in the process of distribution grid operation is particularly important. Named entity recognition is a key technology to construct knowledge graph of distribution grid operation, but there are problems such as difficulty in entity recognition and low accuracy. In response to the appealing situation, this paper proposes a named entity recognition model based on the fine-tuning of RoBERTa-wwm pre-trained language model, and through the comparative experimental analysis, the model in this paper reaches 94.68% of the F1 value on the distribution network operation dataset, which is more advantageous than the existing model. On this basis, the knowledge graph of distribution network operation is constructed, which can assist distribution network operators to make auxiliary decisions and improve the safety and efficiency of operators.