Abstract: The extensive application of deep learning algorithms in the field of partial discharge pattern recognition largely makes up for the shortcomings of traditional works such as difficulty in practice and low efficiency. However, because the unpredictable situations happen on-site, the partial discharge has complex mechanisms, and partial discharge data are rarely available for training, most deep learning algorithms are still difficult to achieve ideal recognition accuracy in actual situations. By analyzing the problems in the features of actual PRPS pattern and comparing the differences with the features of PRPS pattern measured on partial discharge defect models, this paper finds that the actual pattern has more atypical features. Therefore, we put forward a partial discharge recognition method based on the Swin Transformer network. We tested the recognition accuracy of the network on the data from two sources mentioned above, and compared the results with traditional neural networks. The results show that the Swin Transformer can achieve a recognition accuracy of 93.3% on on-site data sets. Compared with traditional methods, the Swin Transformer has higher recognition accuracy. It is also suitable for situations where there is a lack of partial discharge samples and it is more suitable for practical use.