Applications of deep learning methods in the classification task of power quality disturbances (PQDs) are becoming increasingly popular. Aiming at the limited availability of measured label data compared to the ability to generate simulated data in large batches, a disturbance classification method is proposed based on recursive graph theory and pre-trained transfer learning. Firstly, the PQDs signals are transformed into two-dimensional recursive images using the recursive graph algorithm. Then, a VGG-16 deep learning network is pre-trained using a large amount of simulated data, and the model's weight parameters are saved. Finally, through transfer learning method, the fully connected layer of the model is fine-tuned using a limited amount of measured data, enabling deep feature extraction and classification of PQDs signals under the constraint of limited training samples. Simulation and measured data validation demonstrate that the proposed method achieves high classification accuracy even with limited labeled data.