It is of great significance to the development of smart grid on intelligent classification of complex power quality disturbances (PQD). Disturbance feature extraction and location, pattern recognition and classification are the difficulties of power quality disturbance classification. Based on deep learning, Transformer, which paid attention to global information, was combined with convolutional neural network (CNN). It was good at extracting local features. A power quality disturbance classification method based on CNN and Transformer, namely CTranCBA, was proposed. This dual-deep learning model classification mainly extracted the characteristics of power quality disturbance signals by one-dimensional CNN, and used Transformer self-attention mechanism to guide the model to pay attention to the dependence between different positions in the sequence. As a result, it realized the complementarity of local characteristics and global characteristics of disturbance signals, and overcame the problems of unclear recognition and inaccurate classification caused by the limitation of receptive fields. The CTranCBA was compared with Deep-CNN, CNN-LSTM, CNN-CBAM in 23 different power quality disturbance signals. The results show that the CTranCBA is superior in classification accuracy and noise immunity, which can provide a new method for intelligent classification of power quality disturbance.