A lightweight power quality disturbances (PQDs) recognition model that integrates convolutional neural network (CNN) and Transformer (CaT) is proposed to address the high number of parameters and computational complexity in existing deep learning-based models. Depthwise separable convolutions are first employed to extract local features from the disturbance signals. An efficient softthreshold block is then introduced to reduce noise and redundant features without significantly increasing the model's parameters or complexity. The Transformer model is used to capture global features of the disturbance signals. Finally

pooling layers

fully connected layers

and Softmax are applied to complete the recognition PQDs. Simulation experiments demonstrate that the CaT model effectively recognizes PQDs with fewer parameters and floating point operations

achieving high accuracy and strong noise robustness. Its lightweight

end-to-end design also results in shorter inference times compared to other deep learning models.