Traditional power quality disturbances (PQD) classification methods rely on a large amount of labeled data and perform poorly under multiple disturbance coupling and strong noise interference. To solve this problem

this paper proposes a semi-supervised PQD recognition framework that integrates a time-frequency adaptive wavelet network (TFAWNet). First

a teacher network TFAWNet based on wavelet convolution and multi-level attention mechanism is constructed and trained with a small amount of labeled data to obtain the optimal model. Subsequently

the trained and optimized teacher model is used to perform deep reasoning on unlabeled data to generate high-confidence pseudo labels

and the labeled data and pseudo-label data are combined to jointly train a lightweight student network EfficientNet. The experimental results show that with only 50 labeled samples for each type of disturbance

the student network has a test accuracy of 93.27% on the simulation data. After the model is deployed on the edge computing platform

it is verified by measured data that the accuracy is as high as 99.83%

and the average inference time is only 10 ms. These performances further verify the superior performance of the framework in simulation and practice

highlighting the robustness

practicality and efficiency of the model.