Abstract: In view of the problem of low classification accuracy which is caused by the large amount of data and insufficient feature extraction in the current power quality disturbance classification, this paper com-bines compressed sensing and deep learning to propose a power quality disturbance classification method based on improved stacked denoising autoencoders (ISDAE). First, the sparse vector obtained after the original data compressed and sensed is used as the data set. Then the stacked denoising au-toencoder model is constructed, the inverted dropout technology is introduced to improve the gener-alization ability of the network and avoid the occurrence of overfitting. In the fine-tuning stage adaptive moment estimation (Adam) optimization method is introduced to reduce the probability of falling into a local optimum. Finally, simulation analysis is performed on 10 common power quality disturbance signals. It can be found that this method effectively reduces the amount of disturbance data that needs to be processed, and solves the problem of low classification efficiency due to the insufficient feature selection of traditional classification algorithms. To a certain extent, the robustness of the model is improved.