Aiming at the problem of low recognition accuracy of power quality disturbance (PQD) in complex noise environment, this paper combines the soft threshold function with one-dimensional convolutional neural network, and proposes a deep convolutional denoising network for power quality disturbance recognition. The soft-threshold function is inserted into the deep network as a nonlinear transformation layer, and a soft-threshold denoising module is constructed to effectively remove noise or other redundant features. Soft-threshold denoising networks are learnable parameters whose weights can be determined through model training. Compared with the soft threshold function, the proposed denoising module can adaptively calculate its thresholds from different sample inputs. The simulation results show that the proposed method can effectively identify 18 common single and multiple PQD in the case of different noise environment. Compared with other common disturbance identification algorithms, the proposed method has better anti-noise and high recognition accuracy of PQD in complex noise environment.