A novel method combined with adaptive wavelet threshold noise reduction and deep learning is proposed to improve the accuracy of identifying power quality disturbances in strong-noise environments. First, the noise-containing disturbance signals are noise-reduced by a threshold function algorithm based on an improved peak and score level adaptive thresholding and energy optimization. Then, the residual network is used to extract deep features from the noise-reduced disturbance signals, based on which the bidirectional long short term memory network under the multi-attention mechanism is incorporated to establish temporal feature dependency. This constitutes a framework applicable to the recognition of disturbance signals in a noisy environment. Finally, numerical simulations are carried out on 20 types of disturbance signals in different noise environments. It can be seen from the results that the proposed method has good noise robustness and high recognition accuracy in different noise environments