Dissolved oxygen (DO) plays an important role in aquaculture. It has the characteristics of changing with time, instability and nonlinearity, and it has too many influence factors with complex coupling relationship, and it is difficult to be predicted accurately. The traditional long shorttime memory neural network (LSTM) is easy to introduce redundant data. And when it deals with long sequences, the gradient disappears so that it cannot capture very long term dependencies. To solve the problems above, the WT-CNN-LSTM prediction model was proposed. In view of the timing and nonlinearity of dissolved oxygen in aquaculture, the LSTM, which was widely used in time series prediction and had excellent performance, was selected to predict the dissolved oxygen value in two hours later. Aiming at the noise generated by environmental factors, human factors and system factors in the process of data collection, the hybrid wavelet transform (WT) was proposed to reduce noise in the data set so as to provide reliable support for the establishment of accurate prediction model. Moreover, due to the complex aquaculture environment, the dissolved oxygen content was affected by a variety of water quality factors and environmental factors. Therefore, the convolutional neural network (CNN) was used to mine and store the potential information between variables and DO in aquaculture. The result showed that WT-CNN-LSTM had good predictive performance. Its mean absolute error, root mean squared error and determination coefficient were 0.138, 0.229 and 0.954, respectively, which were optimized by 28.87%, 21.03% and 4.61% compared with those of the LSTM model. 