Dissolved oxygen (DO) is one of the key water quality parameters for water products, which reflects changes in water quality in aquaculture. To increase prediction accuracy of dissolved oxygen (DO) in aquaculture, a hybrid model based on ensemble empirical mode decomposition (EEMD), reconstruction using runs test method, the appropriate forecasting method and superposition by BPANN was proposed. Firstly, the dissolved oxygen sequence was decomposed into a series of components by EEMD. EEMD performs very well in noise reduction and detail features extraction. Original dissolved oxygen datasets were decomposed to 9 IMFs and a Res. Then these components were reconstructed into high frequency component， intermediate frequency and low frequency component by the runs test method, which centralized the characteristic information and reduced the difficulty of predicting. Thirdly, based on the changed characteristics of each sequence, the appropriate prediction method was selected. According to the characteristics of high frequency component, which fluctuated violently and varied complicatedly, the least squares support vector machine (LSSVM) optimized by particle swarm optimization (PSO) was used to predict this term because this method had high accuracy and adaptability to predict. Intermediate frequency component was forecasted by extreme learning machine (ELM) because this component presented a periodic characteristic. The fluctuation of low frequency term was gentle and periodic, so it could be modeled by nonlinear regression method. Finally, the predicted values of DO datasets were calculated by using BPANN to reconstruct the forecasting values of all components, which ensured good fitting effect and stability. The model adopted BPANN nonlinear superposition to replace simple adaptive superposition, which achieved better fitting effect. This model was tested in Daitou Huangjiadang Special Aquaculture Farm in Liyang City, Jiangsu Province. Based on the prediction model, the dissolved oxygen changing was predicted for aquaculture pond during June 23 to June 24, 2016. The experimental results demonstrated that the proposed combinatorial prediction model had better prediction effect than the ELM and LSSVM. When it predicted the next 12, 24, 36 h dissolved oxygen values by using the proposed combinatorial prediction model, the root mean square error (RMSE) of the proposed combination prediction model was 0.088 6, 0.093 1 and 0.095 7, respectively, mean absolute errors (MAE) were 0.066 6, 0.071 2 and 0.076 9, respectively; and mean absolute percentage errors (MAPE) were 0.008 9, 0.013 1 and 0.014 2. When dissolved oxygen values of next 48 h were predicted, the RMSE, MAPE, and MAE of the proposed combination prediction model were 0.099 2, 0.078 and 0.015 5, respectively. The RMSE, MAPE and MAE for dissolved oxygen value of the standard extreme learning machine (ELM) were 0.1581, 0.111 2 and 0.027 6, respectively. Three indexes of LSSVM were 0.134 4, 0.102 6 and 0.021 3, respectively, under the same experimental conditions. The proposed model performs well in terms of all the evaluation indexes, so the proposed model is highly competitive with the compared models. It is obvious that the proposed hybrid model has high forecast accuracy and is proven to be an effective way to predict DO. [ABSTRACT FROM AUTHOR]