The dissolved oxygen (DO) in the pond has a very close relationship with the growth of aquatic products such as fish and shrimp. In order to improve the prediction accuracy and effectiveness of DO, a combined prediction model based on empirical mode decomposition (EMD) and least squares support vector machine (LSSVM) of adaptive disturbance particle swarm optimization is proposed. First DO time series are decomposed into several components by EMD, then each component is subject to phase space reconstruction and modeling prediction by LSSVM in the phase space, and finally adaptive disturbance particle swarm optimization is applied for optimization of hyper-parameters of LSSVM and single point iterative method for multi-step prediction. The results show that the model has good prediction effect compared with single LSSVM prediction model. When DO value of next 4h is predicted RMSE, MAPE and MAE are decreased by 13.4%, 11.3% and 1.8% respectively; when DO value of next 24h is predicted, the three indexes are decreased by 12.9%, 12.1% and 2.7% respectively. The study shows that the combined model can extract DO series features effectively and has relatively high prediction accuracy and generalization performance.