Abstract: The dissolved oxygen in aquaculture ponds is crucial to the survival of eriocheir sinensis. Grasping the trend of the dissolved oxygen concentration timely and accurately is the key for the high-density healthy eriocheir sinensis culture. In order to solve the low prediction accuracy and bad robustness of the traditional methods in dissolved oxygen content prediction, the combined ant colony algorithm (ACA) was combined with least squares support vector regression (LSSVR) algorithm to construct a non-linear prediction model for predictin the dissolved oxygen content changing in intensive aquaculture eriocheir sinensis cultures. The hybrid ACA-LSSVR algorithm inherits from advantages of this regression model support vector machine and applies linear least squares criteria to the loss function instead of traditional quadratic programming method. Thus it effectively increases the training speed. But it restricts the prediction accuracy and performance whether the kernel function and parameter calibration are appropriate. Therefore, we choose the Gauss RBF kernel function as the kernel function of the least squares support vector regression model, which needs few parameters to set and can well process high dimensional space nonlinear relationship. Moreover, the global heuristic search ant colony optimization method was used to seek the optimal hyper parameters needed in the least squares support vector regression model, which overcoming the blindness and the impact of human factors of the trial-and-error method in parameter selection of support vector regression model and accelerating model convergence rate. The combinations of the best parameters were obtained automatically after the optimization, from which the nonlinear prediction model between the dissolved oxygen and the impact factors was constructed. Based on the prediction model, the dissolved oxygen changing was predicted for a high-density aquaculture pond during July 20, 2010 to July 28, 2010 in Yixing city, Jiangsu province. Experimental results show that the proposed prediction model of ACO-LSSVR has good prediction effect than the support vector regression machine and BP neural network method. Under the same experimental conditions, the relative RMSE, MSRE, MAE, R2 and the running time differences between the ACA-LSSVR and standard SVR models are 56.1%, 36%%, 30.8%, 0.9% and 1.5651 s respectively. It is clear that ACA optimization is better than the trial-and-error method to get the LSSVR parameters. The relative RMSE, MSRE, MAE, R2 and the running time differences between the ACA-LSSVR and BPNN models are 67.9%, 60%, 50.8%, 4.7% and 2.3464 s respectively. It is obvious that ACA-LSSVR is more accurate than BPNN. The prediction model has low computational complexity, fast convergence rate, high forecast accuracy and generalization ability. It can meet the actual demand for the dissolved oxygen controlling in the high density eriocheir sinensis culture and provide a reference for other areas of water quality predictions.