In order to solve the problem of low prediction accuracy and bad robustness of the traditional forecasting methods in water quality, this paper put forward the prediction model for pH value of aquaculture water quality based on the principal component analysis (PCA) and least squares support vector machine (LSSVM), which the hyper-parameters is optimized by modified cultural artificial fish-swarm algorithm(MCAFA). The dimension of aquiculture ecologic environmental data was reduced by principal component analysis; double evolutionary mechanism of cultural algorithm for reference was applied and LSSVM was taken as an artificial fish; belief space was used to guide the shoal evolution step size, global search direction and Cauchy mutation to improve the diversity of the artificial fish swarm; so the optimal hyper-parameters nonlinear pH value prediction model was automatically obtained. Based on the prediction model, the water quality on-line monitoring was predicted for a high-density aquaculture pond from September 1, 2011 to September 4, 2011 in Yixing city, Jiangsu province. Experimental results show that the PCA-MCAFA-LSSVM prediction model has good prediction effect than the ant colony algorithm LSSVM and genetic algorithm LSSVM. The absolute error of the 93.05% test samples is less than 8%, and the max absolute error is only 1161%; the root mean square error, average absolute relative error and the running time are 0.0474, 0.0041 and 4.367s respectively, which are better than those from the other models. It is obvious that PCA-MCAFA-LSSVM prediction model has low computational complexity and high forecast accuracy. It can provide the decision basis for the water quality controlling in the high density eriocheir sinensis culture.