Abstract: Constructed wetland systems are effective in treating sewage. They have simple process and low investment/operation costs. However, there are many factors influencing the water quality, and these factors are often nonlinear. Therefore, it is difficult to model these factors so as to predict water quality. Some prediction methods are too complicated, while others have a relatively low prediction accuracy. The artificial neural network is an efficient, new method in predicting a variety of non-linear models. On the basis of laboratory experiments, three kinds of trained wavelet neural networks were used to predict water quality along the Phragmites australis subsurface flow constructed wetlands, including water temperature, ρ(DO), pH, Eh andρ(COD_Cr). The prediction results showed that the averagerelative error of the water temperature≤4.21%, pH≤1.36%, ρ(DO)≤9.77%, Eh≤6.50%, ρ(COD_Cr)≤17.76%. The results indicated that the wavelet neural networks model can effectively predict various items in water quality of constructed wetlands.