According to the water quality monitoring data from one of the P. vannamei breeding base in Hangzhou, a mathematical model of multi-layer feed forward neural network was established to predict and evaluate the quality of aquaculture water. The topology of the model was 40-14-4, that was, the temperature, pH value, dissolved oxygen, and oxidation-reduction potential were the input variables in m=10 consecutive time units, the number of hidden layer nodes was 14, and the output layer was 4. In order to overcome the disadvantages of traditional back propagation （BP） neural network including slow convergence speed, easy to fall into the shock, poor generalization ability and so on, the adaptive variable step size BP-neural network learning algorithm based on the fuzzy method had been developed, which can reduce the learning time of BP-neural network, improve the convergence efficiency and network stability.

　　The computed results for water quality showed good agreement with the measured values. The correlation coefficients for the temperature, pH value and the oxidation-reduction potential during the training and testing processes were better than 0.975. The comparison between the original measured and predicted values of the BP-neural network showed that the relative errors, with a few exceptions, were lower than 2.5%.

　　In conclusion, the BP-neural network model can well reveal the complicated non-linear relationship between the input and output water quality variables in intensive P. vannamei tanks, and the improved BP-neural network (FABPM) method based on fuzzy method has the characteristics of fast convergence, high accuracy and good stability. It provides a new method for the prediction and evaluation of water quality in aquaculture.