BP neural network is a feed2forward neural network that is learned according to error backpropagation algorithm. BP neural network with excellent nonlinear approximation ability is widely applied to various fields. The excellent nonlinear approximation ability of BP neural network is ensured by determining the topology and structural parameters properly , learning efficient training data set with good typical characters , searching the global minimum solutions and escaping overlearning during learning. According to the recently research results of BP neural networks modelling , some aspects on comprehensive assessment of water environmental quality using BP neural networks were presented and studied in this paper. The main problems are too few of training set data , no verification ( validation ) set data , too large in network topology , which thus resulted in overfitting and overlearning in training and poor generalization of the neural networks model set up. The necessary modelling conditions for BP neural networks were concluded. The many BP models , presented before , for water environmental quality were set up under the conditions inconsistent with the necessary modelling conditions . The case study shown that the model set up under conditions disagreement with the necessary modelling conditions possessed poor generalization, prediction capability, and possibly induced multimodal in connection weights.