Abstract: Existing approaches to evaluate ship traffic flow are difficult to use when the relationships among the indicators are not clear.These methods are subjective and the results are usually deviated from the real cases.In order to reduce effects of experts' subjectivity when evaluating the severity of marine traffic flow conflicts, a model based on BP neural network is proposed in this paper, in which the precision of Trainlm function is calculated and the number of iterations is determined using network training.In order to avoid the influences of the data on training efficiency and accuracy of BP neural network, a new model based on clustering analysis and BP neural network is proposed.The training data is classified according to Euclidean metric in order to carry out neural network training.This model is then used to evaluate the severity of the conflicts of water traffic flow.A case study of 9 channels is then performed to evaluate the proposed model.Comparing the processed data by clustering analysis with the original data, the evaluation error is reduced to 23.74% from 42.05%, which verifies the feasibility of BP neural network in this case.The modified model, which combining BP neural network with cluster analysis, has higher precision and more objectivity.