This paper deals with the black-box modeling of 3-DOF nonlinear maneuvering motion of surface ships by using system identification method based on BP neural network. A Mariner Class vessel is taken as the study object. The time series used in training and testing the network is the simulated data of a series of maneuvers, which is obtained by numerically solving the Abkowitz model using fourth-order Runge-Kutta method. A three-layer neural network is built to solve this multivariable regression fitting problem, and only one network model is trained to predict various ship maneuvers. Taking the mean squared error (MSE) as the loss function, the network’s weights are optimized by Levenberg-Marquardt (LM) algorithm effectively. The trained network is evaluated by several simulation tests, and it is shown that the network achieves good prediction ability and can predict the maneuvering motion as long as the control inputs and initial states of the ship motion are known.