Abstract: Directly mapping primitive samples to yaw velocity is infeasible when constructing black-box model of ship maneuvering by means of SVR（Support Vector Regression）. Two ways of building the sample structure are devised to solve this problem. The prediction accuracy difference between the two ways is discussed. The parameters of the kernel function are optimized through grid search cross validation. The impact of variation of parameters to prediction accuracy is demonstrated to justify the effort of optimization. Taking KVLCC2 as an example, the SVR black-box model is trained to predict 10°/10°, 20°/20° zigzag tests and 35° turning circle maneuver. The effectiveness and generalization performance of the proposed model are verified through predicting 15°/15° zigzag test and 15°, 25°turning circle maneuver.