A large terminal set of model predictive control results in a large region of attraction of the closed-loop systems, which can help to reduce the computational burden of the involved optimization problem. In this paper, a novel scheme is proposed to obtain a terminal set and a terminal penalty of nonlinear model predictive control. Firstly, the nonlinear system is approximated through the Koopman operator theory, whereby a linear system with unknown but bounded disturbances is generated. Then, a polytopic terminal set is obtained accordingly, where the nonlinear system is described by a linear model with disturbances. The effectiveness of the proposed scheme is demonstrated using a benchmark problem.