This paper presents a Model Identification scheme for a class of nonlinear dynamic systems. A novel Grey-Box Neural Network Model (GBNNM), including Multi-Layer Perception (MLP) Neural Network (NN) and integrators, is proposed to approximate both the nonlinearity and dynamics of the object system. A self-defined exciting strategy is introduced into NN training to improve its generalization ability. Unlike previous NN based model identification methods, GBNNM directly inherits system dynamics and models nonlinearities separately. This accords well with the expected model and is easy to implement. Then, the proposed scheme is applied in a high-fidelity Reaction Wheel (RW) in Satellite Attitude Control System (SACS). The proposed scheme using GBNNM is compared with those using static NN or dynamic NN in the same scenario. Results demonstrate the effectiveness and superiority of the scheme.