Neural networks have attracted much interest in system identification and control communities because they can learn any nonlinear mapping. However, from a user's point of view, when neural networks are used as models for controller design, they do not have structures of easy use. This paper introduces a new neural network based prediction model for control of nonlinear systems. Distinctive features of the new model to conventional neural-network based ones are that it has not only meaningful interpretation on part of its parameters but also is linear for the input variables. The former feature makes parameter estimation easier and the latter allows us to derive a nonlinear controller directly from the identified prediction model. The modeling and the parameter estimation are described in detail. The usefulness of the new model is demonstrated by using numerical simulations.