Neural network prediction is a very challenging subject in the presence of disturbances. The difficulty comes from the lack of knowledge about perturbation. Most papers related to prediction often omit disturbances but, in a natural environment, a system is often subject to disturbances which could be external perturbations or also small internal parameters variations caused, for instance, by the ageing of the system. The aim of this paper is to realize a neural network predictor of a nonlinear system; for the predictor to be effective in the presence of varying perturbations, we provide a neural network observer in order to reconstruct the disturbance and compensate it, without any a priori knowledge. Once the disturbance is compensated, it is easier to realize such a global neural network predictor. To reach this goal we model the system with a State-Space Neural Network and use this model, completed with a disturbance model, in an Extended Kalman Filter.