A methodology is presented to obtain approximate models from input-output data, particularly oriented to implement a model-predictive control scheme. Causal, time-invariant nonlinear discrete systems with a certain type of continuity condition called fading memory are dealt with. To synthesize the nonlinear model a finite-dimensional linear dynamic part (discrete Laguerre polynomials) is used, followed by a nonlinear nonmemory map (single hidden-layer perceptron). Results of the application to approximate and control a binary distillation column are presented.