A novel approach to the design of a model predictive control (MPC) algorithm for a complex plant (with energy integration and mass recycles) is given. Sensitivity analysis and steady-state optimization are used to determine the manipulated variables that have the strongest influence on the objective function of the operation. This allows a reduction of the number of variables that are optimized on-line, as well as the use of detailed, first-principle-based models in the MPC algorithm, thus resulting in more reliable predictions. Moreover, the same algorithm can be used to control plants of different size, without the need of a new calibration of the parameters of the model. The application of this procedure to a pressure-swing distillation unit is given as an example. (c) 2006 American Institute of Chemical Engineers.