This paper addresses the problem of implementing predictive controllers for supervisory level control systems. In this configuration the manipulated variables calculated by the Predictive Controller are used as command signals for the Distributed Control Systems, which provide references to the operator-tuned local PID controllers that act on the physical system. This structure introduces the problem of loosing of performance if the inner-loop controllers are re-tuned. The paper discusses the solution to this problem based on the use of a two-degrees-of-freedom structure in the inner loop, that separates open and closed-loop properties. Both design guidelines and robustness issues are discussed.