An orthogonal collocation method is used for the reduction of a large nonlinear model of a multi-stage flash (MSF) desalination plant. The real plant located in Saudi Arabia consists of 19 stages in the heat recovery section and three in the heat rejection section. It is shown that three collocation points, i.e., two for the heat recovery section and one for the heat rejection section, are sufficient for a good representation of the full-order model for both steady-state and dynamic behavior. The reduced model is then used for a robust controller design of the plant using a constrained nonlinear model predictive control (NLMPC) strategy. Numerical simulations show that the controlled system obtained when the reduced model is used for output prediction yields the same performances as with the full-order model but with a considerable reduction in the computational time and a much easier tuning of the controller.