A model predictive control at the supervisory level is proposed for refrigeration systems using price and temperature predictions. The control objective is to minimize the overall energy cost within the prediction horizon. The method is mainly developed for demand-side management in the future smart grid, but a simpler version can be applied in the current electricity market. Due to the system nonlinearity, the minimization is in general a complicated nonconvex optimization problem. A new supervisory control structure as well as an algorithmic pressure control scheme is put forward to rearrange the problem to facilitate convex programming. A nonlinear continuous time model validated by real data is employed to simulate system operation. The results show a considerable economic saving as well as a trade-off between the saving level and design complexity.