Adaptive neural model predictive control for the grape juice concentration process

The four-stage evaporator is the core of the process in the manufacture of concentrated grape juice. The dynamic features of this process are very complex due to inputs and outputs constraints, time delays, loop interactions and the persistent unmeasured disturbances that affect it. Therefore, this kind of process requires a robust control in order to assure a stable operation taking into account the changes in the organoleptic properties of the raw material and, to guarantee the quality of the concentrated product. This work proposes an adaptive neural model to control of a four-stage evaporator in a grape juice concentration plant. In order to obtain a more accurate process description the neural model is trained with data from simulation of a phenomenological model and afterwards, is validated with actual plant data. This strategy allows to carry out the training without to introduce disturbance in the real plant. Neural networks of different size are trained and the performance of one of the neural models is compared with the first principles model. In a last step, the performance of a model predictive control based on the neural model is evaluated for disturbance rejection and compared with a MPC controller based on the phenomenological model and with a PI controller. The achieved results allow us to conclude that the developed neural model predictive control is adequate to control effectively the four-stage evaporator.