| Scope and method of study. In this work, a Nonlinear Model Predictive Control (NMPC) strategy named as Grouped Neural Network Model Predictive Control (GNNMPC), based mainly on heuristics and experience is proposed to reduce the computation complexity for online application of NMPC. The unique way of modeling in GNNMPC, is to provide prediction of the future process outputs at selected sample intervals (in contrast to the conventional way of predicting the future outputs at each sample intervals) within the prediction horizon. One neural network (NN) model is used to provide prediction of the process outputs at one selected sample interval. The Grouped Neural Network (GNN) model, therefore, comprises a group of NNs, independently trained and implemented, to represent the dynamics of the process. This modeling approach does not make assumptions on the characteristics of the process and therefore is a general modeling approach. The separate and independent training of each NN decreases the complexity and effort in the training stage. Additionally, in the optimization problem, the number of decision variables, i.e., the future control actions, is significantly decreased by only making future control action change at subjectively selected future control time intervals.; Findings and conclusions. The proposed GNNMPC was demonstrated by both simulations and experiments to be capable of controlling a binary methanol-water distillation process, which has such control-challenging characteristics as modeling error, severe interactions, static and dynamic nonlinearity, measured and unmeasured disturbances, constraints on the process inputs as well as the operating conditions. GNNMPC has shown to be very effective in setpoint tracking, disturbance rejection, and constraint handling. |
