Neural network-based controller design of a batch reactive distillation column under uncertainty

This paper presents the use of neural network-based model predictive control (NNMPC) incorporated with a neural network (NN) estimator for handling the predefined optimal policy tracking of a batch reactive distillation. The predefined optimal policy has been determined by dynamic optimization strategy. Then, the NNMPC incorporated with the NN estimator has been implemented to provide tracking of the obtained optimal policy. The NN model in the MPC algorithm gives as a one-step-ahead prediction of states, and it is therefore used in every iteration over a prediction horizon. Thus, the measured distillate composition at current time, needed as one of NN model inputs, is needed. However, the composition measurement is rarely available online in practice. Hence, an NN estimator is developed to estimate the current composition from the available measured composition with delay of 10 min. Both NNs are trained based on LevenbergMarquardt algorithm. It has been found that the NNMPC provides satisfactory control performance for set point tracking problems. The robustness of the NNMPC is investigated with respect to parametric plant uncertainties and temperature measurement noise. Comparisons are made with a proportional integral derivative (PID) controller incorporated with the NN estimator. The results show that the NNMPC provides better control performance than the PID controller in all cases. (C) 2011 Curtin University of Technology and John Wiley & Sons, Ltd.