Neural network model-based on-line re-optimisation control of fed-batch processes using a modified iterative dynamic programming algorithm

A strategy to overcome the problems of unknown disturbances and model-plant mismatches in fed-batch process optimal control through on-line re-optimisation is presented in this paper. To address the difficulty in developing detailed first-principle models and the time-consuming nature of optimisation based on nonlinear differential equations, neural network-based discrete-time models are used to model fed-batch processes from process operation data. Based on the neural network model, an “optimal” feeding policy is calculated off-line and its first stage is implemented to the batch process. At the end of the first stage, the process output variables are measured and any differences between the measured values and neural network model predicted values reflect the existence of unknown disturbances and model-plant mismatches. Due to the existence of unknown disturbances and model-plant mismatches, the off-line calculated “optimal” feeding policy for the remaining batch period may no longer be optimal and should be re-optimised. Based on the mid-batch process measurements, re-optimisation is carried out for the remaining batch period. A modified iterative dynamic programming algorithm based on discrete-time nonlinear models is developed to solve the on-line re-optimisation problem. The proposed scheme is illustrated on simulations of an ethanol fermentation process.