Neural Network Based Model Predictive Control of Batch Extractive Distillation Process for Improving Purity of Acetone

Abstract

In a pharmaceutical industry, batch extractive distillation (BED), a combination process between extraction and distillation processes, has been widely implemented to separate waste solvent mixture of acetone-methanol because of minimum-boiling azeotrope properties. Normally, water is used as solvent and semi-continues mode is proposed to improve purity of acetone. The solvent is charged into the BED column until the purity of a desired product is achieved. After the total reflux start-up period is ended, a dynamic optimization strategy is applied to determine an acetone distillate composition profile maximizing the weight of the distillate product (acetone). The acetone distillate composition profile is used as the set point of neural network model-based controllers: the neural network direct inverse model control (NNDIC) and neural network based model predictive control (NNMPC) in order to provide the acetone composition with the purity of 94.0% by mole within 9.5 hours. It has been found that although both NNDIC and proportional integral derivative (PID) control can maintain the distillate purity on its specification for the set point tracking and in presence of plant uncertainties, the NNMPC provides much more satisfactory control performance and gives the smoothest controller action without any fluctuation when compared to the NNDIC and PID.