Nonlinear Process Modelling and Optimization Control Using Computational Intelligence Techniques

Abstract

With the ever-increasing global competition and customer requirement, chemical industrial processes face increasing pressure of maintaining high product quality and reducing production costs. To reduce the production cost and limit the amount of off-specification products production, close control of polymer product quality is required. The melt index (MI) of polymer is an important product quality indictor that needs to be closely monitored and controlled. However, it is difficult to measure MI in real-time. To overcome these issues, a reliable data-driven soft sensor for MI using deep belief network (DBN) is developed in this study. The training of DBN involves two phases: initial unsupervised training using input data only and supervised fine-tuning. It can capture profuse information from latent operational inputs. It is shown that DBN gives better performance in estimating MI than conventional neural networks. To further enhance the accuracy and robustness of DBN models, a bootstrap aggregated deep belief network (BAGDBN) is developed in this study. Several DBN models are developed from bootstrap resampling replications of the original modelling data and are combined to form a BAGDBN model. The effectiveness of the proposed model is demonstrated on two application examples, inferential estimation of polymer MI in an industrial polypropylene polymerization process and dynamic modelling of water level in a conical water tank. It is shown that BAGDBN has a much better generalisation capability than a single DBN model. A reliable optimal control strategy for a batch process is developed using BAGDBN model. To enhance the reliability of the optimal control policies, the model prediction confidence bound is incorporated into the optimisation objective function. Wide model prediction confidence bounds are penalised in order to enhance the reliability of the obtained optimal control policy. Application to a batch reactor demonstrates the effectiveness of the proposed reliable optimisation control strategy.