Modelling and Optimisation of Batch Processes Using Computational Intelligence Techniques and Statistical Learning Approach

Abstract

Batch processes are commonly used for the manufacturing of high-value-added products such as specialty chemicals and pharmaceuticals. The efficient operation of batch processes is of great importance to produce high-quality products with minimal consumption of energy and resources. Batch process optimization and control are essential for achieving this. One common approach to the modelling and optimization of batch process systems is to use the first principle concept in building the models of the process using mass balances, energy balances and reaction kinetics of the batch process operations. However, this type of mechanistic model is difficult to develop due to the complexity of the process operations, multiple variables, and batch-to-batch variations involved. The development of mechanistic models is time and effort-challenging, which may not be feasible for batch processes where frequent changes in product specifications occur and a type of product is usually manufactured for a limited time in response to the dynamic market demand. Computational intelligence techniques address these shortcomings by making use of datadriven concepts in the modelling and optimization of batch processes. With the development and progress in research, data-driven modelling is becoming the more widely used method in modelling and analyses of batch/fed-batch process operations. Extreme learning machine (ELM) is a type of data-driven modelling technique with a fast-training process and can be used for modelling a different kind of process operation like the conventional neural network (NN). ELM has been established to be successful in modelling nonlinear (complex) batch operations as it provides good generalization performance at fast learning speed and gives accurate long-range or multi-step ahead prediction performance. However, it has its shortcomings as well, hence the need to combine other statistical learning techniques to improve its general prediction capabilities. This work presents the modelling and optimisation of fed-batch processes using different data-driven modelling techniques such as the ELM, Bootstrap Aggregated ELM, and Iterative Learning Control. It also presents the strategy of merging extreme learning machine (ELM) and recursive least square (RLS) techniques in modelling and batch-to-batch optimization of fed-batch processes. To cope with the batch-to-batch variations due to unknown disturbances such as unknown process condition drift, the RLS algorithm is integrated with the ELM to update the output iii layer weights recursively from batch to batch. This is because the number of hidden neurons selection together with the output layer weights computation are major criteria towards accurate model predictions in ELM. The recursive least square (RLS) adapts to the current process operation by recursively solving the least-squares problem in the considered model. RLS estimation algorithm nullifies the model plant mismatches caused by the occurrence of unknown disturbances. The offline trained output layer weights of the ELM are used as the initial parameter estimation in RLS. After updating the ELM model, optimisation is carried out to update the feeding policy for the next batch. The proposed technique is thus applied to two fed-batch case studies including a simulated fed-batch reactor process and a simulated baker’s yeast fermentation process. The results obtained from the use of the proposed technique show that the proposed technique can accurately cope with unknown disturbances and improve process operation from batch to batch.