| dc.description.abstract | Bio-fermentation process is difficult to model given its use of living micro-organisms to produce useful products via complex reaction mechanisms. Their kinetics are hard to characterize; hence, approximate formulations are used when building a first-principles model. Consequently, such a model will be of poor accuracy. Recently, there is a lot of interest towards data-driven modeling as the amount of data collected, stored, and utilized is growing tremendously due to the advent of super-computing power and data storage device. Additionally, data-driven models are simple and easy to build but their utility is hugely restricted by the amount and quality of data used to develop them. Therefore, hybrid modeling is an attractive alternative to purely data-based modeling, wherein it combines a first-principles model with a data-based model resulting in improved accuracy and robustness. In this work, we develop a three-step method to build a hybrid model for a full-scale bio-fermentation process with a volume of over 100,000 gallons. Firstly, we improved the accuracy of the first-principles model via incorporating mathematical terms in its equations which are based on obtained process knowledge from a literature study. Secondly, we performed local and global sensitivity analysis to identify sensitive parameters in the improved first-principles model that have considerable influence on its prediction capability. Finally, we developed a deep neural network (DNN) based hybrid model by integrating the improved first-principles model with a DNN which is trained to predict the identified model parameters. The resulting hybrid model is more accurate and robust than the (original and improved) first-principles models as it is equipped with a trained DNN to predict the uncertain parameters and process states accurately.
Based on the developed hybrid model, a hybrid model-based observer was developed to track the different states present in the process. As the available measurements were fairly accurate, the open-loop observer was re-initialized with a new set of measurements whenever they become available. This method is computationally less demanding and was able to accurately estimate the states. Next, we build an optimal control algorithm on GAMS software to estimate the optimal operating conditions of the fermenter in real-time. This is carried out in order to maximize the product amount and minimize the cost by manipulating the inputs and taking practical constraints into account. The resulting control algorithm was able to improve the profitability and the productivity of the full-scale bio-fermentation process. | |
