Cascade Gaussian Process Regression Framework for Biomass Prediction in a Fed-batch Reactor

Abstract

Model-based control of a fed-batch bioreactor requires an accurate dynamic model of the bioprocess. Process dynamics in a bioreactor can be highly non-linear making it difficult to identify phenomenological models with large numbers of model parameters, especially in real time. In the present work, the Gaussian process regression (GPR) algorithm is used to build a fed-batch bioreactor model using a cascade structure. This model predicts the biomass concentration in response to a given substrate feed-rate profile using three cascaded GPR sub-models, each predicting hold-up, dissolved oxygen (DO) and biomass respectively. A mathematical model of an industry fed-batch fermentation process is used to depict the kinetics in a bioreactor. Firstly, open-loop sub-models are trained and tested with data generated using the mathematical model. Later, these fine-tuned open-loop sub-models are integrated sequentially into a closed-loop cascaded GPR structure. The cascaded GPR model is validated in a closed-loop environment with the solution obtained using a mathematical model. Various model performance metrics such as RMSE, MAE and MAPE are calculated to determine the accuracy of each sub-model and final cascaded GPR fed-batch bioreactor model.