(256f) A Parametric Hybrid Nonlinear Model For Model Predictive Control Of A Batch Fermentation Process In A Bio-Reactor

Model Predictive Control (MPC) is now a widely accepted control methodology in many industrial control applications. Advanced Process Control applications require high-fidelity models. One of the main challenges in the implementation of MPC strategies has been the development of robust, accurate, and computationally efficient models. This challenge has been significantly greater in processes that are described by nonlinear models and involve varying dynamics. MPC applications typically use empirical models developed using historical data to capture the process characteristics. These models are subject to problems such as correlated, incomplete, and noisy data, and standard empirical modeling or training techniques can not use available first-principles information. First Principles models are difficult to use in MPC applications because they are often incomplete, computationally inefficient, and do not capture the specific process nuances captured in the historical data. This work describes a modeling framework that captures the advantages of both modeling paradigms and leverages them in an optimization-based training algorithm. This framework combines first-principles knowledge with neural network models in two ways: by allowing constraints reflecting first-principles information to be imposed on the optimization training problem, and by allowing the training models to be composed of neural network and first-principles models. The approach generates models that have better predictive capabilities and have better extrapolation characteristics than when using empirical models alone. The modeling framework is demonstrated through the development of a parametric hybrid nonlinear model for a batch fermentation process in a bioreactor by using both empirical data and fundamental process knowledge. The developed model is then used for the model predictive control of the bioreactor. The accuracy and computational efficiency of the model developed using this strategy underscores the efficacy of the framework for producing models that are useful for model-based control and optimization applications.