Efficient and Simple Gaussian Process Supported Stochastic Model Predictive Control for Bioreactors using HILO-MPC

Model-based control of biotechnological processes is, in general, challenging. Often the processes are complex, nonlinear, and uncertain. Hence modeling tends to be complex and is often inaccurate. For this reason, non-model-based control strategies developed via fask, bench-scale, or pilot plant experiments are often applied in the biotechnology industry. Model-based control and optimization techniques can increase processes’ performance and automation level, thereby decreasing costs and guaranteeing the desired specifications. These rely on a model of the process to make predictions and optimize the inputs to the plant. To improve the quality of the models, it is often helpful to use combined first principle and data-driven models together in a hybrid modeling approach which increases the model prediction capabilities. The residual uncertainty of the hybrid model should be taken into account in the control level to satisfy the process specifications and constraints. This paper proposes to use a stochastic model predictive control scheme that exploits a hybrid, Gaussian processes-based model. We outline the effectiveness of the stochastic model-based approach in combination with a suitable Kalman filter for state estimation considering an example biotechnological process. Furthermore, we underline that appropriate tools exist that allow the simple application of such methods even for the novice user. To do so, we use an open-source Python package — HILO-MPC, which allows the simple yet efficient formulation and solution of machine learning-supported optimal control and estimation problems