Nonlinear Model Predictive Control based on multi-scale models: is it worth the complexity?

Abstract

Nonlinear model predictive control (NMPC) has been used to control a variety of bioprocesses. As NMPC requires full state information, a nonlinear moving horizon estimator (NMHE) is used to reconstruct the states from the available measurements. In context of bioprocesses, both NMPC and NMHE typically use macroscopic mass balance type models. These models ignore all intracellular information that is available via, e.g., Genome-Scale Metabolic networks. However, the use of these models adds significantly to the computational complexity of the controller. In this paper, we aim to assess whether incorporating metabolic networks in an NMPC framework leads to improvement in the tracking performance of a continuous bioreactor. To avoid complex bilevel optimisation problems, our NMPC scheme utilises the dynamic Metabolic Flux Analysis framework. In this framework, the intracellular fluxes are estimated from the measurements of extracellular metabolites. This is achieved by an NMHE scheme. The tracking performance of this metabolic network-based NMPC-NMHE is compared to a simple macroscopic model based NMPC in which all the states, i.e., the extracellular metabolites are assumed to be measured. In our case study, the tracking performance of the metabolic network-based NMPC is marginally better than the performance of the macroscopic model based NMPC. The added advantage of metabolic network based NMPC comes from the insight they offer from the estimated fluxes. Moreover, the performance and the insights obtained depend on the complexity of the metabolic network used.