Automatic Control of Bioprocesses - Measurement, Monitoring, Modelling, and Control of Bioprocesses - page 54

without online adaption of the weights in the layers of the ANN. Based on the
estimation and a mass balance equation, the feed rate was calculated by the
controller to maintain the glucose concentration at the desired set-point. During
controlled substrate fermentations with concentrations between 0.8 and 1.0 gL -1
and specific growth rate around 0.2 h -1 , the functional efficiency of the control
algorithm was demonstrated. The calculation time for the weight adaption is 1-2 s,
allowing online implementation. They admit the need from a priori offline data,
mirroring different cultivation behaviors for training purposes, although in
industrial plants such information is usually available. Furthermore, they carried
out simulation studies with more than one measured variable and concluded that
more than one measured variable will significantly increase the precision of the
control. Especially the online adaption of the weighting factors of the ANN seems
promising, leading to a broader range of applications even outside the training
domain.
3.6 Model Predictive Control
In the model predictive control (MPC) strategy a dynamic model of a process is
applied to simulate the future evolution of the process depending on possible
simulated values of the controlled variable. Typically the future evolution will
only be calculated up to a predefined prediction horizon. Using an optimization
algorithm the best value of the controlled variable is calculated using a cost
function. Due to the fact that a differential equation system must be solved online,
MPC is computationally demanding. Therefore, for MPC a state estimator as well
as a controller is required.
Improved understanding of penicillin formation mechanisms, morphological
features, and the role of mycelia in the synthesis led Ashoori et al. [ 57 ]to
implement a detailed unstructured model of penicillin production in a fed-batch
fermenter. This model was used to implement a nonlinear MPC (NMPC) to control
the feed rate to increase penicillin formation. As the controller input they apply
online measurements of pH and temperature. They propose a novel cost function,
applying the inverse of the product rather than the common quadratic relation. This
is implemented to avoid ordinary differential equation solver problems where it is
not possible to guarantee the efficiency of set-point tracking. They compare the
control performance with a regular autotuned PID controller and identify the
NMPC as superior with higher process yields. The NMPC controls the acid as well
as the base flow, and the cooling water system. Due to the more sophisticated
model, the control achieves better performance than in a previous work by Birol
et al. [ 58 ]. To address the computational cost of this more detailed model, they
propose the application of a locally linear model tree (LoLiMoT) to simplify the
original nonlinear model, as described in the next section.
Certainly due to the high computational power that needs to be provided for a
MPC, many NMPC approaches are still only proven by simulation but have not yet