interconnected modules. Each module is expressed by an input-output relationship based
on a particular modelling technique.
Several authors have used the very simple complementary hybrid model structure
represented in Figure 3.4a (e.g. Wilson and Zorzetto, 1997; Psichogios and Ungar, 1992;
Montague and Morris, 1994; Feyo de Azevedo et al., 1997) based on the use of artificial
neural networks for describing the microorganism kinetics embedded into mass balance
equations.
Feyo de Azevedo et al. (1997) analysed the practical advantages and disadvantages of
such a structure in terms of model accuracy and robustness. Neural networks are
powerful for kinetic modelling provided sufficient data is available covering the whole
operating region. Model predictions are most significantly degraded whenever the neural
network operates outside the input space used in the training phase. These aspects are
illustrated in Figure 3.5.
To improve robustness of the above mentioned hybrid structure, Simutis et al. (1995)
suggested to include a safety model which should be used whenever the ANN is
operating in extrapolation conditions. In this way the global extrapolation properties of
the model are improved. This type of approach is illustrated in Figure 3.6, where a
Monod-type model and an ANN model compete for the modelling of a baker’s yeast
fermentation.
BIOPROCESSES AUTOMATION AND CONTROL
One of the most important tasks for bioprocess automation and control is to force the
process state to follow an optimal path defined according to a pre-established economic
target. The economic performance of the production process is most often quantified by
Multiphase bioreactor design 74