ABSTRACT
Present work at the Industrial Control Centre involves the exploration and development of the most suitable techniques for the modelling and control of fermentation processes. Techniques such as Kalman filters, kinetic models and multi-layer perceptrons have been evaluated previously [1, 2, 3]. In this contribution, work performed for the modelling of an industrial fed-batch fermentation process using the Group Method Data Handling Neural Network (GMDHNN) will be presented [4]. Fermentation processes present the neurocomputing community with a number of challenges due to the non-linearities and dynamics of such processes. The architecture of a GMDHNN is introduced as well as the series of experiments performed using the available data. The experiments address the suitability off the GMDHNN as a selector of inputs, as a one step ahead predictor, and as a long term predictor. Current results seem to be promising. Networks generated for the one step ahead prediction of the residual carbon in the fermentation process provide accurate estimations but the long term prediction presents more problems.
