|
Poster Presentation 3-29
Hybrid Modeling with Neural Networks for Alcoholic Batch Fermentation
Elver Radke, Aline C. Costa and Rubens Maciel Filho
Laboratory of Optimization, Design and Advanced Control - LOPCA Department of Chemical Process – School of Chemical Engineering State University of Campinas – UNICAMP Campinas, SP – Brasil
Fax: +55 (0)19 3788-3910; E-mail: elradke@lopca.feq.unicamp.br
It is necessary to study ways of minimizing ethanol production costs to keep it economically competitive with the fuels currently used. Implementation of efficient control systems in alcohol industry plants is critical to achieve this objective. However, it is difficult to estimate the primary variables of the process due to the lack of sensors capable of measuring variables, in real time and with low costs. A solution for this problem is the use of software sensors, so the development of models capable of describing the process with a high degree of reliability is necessary. However, the modeling of bioprocessing is difficult due to the nature of the biochemical reactions involved being expensive, slow, difficult and models detailed in complex kinetic studies are not always available. The present work considers the study of neural network hybrid modeling of an alcoholic batch fermentation process, capable of describing the most representative features of the process, and being usable for optimization, control or as software sensors. The process balance equations are combined with neural networks, which are responsible for describing the kinetic rates of biomass growth, substrate consumption and product formation. From experimental studies carried out in a temperature range of 28 to 40°C, data for the training were obtained. With few experimental data available, the best data fit was obtained from the Sigmoidal-Boltzman function. To accomplish the testing, the algorithms of Levenberg-Marquardt and Bayesian-framework were used, all with the backpropagation method to update the connections weights. It was possible to estimate how many nodes (neurons) were required to achieve a suitable process representation. The input data for the neural networks were the concentrations of Biomass, Substrate and Alcohol (X, S and P), while the desired outputs were the kinetic rates of growth, substrate consumption and alcohol production. The hybrid neural model inputs and outputs were (X, S, P)t and (X, S, P)t+1, respectively, and yielded a good performance, especially considering the speed of its development compared to conventional models. In general, this technique appears to be very versatile and to have great potential for biotechnological process representation.
|