Author:
Dobrowolski Maciej,Iciek Jan
Abstract
Part 1 of this article (Dobrowolski and Iciek, 2012) demonstrates how the predictions of classic mathematical models of sugar juice colour during their concentration in evaporator stations prove to be effective only for preliminary evaporator stages (1, 2, 3) – that is, while the formation of juice colour is still relatively weak. Additionally, the application of classic models to simulate juice-colour changes during redesign or modification of the evaporator station reflects poor compatibility with measurement data, specifically for the Smejkal model, Vukov model and de Visser model (Dobrowolski and Iciek, 2012). Therefore, two neural models were developed for raising thin juice colour in an evaporator station, and are presented in this part of the article. They are a neural (classic) model without feedback and a recurrent neural model (with feedback). Because the classic neural model did not sufficiently simulate juice-colour formation in the evaporator station, the recurrent neural model was developed to use feedback from measured juice colour. The recurrent neural model for colour measurements requires only the thin juice colour and process conditions in sequential apparatuses of the evaporator station. In addition, this model includes the initial pH value of thin juice and the initial concentration of invert sugar. The prediction results from the recurrent model correlate with the measurement results at the level of R2sim.ANN = 0.98 (for n ~ 780 industrial measurements). For this model, the mean prediction error of juice colour formation is øsim.ANN = 98 IU – a value several times lower than analogous errors resulting from the use of classic mathematical models. The data for design and accordingly revision of classic and neural models were collected during the research performed in four sugar factories operated by Südzucker Polska between 2005 and 2009. The proposed neural model (ANN) of rising sugar juice colour in sugar evaporator stations, presented in this paper, has been compared with three classic models (those of Vukov, de Visser and Smejkal).
Publisher
Verlag Dr. Albert Bartens KG
Reference46 articles.
1. Beale, M.; Hagan, M.; Demuth, H. (2010): Neural Network Toolbox™ 7. Mathworks.
2. Beyou, S.; Grondin-Perez, B.; Benne, M.; Damour, C.; Chabriat, J.-P. (2009): Control improvement of a C sugar cane crystallization using an auto-tuning PID controller based on linearization of a neural network. World Academy of Science, Engineering and Technology 54, 190–195
3. Biyanto, T., Ramasamy, M., Zabiri, H. (2007): Modeling heat exchanger using ANN. International Conference on Intelligent and Advanced Systems, 120–124
4. Dobrowolski, M. (2011): The use of artificial neural networks in the sugar industry. Techical University of Lodz, PhD Thesis
5. Dobrowolski, M.; Iciek, J. (2000a): The possibility of application of artificial neural networks in the sugar industry – prediction of the beet sugar yield before the campaign. The Sugar Newspaper / Gazeta Cukrownicza 12, 201