Advanced search
Start date
Betweenand


Optimization of an industrial process for isoprene production using neural networks.

Full text
Author(s):
Rita Maria de Brito Alves
Total Authors: 1
Document type: Doctoral Thesis
Press: São Paulo.
Institution: Universidade de São Paulo (USP). Escola Politécnica (EP/BC)
Defense date:
Examining board members:
Cláudio Augusto Oller do Nascimento; Saul Gonçalves D'Ávila; Roberto de Campos Giordano; Frank Herbert Quina
Advisor: Cláudio Augusto Oller do Nascimento
Abstract

This work describes the application of a three-layer feed-forward neural network (NN) in different areas of chemical engineering. The main objective of this study is to model, simulate and optimize a real industrial plant, using NN by replacing phenomenological models. The industrial process studied is the isoprene production unit from BRASKEM. The chemical process consists basically of a dimerization reactor and a separation column train. Since NNs are able to extract information from plant data in an efficient manner, for this work, the neural network model was built directly from historical plant data, which were collected every 15 minutes during a period of one year. These data were carefully analyzed in order to identify and eliminate gross error data and non-steady state operation data. The modeling using NN was carried out by parts in order to get information on intermediate streams. Then, the global model was built, by interconnecting each individual model, and used to simulate and optimize the process. The optimization procedure carries on a detailed grid search of the region of interest, by a full mapping of the objective function on the space of decision variables. The second stage of this work deals with the azeotropic prediction using also the neural network approach. The objective of this step was to obtain a better understanding of the system behavior in the isoprene extraction section. Since all the cases studied are non-linear, complex andmultivariable systems, NN approach appears as a technique of interest due to its capability of learning the system without knowledge of the physical and chemical laws that govern it. Comparisons between the model\'s prediction and the experimental data were performed and reasonable results were achieved from an industrial point of view. ) Using neural network approach provides more comprehensive information for an engineer\'s analysis than the conventional procedure. This work shows that the use of NN methodology is promising for several industrial applications, such as data analysis, modeling, simulation and optimization process, as well as thermodynamics properties prediction. However, success in obtaining a reliable and robust NN depends strongly on the choice of the variables involved, as well as the quality of available data set and the domain used for training purposes. (AU)