Controle inferencial em colunas de destilação utilizando redes neurais artificiais dinâmicas

Abstract

The measurement of the composition variable in real time is of paramount importance for the control of the industrial processes, due to the need to follow the final quality of the product and to the possibility of applying an effective control system. However, a number of obstacles make this goal impractical, either for reasons where the appropriate sensor does not exist or because these analytical measurements are too expensive, such as the online process analyzer. Some models of system identification are used for inference of chemical variables, among which artificial neural networks (RNA), with virtual sensors estimating difficult-to-measure variables through easily measurable input variables. Thus, the work proposes a methodology for the construction of virtual sensors implemented in software with the objective of estimating the output composition in a productive process of ethylbenzene (EB) and use it in process control. The production process in question was interesting for the present work, considering the occurrence of transient effects in the compositions of high purity in the top and bottom streams of the second distillation column. In addition, the production of ethylbenzene represents the beginning of the production chain of styrene, a monomer that has a large market in Brazil and worldwide, requiring more than 25 million tons per year worldwide. The results of the work range from the design of the unit to the construction of virtual sensors and control proposal. Therefore, it was verified that the overall balance of the unit presented results consistent with those of the literature, the variable selection algorithm showed efficiency, the virtual sensor using artificial neural networks presented satisfactory results and the control structure minimized the transient effects of the process. Keywords: virtual sensor, simulation, distillation column, variable selection, neural networks