COMPOSITION INFERENCE USING NEURAL NETWORKS AND INFERENTIAL CONTROL IN THE ETHYLBENZENE PROCESS
Emanuella Francisca de Lacerda Vieira1; Paulo Romero de Araujo Mariz1; Marcelo da Silva Pedro1; Jonas Laedson Marinho da Silva Santos1; Arioston Araújo de Morais Júnior1; Leopoldo Oswaldo Alcazar Rojas1
1 Universidade Federal da Paraíba
doi:10.20906/CPS/CILAMCE2017-0475
Resumo
For the implementation of process control systems focused on optimization, such as advanced control, it is necessary to continuously measure the properties of various plant streams, like the product composition of a process unit. However, in most cases, there are no real-time measurements of these quantities, or these data are only available at a very low frequency. Therefore, soft sensors are increasingly applicable in this area, and they consist of algorithms that can be estimate, from measurable variables at high frequencies (such as flow rates, pressures, temperatures, other compositions available), composition of products whose direct measurement is not possible or can not be performed at the desired frequency. The use of Artificial Neural Networks (ANNs) for the construction of soft sensors has been gaining more and more space, due to their non-linearity, generalization and adaptability of complex problems. This work aims to construct a methodology of soft sensors for a productive process of ethylbenzene (EB), which consists of two reactors in series and two columns of distillation with liquid recycle streams, and was interesting for the study due to occurrence of transient effects in the high purity compositions of the top and bottom streams of the second process distillation column. A methodology was built in stages, being the first stage of modeling, followed by simulation (software Aspen Plus and DynamicsTM) and equipments sizing involved in the process. Then, a key-variable selection algorithm with multivariate statistics was developed with the methods of all possible regressions and stepwise regression. Thereafter, were used two types of ANN architectures: multi-layer perceptron and Elman's recurrent network, to estimate and control compositions using Matlab® software. The results obtained showed that the ANNs were able to approximate the values of the compositions of the real values obtained by the mathematical model, and the estimates were evaluated by means of the mean square error (MSE) and root m
Palavras-chave: Neural Networks; Inferential Control; Soft Sensors; Real-time measurement