Sensores virtuais empregando modelos de inferência da composição(responsável pela qualidade) dos produtos de uma coluna de destilação correspondem a medidores implementados em software, capazes de estimar, em tempo real, a composição dos produtos da mesma, a partir de informações do tipo temperaturas e pressões em diversos pontos da coluna e vazões de entrada, de saída e de reciclo. O objetivo deste trabalho é obter esse tipo de sensor para uma coluna de destilação, capaz de estimar instantaneamente a composição dos produtos no topo de uma coluna de destilação multicomponente com condensador parcial, empregando redes neurais artificiais. Foi desenvolvido um simulador dinâmico baseado em modelo não-linear da coluna para aquisição de dados. Neste projeto foi incluído um estudo sobre a influência do treinamento parcial no desempenho do sensor virtual. A idéia é estudar o desempenho para o caso de um sensor virtual treinado de antemão, com dados coletados a partir de um simulador da coluna. Este procedimento disponibiliza um sensor operacional, treinado através de um conjunto de dados simulados ou através de um pequeno conjunto de pontos e retreinado, quando dados reais ou um conjunto maior de dados estiver disponível. Outra contribuição importante é o estudo realizado sobre os principais erros que podem ocorrer neste tipo de sensores, que são raramente tratados em publicações científicas. É também proposta uma metodologia para detecção e correção destes erros que foram encontrados e que afetam o comportamento do sensor, alterando sua precisão e capacidade de ser utilizado em um controle inferencial da planta. / Soft sensors for composition inference models (that are responsible for the quality) of distillation column products, correspond to virtual instruments implemented in software. This software is able to estimate, in real time, the composition of the output products of the column, based on information such as temperature and pressure on several points of the column and on input, output and recycle flow. The purpose of this work is to obtain a soft sensor that estimates the instantaneous composition of the product at the top of a multicomponent distillation column with a partial condenser, employing artificial neural networks. The chosen architecture was the feedforward neural network with three layers. It was chosen based on many tested options. It was developed a dynamical simulator of this column for data acquisition based on a non-linear model. In this study, it was included an investigation about the influence of partial training in the performance of the soft sensor. The goal is to study the results achieved in the case of a soft sensor trained beforehand, with data acquired from the simulator of this column. This procedure makes possible to have an operational soft sensor, trained based on a simulated data set or on a small amount of points and then retrained when a real or larger data set is available. Another important contribution is the study performed about the main errors that may appear in this kind of sensor. These errors are rarely mentioned in scientific papers. It also aims at implementing techniques to enable detection and correction of those errors that the soft sensor may present, and that affect the performance of the soft sensor, changing its precision and making it inadequate for inferential control.

Colunas de destilação de alta pureza são sistemas de difícil controle. Apresentam longo tempo de resposta, comportamento altamente não linear e grande interação entre as variáveis. Os controladores preditivos são muito utilizados para o controle de colunas de destilação. Porém, em colunas de alta pureza, a incorporação de um único modelo linear geralmente acarreta em um controle de fraco desempenho. Isto pois, a representação do processo é deficiente, já que não considera variações nos ganhos e nas dinâmicas, típicas de um comportamento não linear. Estas incertezas podem, inclusive, provocar a instabilidade do controle o que resultaria em produtos que não atendam à especificação. Este trabalho tem por objetivo avaliar o desempenho dos controladores de horizonte de predição infinito com um modelo interno e com múltiplos modelos tendo o HYSYS(TM) como simulador de uma coluna de separação benzeno/tolueno e o MATLAB(TM) como ambiente para o controle supervisório. Observa-se que o controlador com apenas um modelo não é capaz de estabilizar o processo para perturbação nos valores de referência das variáveis controladas, ao contrário do controlador com múltiplos modelos. / High-purity distillation columns are systems which are typically difficult to control. The main reason for this is a strongly nonlinear and interactive system associated with a very sluggish response. Model Predictive Control is widely used for control of distillation columns. However, for high-purity columns, the use of a single linear model in the controller usually leads to a poor performance of the control system. The reason for this is the poor system representation, since variation in the system gains and time constants are not taken into account in the computation of the control law. Model uncertainties can produce instability in the control system and consequent deterioration of the product quality. The goal of this work is to evaluate the performance of infinite horizon MPC with a single internal model and with multiple models. HYSYS(TM) is used as simulator for the benzene/toluene column, and MATLAB(TM) is used as a platform for supervisory control. It is observed that the controller with a single model is not capable of stabilizing the process for disturbance in the set point of the controlled variables. Opposite to that behavior the controller with multiple models has a good performance.

In Brazil, especially in Rio Grande do Sul, there is a significant increase in ethanol production in small scale, i.e. in micro distilleries. These small-scale distilleries benefit small farmers, decentralize ethanol production and are based on the economic, social and environmental issue, but they do not hold enough technology that enables high performance in their processes, especially in the distillation stage. View of all this, it has developed this work, which aims to study the performance of a distillation column in permanent regime, developed and built at the Laboratory for Post-Graduate Processes Engineering using as feed a commercial ethanol solution and water, varying the concentration and the input flow into the column. It was opted for a standardized solution in order to achieve free results from interference by possible impurities. The concentration and flow of feed were set as independent variables and their limits established respectively in 5 and 9 °GL (% by volume) and 1848-4385 g/h. The concentrations of the distillate and of the bottom product were defined as dependent variables. Eleven tests were carried out according to the methodology of planning of experiments by using a central composite rotatable delineation, with triplicate at the central point. Three samples of distillate and bottom product were collected in flasks with lids, at intervals of fifteen minutes. Data were tabulated in spreadsheets from excel. Statistical analysis, using the software Statistica 7th version, was performed to evaluate whether the parameters: ethanol concentration and input flow of feed product interfere in the obtaining ethanol in patterns determined by ANP Resolution n°. 7, February 9th, 2011, i.e., Hydrous Ethanol Fuel (HEF) with ethanol concentration between 92.5 - 93.8°INPM. It was observed that, for the distillate, feed concentration and input flow, both in the quadratic form, are significant, i.e., there is the influence of these two variables in the obtaining of ethanol with higher alcoholic strength, but it was not able to reach the standard established by resolution of ANP. For the bottom product, it was found that neither concentration nor the input flow influence in their alcoholic concentration, possibly by design of the studied column. / No Brasil, principalmente no Rio Grande do Sul, há um aumento significativo na produção de etanol em pequena escala, ou seja, em microdestilarias. Essas microdestilarias beneficiam os pequenos agricultores, descentralizam a produção de etanol e tem como base a questão econômica, social e ambiental; porém não detêm tecnologia suficiente que permita um alto rendimento nos seus processos, em especial na etapa de destilação. Diante de tudo isso, desenvolveu-se este trabalho, que tem o propósito de estudar o desempenho de uma coluna de destilação em regime permanente, desenvolvida e construída no Laboratório de Pós-Graduação em Engenharia de Processos, utilizando como alimentação uma solução de etanol comercial e água, variando a concentração e a vazão de entrada na coluna. Optou-se por uma solução padronizada a fim de se alcançar resultados livres da interferência de possíveis impurezas. A concentração e a vazão de alimentação foram definidas como variáveis independentes e seus limites estabelecidos respectivamente em 5 e 9 °GL (% em volume) e 1848-4385 kg/h. As concentrações do destilado e do produto de fundo foram definidas como variáveis dependentes. Foram realizados onze ensaios de acordo com a metodologia de planejamento de experimentos, através da utilização de um delineamento composto central rotacional, com triplicata no ponto central. Foram coletadas, em frascos com tampa, três amostras de destilado e produto de fundo, em intervalos de quinze minutos. Os dados foram tabulados em planilhas do excel. A análise estatística, através do programa Statistica versão 7, foi realizada para avaliar se os parâmetros teor alcoólico e vazão de entrada do produto de alimentação interferem na obtenção de etanol nos padrões determinados pela Resolução ANP nº 7, de 9 de fevereiro de 2011, ou seja, Etanol Hidratado Combustível (EHC) com teor alcoólico entre 92,5-93,8 °INPM. Observou-se que, para o destilado, a concentração de alimentação e a vazão de alimentação, ambas na forma quadrática, são significativas, isto é, há a influência dessas duas variáveis na obtenção de etanol com maior teor alcoólico, porém não se conseguiu alcançar o padrão estabelecido pela Resolução da ANP. Para o produto de fundo, constatou-se que nem a concentração, nem a vazão de alimentação influenciam no seu teor alcoólico, possivelmente pelo desenho da coluna estudada.

Colunas de destilação de alta pureza são sistemas de difícil controle. Apresentam longo tempo de resposta, comportamento altamente não linear e grande interação entre as variáveis. Os controladores preditivos são muito utilizados para o controle de colunas de destilação. Porém, em colunas de alta pureza, a incorporação de um único modelo linear geralmente acarreta em um controle de fraco desempenho. Isto pois, a representação do processo é deficiente, já que não considera variações nos ganhos e nas dinâmicas, típicas de um comportamento não linear. Estas incertezas podem, inclusive, provocar a instabilidade do controle o que resultaria em produtos que não atendam à especificação. Este trabalho tem por objetivo avaliar o desempenho dos controladores de horizonte de predição infinito com um modelo interno e com múltiplos modelos tendo o HYSYS(TM) como simulador de uma coluna de separação benzeno/tolueno e o MATLAB(TM) como ambiente para o controle supervisório. Observa-se que o controlador com apenas um modelo não é capaz de estabilizar o processo para perturbação nos valores de referência das variáveis controladas, ao contrário do controlador com múltiplos modelos. / High-purity distillation columns are systems which are typically difficult to control. The main reason for this is a strongly nonlinear and interactive system associated with a very sluggish response. Model Predictive Control is widely used for control of distillation columns. However, for high-purity columns, the use of a single linear model in the controller usually leads to a poor performance of the control system. The reason for this is the poor system representation, since variation in the system gains and time constants are not taken into account in the computation of the control law. Model uncertainties can produce instability in the control system and consequent deterioration of the product quality. The goal of this work is to evaluate the performance of infinite horizon MPC with a single internal model and with multiple models. HYSYS(TM) is used as simulator for the benzene/toluene column, and MATLAB(TM) is used as a platform for supervisory control. It is observed that the controller with a single model is not capable of stabilizing the process for disturbance in the set point of the controlled variables. Opposite to that behavior the controller with multiple models has a good performance.

Sensores virtuais empregando modelos de inferência da composição(responsável pela qualidade) dos produtos de uma coluna de destilação correspondem a medidores implementados em software, capazes de estimar, em tempo real, a composição dos produtos da mesma, a partir de informações do tipo temperaturas e pressões em diversos pontos da coluna e vazões de entrada, de saída e de reciclo. O objetivo deste trabalho é obter esse tipo de sensor para uma coluna de destilação, capaz de estimar instantaneamente a composição dos produtos no topo de uma coluna de destilação multicomponente com condensador parcial, empregando redes neurais artificiais. Foi desenvolvido um simulador dinâmico baseado em modelo não-linear da coluna para aquisição de dados. Neste projeto foi incluído um estudo sobre a influência do treinamento parcial no desempenho do sensor virtual. A idéia é estudar o desempenho para o caso de um sensor virtual treinado de antemão, com dados coletados a partir de um simulador da coluna. Este procedimento disponibiliza um sensor operacional, treinado através de um conjunto de dados simulados ou através de um pequeno conjunto de pontos e retreinado, quando dados reais ou um conjunto maior de dados estiver disponível. Outra contribuição importante é o estudo realizado sobre os principais erros que podem ocorrer neste tipo de sensores, que são raramente tratados em publicações científicas. É também proposta uma metodologia para detecção e correção destes erros que foram encontrados e que afetam o comportamento do sensor, alterando sua precisão e capacidade de ser utilizado em um controle inferencial da planta. / Soft sensors for composition inference models (that are responsible for the quality) of distillation column products, correspond to virtual instruments implemented in software. This software is able to estimate, in real time, the composition of the output products of the column, based on information such as temperature and pressure on several points of the column and on input, output and recycle flow. The purpose of this work is to obtain a soft sensor that estimates the instantaneous composition of the product at the top of a multicomponent distillation column with a partial condenser, employing artificial neural networks. The chosen architecture was the feedforward neural network with three layers. It was chosen based on many tested options. It was developed a dynamical simulator of this column for data acquisition based on a non-linear model. In this study, it was included an investigation about the influence of partial training in the performance of the soft sensor. The goal is to study the results achieved in the case of a soft sensor trained beforehand, with data acquired from the simulator of this column. This procedure makes possible to have an operational soft sensor, trained based on a simulated data set or on a small amount of points and then retrained when a real or larger data set is available. Another important contribution is the study performed about the main errors that may appear in this kind of sensor. These errors are rarely mentioned in scientific papers. It also aims at implementing techniques to enable detection and correction of those errors that the soft sensor may present, and that affect the performance of the soft sensor, changing its precision and making it inadequate for inferential control.

The target of diploma thesis was to improve author’s theoretical and practical design knowledge of process engineering. In the diploma thesis a Simplified Basic Engineering Project of distillation unit has been developed. It was necessary to become familiar with the process technology and formal requirements of Basic Engineering Project. The most important parts of Basic Engineering Project are material and heat balances. Material and heat balances are the corner stones of distillation unit equipment design. Parameters of process equipment are presented in the form of equipment datasheets. The specification of pipelines is based on material and heat balances too. Inevitable part of Basic Engineering are Process Flow Diagram and Piping and Instrumentation Diagram (PFD & PID). Process Flow Diagram and Piping and Instrumentation Diagram include all equipment, piping and basic control loops.

This diploma thesis is about design of implementation of the ISA-95 standard for model of distillation column. The thesis consists of two parts, theoretical part and empiric part. In theoretical part, based on scientific literature there are defined terms about regarding standard ISA-95 in which are described individual models of standard ISA-95. Next chapter of diploma thesis is focused on FactoryTalk Services Platform, in which are introdced applications made by RockWell Automation, which are applied for this standard. Then there is description of distilation and model of distilation column. Last chapter of theoretical part is focused on industrial EtherNet/IP. This chapter blends into empiric part. In the next chapter of empiric part, there is design of implementation of standard ISA-95 using applications made by Rockwell Automation. Next chapter is focused on realization of the desing. Chapter of empiric part evaluates the entire project.

Industrial automation systems normally consist of four different hierarchy levels: planning, scheduling, real-time optimization, and control. At the planning level, the goal is to compute an optimal production plan that minimizes the production cost while meeting process constraints. The planning model is typically formulated as a mixed integer nonlinear programming (MINLP), which is hard to solve to global optimality due to nonconvexity and large dimensionality attributes. Uncertainty in component qualities in gasoline blending due to measurement errors and variation in upstream processes may lead to off-specification products which require re-blending. Uncertainty in product demands may lead to a suboptimal solution and fail in capturing some potential profit due to shortage in products supply. While incorporating process uncertainties is essential to reducing the production cost and increasing profitability, it comes with the disadvantage of increasing the complexity of the MINLP planning model. The key contribution in the planning level is to employ the inventory pinch decomposition method to consider uncertainty in components qualities and products demands to reduce the production cost and increase profitability of the gasoline blend application. At the control level, the goal is to ensure desired operation conditions by meeting process setpoints, ensure process safety, and avoid process failures. Model predictive control (MPC) is an advanced control strategy that utilizes a dynamic model of the process to predict future process dynamic behavior over a time horizon. The effectiveness of the MPC relies heavily on the availability of a reasonably accurate process model. The key contributions in the control level are: (1) investigate the use of different system identification methods for the purpose of developing a dynamic model for high-purity distillation column, which is a highly nonlinear process. (2) Develop a novel hybrid based MPC to improve the control of the column and achieve flooding-free control. / Dissertation / Doctor of Philosophy (PhD) / The operation of a chemical process involves many decisions which are normally distributed into levels referred to as process automation hierarchy. The process automation hierarchy levels are planning, scheduling, real-time optimization, and control. This thesis addresses two of the levels in the process automation hierarchy, which are planning and control. At the planning level, the objective is to ensure optimal utilization of raw materials and equipment to reduce production cost. At the control level, the objective is to meet and follow process setpoints determined by the real-time optimization level. The main goals of the thesis are: (1) develop an efficient algorithm to solve a large-scale planning problem that incorporates uncertainties in components qualities and products demands to reduce the production cost and maximize profit for gasoline blending application. (2) Develop a novel hybrid-based model predictive control to improve the control strategy of an industrial distillation column that faces flooding issues.

Thesis (PhD)--Stellenbosch University, 2012. / ENGLISH ABSTRACT: Distillation column design and operation require understanding of both the hydrodynamic and thermodynamic behaviour and limitations. One of the hydrodynamic aspects that negatively influence separation efficiency in the distillation column is entrainment of the liquid with the rising vapour or gas. Inaccurate entrainment predictions will lead to poor separation efficiencies in the column and consequently over design of the column diameter and/or height has to be incorporated. This has a significant impact on the capital cost due to the size and scale of industrial columns. Therefore, small improvements in entrainment prediction will lead to large savings in capital investment. Previous research published in the open literature focused primarily on the influence of gas and liquid flow rates and, tray geometry on entrainment for the air/water system. Consequently the non-air/water database is small and consists of data obtained from various tray and column geometries. As a result the accuracy of current entrainment prediction models is questionable for systems other than air/water. Therefore, the first objective of this work was to investigate whether current prediction models perform well for systems other than air/water. To prove this air/water, air/ethylene glycol and air/silicon oil data were measured and compared with current prediction correlations. It was found that current prediction models perform poorly for the air/ethylene glycol and air/silicone oil systems. At the same time a new observation was made with regard to froth development and behaviour inside the column. The observation shows that liquid flow rate has a nonmonotonic influence on entrainment, caused by the short (475mm) tray flow path. The second objective was to examine the influence of gas physical properties on entrainment. New entrainment data were measured by individually contacting air, CO2 and SF6 with water and ethylene glycol, while n-butanol was contacted with CO2 and SF6. The data was compared with current prediction models which performed poorly for SF6 results. This shows the inability of these models to predict entrainment for gas systems with high densities. Modified Reynolds and Froude numbers were developed to show the influence of gas physical properties on entrainment. Low modified Reynolds numbers and large modified Froude numbers resulted in high entrainment. The third objective was to determine the influence of liquid physical properties on entrainment. New entrainment data were measured using CO2 with Isopar G, n-butanol, water, silicone oil and ethylene glycol. Current prediction models compared poorly to the data and did not include the influence of liquid viscosity on entrainment. It was found that viscosity had an intricate non-monotonic influence on entrainment. The fourth and final objective was to correlate the influence of gas and liquid properties on entrainment as determined by the previous two objectives. To make the dataset more complete, entrainment was measured for four tray spacings using CO2/Isopar, CO2/nbutanol, air/ethylene glycol, CO2/ethylene glycol, air/silicone oil and CO2/silicone oil (over 1700 data points). Two new correlations are presented to predict the fraction of liquid entraining with the rising gas (L’/G with R2 = 85%) and the fraction of liquid entering the tray that entrains (L’/L with R2 = 92%). The performance of the L’/G correlation (R2 = 85%) is vastly superior to two other prominent correlations (R2 = 61% and 23%). This correlation can be implemented to predict entrainment successfully for different tray geometries by combining the predicted influence of tray geometry, by Kister and Haas (1988), with results from the newly developed correlation. All four objectives are presented as manuscripts for journal publication and serve as alone standing documents. / AFRIKAANSE OPSOMMING: Distillasie kolom ontwerp en bedryf vereis begrip van beide die hidrodinamiese en termodinamiese gedrag en beperkings. Een van die hidrodinamiese aspekte wat skeiding doeltreffendheid negatief beïnvloed in die distillasie kolom is meesleuring van die vloeistof met die stygende dampe of gas. Onakkurate meesleuring voorspellings sal lei tot swak skeiding doeltreffendheid in die kolom en gevolglik word die ontwerp van die kolom deursnee en / of hoogte beinvloed. Dit het 'n beduidende impak op die kapitale koste as gevolg van die grootte en skaal van industriële kolomme. Klein verbeterings in meesleuring voorspelling sal dus lei tot groot besparings in kapitaal belegging. Vorige navorsing gepubliseer in die oop literatuur het hoofsaaklik gefokus op die invloed van gas- en vloeistof vloeitempos en plaat geometrie op meesleuring vir die lug/water sisteem. Gevolglik is die nie-lug/water databasis klein en bestaan van die data wat verkry is uit verskeie plaat en kolom-geometrieë. As gevolg is die akkuraatheid van die huidige meesleuring voorspelling modelle vir stelsels anders as lug/water te betwyfel. Daarom is die eerste doel van hierdie werk om ondersoek in te stel of die huidige voorspelling modelle goed presteer vir stelsels anders as lug/water. Om dit te bewys was lug/water, lug/etileenglikol en lug/silikon olie data gemeet en vergelyk met die huidige voorspelling korrelasies. Daar is bevind dat die huidige voorspellings modelle swak presteer vir die lug/etileenglikol en lug/silikon olie. Op dieselfde tyd was 'n nuwe waarneming gemaak met betrekking tot dispersie ontwikkeling en gedrag binne die kolom. Die waarneming toon dat vloeistof vloeitempo 'n nie-monotoniese invloed op meesleuring het, veroorsaak deur die kort (475mm) plaat vloei pad lengte. Die tweede doelwit was om die invloed van gas fisiese eienskappe op meesleuring te ondersoek. Nuwe meesleuring data was gemeet deur individuele kontak van lug, CO2 en SF6 met water en etileenglikol, terwyl n-butanol slegs met CO2 en SF6 inkontak gebring was. Die eksperimentele resultate word vergelyk met die huidige voorspellings modelle wat swak presteer invergelyking met SF6 resultate. Dit toon die onvermoë van hierdie modelle om meesleuring vir gas stelsels met hoë digthede te voorspel. Gemodifiseerde Reynolds en Froude getalle was ontwikkel om die invloed van gas fisiese eienskappe op meesleuring aan te toon. Lae gemodifiseerde Reynolds getalle en groot gemodifiseerde Froude getalle lei na hoë meesleuring. Die derde doelwit was om die invloed van vloeistof fisiese eienskappe op meesleuring te bepaal. Nuwe meesleuring data is gemeet deur gebruik te maak van CO2 met Isopar G, nbutanol, water, silikon olie en etileenglikol. Huidige voorspellings modelle vergelyk swak met die data en sluit nie die invloed van vloeistof viskositeit op meesleuring in nie. Daar is gevind dat viskositeit 'n ingewikkelde nie-monotoniese invloed op meesleuring het. Die vierde en finale doelwit was om die invloed van die gas en vloeistof eienskappe op meesleuring soos bepaal deur die vorige twee doelwitte te korreleer. Om die datastel meer volledig te maak, is meesleuring vir vier plaat spasiërings met CO2/Isopar, CO2/n-butanol, lug/etileenglikol, CO2/ethylene glycol, lug/silikon olie en CO2/silikon olie (meer as 1700 data punte gemeet). Twee nuwe korrelasies word aangebied om die fraksie vloeistof wat meegesleur word met die stygende gas (L’/G met R2 = 85%) en die fraksie vloeistof wat die plaat binnetree wat meegesleur word (L’/L met R2 = 92%) te voorspel. Die prestasie van die L’/G korrelasie (R2 = 85%) is aansienlik beter as twee ander prominente korrelasies (R2 = 61% en 23%). Hierdie korrelasie kan suksesvol geïmplementeer word om meesleuring vir verskillende plaat geometrieë te voorspel deur die voorspelde invloed van plaat geometrie deur Kister en Haas (1988), met die resultate van die nuut ontwikkelde korrelasie te kombineer. Al vier doelwitte word as manuskripte vir joernaal publikasie aangebied en dien as alleenstaande dokumente.

O objetivo principal desta tese foi a implantação de uma nova estratégia para a integração da otimização em tempo real (RTO), com o controle preditivo multivariável em uma unidade de processo industrial. A solução proposta pode ser considerada como uma estratégia de uma camada, na qual os problemas de controle e otimização econômica são resolvidos simultaneamente, na mesma camada da estrutura de controle. Supondo que o objetivo econômico a ser maximizado (minimizado) seja uma função côncava (convexa) das entradas e saídas de processo, o controlador MPC com otimização econômica (OMPC) foi obtido através da inclusão do gradiente reduzido do objetivo econômico, na função objetivo do controlador preditivo. Esta abordagem foi testada inicialmente através da simulação do conjunto reator regenerador de uma Unidade de Craqueamento Catalítico Fluido (UFCC). O controlador otimizador foi implementado com sucesso em uma coluna de destilação de tolueno, na Unidade de Recuperação de Aromáticos da refinaria de Cubatão da Petrobras. Este controlador está em funcionamento contínuo por cerca de um ano, sem qualquer problema relatado. Para a determinação das condições ótimas, um modelo rigoroso de coluna de destilação multicomponentes no estado estacionário é incluído no controlador preditivo para permitir o cálculo online do objetivo econômico. A trajetória prevista para o sistema de destilação até o ponto ótimo é calculada utilizando-se um modelo linear dinâmico, o qual foi obtido através de testes em degrau na planta real. O ponto ótimo obtido através da estratégia proposta leva em consideração as restrições nas entradas manipuladas e a faixa de controle para as saídas. O problema de otimização resultante para cálculo das ações de controle é uma QP, que pode ser facilmente resolvida com os solvers disponíveis. O MPC com otimização econômica foi implementado como um módulo do pacote SICON (Sistema de Controle da Petrobras). / This thesis was mainly aimed at the implementation of a new strategy for the integration of real time optimization (RTO) with multivariable predictive control in an industrial process system. The proposed strategy can be considered as a one-layer strategy where the control and economic optimization problems are solved simultaneously in the same layer of the control structure. Assuming that the economic objective to be maximized (minimized) is a concave (convex) function of the process inputs and outputs, the optimizing model predictive control (OMPC) was obtained through the inclusion of the reduced gradient of the economic objective in the control objective of the predictive controller. The approach was initially tested through the simulation of the reactorregenerator of a Fluid Catalytic Cracking Unit (FCCU). The optimizing controller has been successfully implemented in a toluene distillation column at the Aromatic Recovery Unit of the Cubatão refinery of Petrobras. This controller has been in continuous operation for about one year without any reported problem. For determining the optimum operating conditions, a steady-state rigorous multicomponent distillation model is included in the predictive controller to allow the on-line computation of the economic objective. The predicted trajectory of the distillation system towards the optimum point is computed with a linear dynamic model that was obtained through step tests in the real plant. The optimum point that is achieved with the proposed strategy takes into account the constraints in the manipulated inputs and the zone control of the outputs. The resulting optimization problem that produces the control actions is a QP that can be easily solved with available solvers. The optimizing MPC was implemented as a module of the SICON (Petrobras Control System) package.