Multivariable Control and Real-Time Optimization - An Industrial Practical View (2005)
Multivariable control (mvc) and real-time optimization (rto) are now available. A perception exists that implementing advanced control on a process unit solves once and for all the optimization problem of that unit. Mvc generates benefits in three ways: stabilizing the process, protecting the process from violating operating constraints and using available degrees of freedom.
Multivariable Control and Real-Time Optimization - An Industrial Practical View (2005)
Multivariable control (mvc) and real-time optimization (rto) are now available. A perception exists that implementing advanced control on a process unit solves once and for all the optimization problem of that unit. Mvc generates benefits in three ways: stabilizing the process, protecting the process from violating operating constraints and using available degrees of freedom.
real-time optimizationan industrial practical view Here's what can be achieved with each technology and how they are integrated at this plant O. ROTAVA and A. C. ZANIN, Petrobras, Brazil P rocess optimization, once mainly a subject of academic interest, was made possible when tools such as multivari- able control (MVC) and real-time optimization (RTO) became available. These tools are a consequence of the move from analog to digital instrumentation and development of the distributed control system (DCS). Among people not familiar with control theory, and even among process engineers, a perception exists that implementing advanced control in the form of a multivariable controller on a process unit solves once and for all the optimization problem of that unit. To the disappointment and frustration of control engineers that implement such systems and managers responsible for automation investments, evaluating multivariable controller payback after it has operated for a while shows that profits are smaller than those planned. MVC generates benefits in three ways: stabilizing the process (decoupling the manipulated variables), protecting the process from violating operating constraints and using available degrees of freedom. This allows some constraints to become active, optimiz- ing the ptocess by maximizing yield oFthe most valuable products and minimizing expenditure oi raw materials and utilities. The first and second points are usually true: MVC stabilizes the process and protects it from violating constraints. Hence, the frustration is largely due to no satisfaction From point three. Such dissatisfaction is based on a real and a False assumption. The real assumption: By lack oFconstant tollow-up by a process engineer, MVC operation tends to degrade with time. Eventual changes in plant operation and production objectives require changes in the controller structure, sometimes immediately after commissioning. These changes are not done because experts are not available. The most common reasons for degradation are: inadequate limits of the manipulated variables, conflicting specifications of the con- trolled variables and inadequate tuning. The false assumption is that MVC gives a complete solution to the unit optimization problem. Actually, MVC is required to perform functions that are not in its designed scope. Examples are distributing the heat load between reflux and pumparounds in an atmospheric distillation unit and defining the reaction temperature of maximum conversion in an FCC unit. These features are not covered by MVC. Only an RTO system, based on a rigorous process model, takes care of both of them. This unfair view oFwhat can be achieved with MVC in terms of ptocess optimization allows continuous questioning of the heavy investments applied in upgrading instrumenta- tion. In an MVC and RTO implementation project, invest- ments in instrumentation and a DCS, if not already in place, represent the largest cost. On the other hand, it is important that such investments should be made anyway, even without implementing MVC and RTO because of analog instrumenta- tion obsolescence. Nevertheless, upgrading instrumentation and DCS implemen- tation do not define a competitive advantage because competitors are obliged to do the same. The convenience oF implementing MVC and RTO tools provides the real advantages. The objective of this article is to show clearly through examples what can be achieved in terms of optimization ftom MVC and what can only be obtained through RTO via a rigorous process model. Suppliers of automation and optimization technology usually split total benefits related to petroleum refining operation optimi- zation equally between MVC and RTO applications. According to Cutler and Perty,' available benefits for online optimization plus advanced control can amount to 610% of the added value of a given process. Multivariable control: steady-state quadratic opti- mization and dynamic control. Iwo distinct algorithms {Fig. 1) executed at the same frequency perform MVC. These algorithms are associated with solving the process control static and dynamic problems. The algorithm that corresponds to solving the static problem, executed First, searches a set of optimum values of the manipulated variables, maximizing a quadratic objective function, namely operating proFit. Solving this problem must satisfy constraints established by maximum and minimum values oi the controlled and manipulated variables. Controlled variables are evaluated at the steady state produced by the linear dynamic process model, with manipulated variables and disturbances as inputs. The result- ing optimum operating point is then sent to the objective ftmction of the dynamic problem. The second algorithm solves the dynamic problem and accom- plishes two tasks. The first consists of keeping the process inside the operating region, compensating for the frequent disturbances HYDROCARBON PROCESSING JUNE 2005 61 SPECIALREPORT PROCESS AND PLANT OPTIMIZATION subject to: Y'""" <Y'<Y that enter the plant and would lead the unit to violate its con- straints. The second consists oi implementing the optimum val- ues infotmed by the static solution oft he first algorithm. The mathematical procedure that satisfies the requirements solves a least-squares ptoblem. The quadratic optimization (first task) solves the following problem: (1) (2) (3) (4) /here: w,,, = vector of manipulated variables values at present time Us - vector of manipulated variables values at steady state W\ = diagonal matrix of economic coefficients ofthe manipu- lated variables (partial derivatives ofthe economic func- tion in relation to the manipulated variables) W2 - diagonal matrix of suppression factors ofthe manipu- lated variables Wi, - diagonal matrix of slack variable weights Ys = vector of predictions ofthe controlled variables on the steady state (those that must strictly satisfy the chosen boundaries) Y^' = vector of controlled variable predictions at steady state (those whose boundaries do not need to be strictly satis- fied) SCV= vector of slack variables (represents how much each variable in Y^' surpasses its boundary; they are added to guarantee existence of a solution to the optimization problem). The dynamic control (second task) is formulated as: (6) submitted to the following constraints: -A""-(y7')<A(;T)<A(;T); y- 1 nl (7) u""' [jT) <u^+ A ( / T ) < H-" ( ; T) ; ; = \,...,nl (8) where: H / = control horizon nr = prediction horizon 7"= sampling or algorithm period execution A = vector of control action size u* = vector of optimal manipulated variable values calculated by the linear optimizer W4 = diagonal matrix of weights on the dynamically controlled ^ Linear optimizer (1 min) - steady-state linear model - QP or LP algorithm Optimum values: 0*', y'' ^ Controller (1 min) -dynamic linear model - predictive multivariable algorithm * Setpoints (manipulated), u Variables, DCS - regulatory control Two distinct algorithms (linear optimizer and controller) perform multivariable control. variables tt^^ = diagonal matrix of manipulated variable supression fac- tors W(, = matrix of weights to lead the manipulated variables to their respective optimum values yi= vector of upper or lower boundaries ofthe dynamically controlled variables yj, = vector oi linear prediction ofthe dynamically controlled variables. Operating region. Performance of MPC: optimum on t he constraints. The term operating region designates the polyhedron ot dimension equal to the number of manipulated variables. This region encloses all feasible unit operating condi- tions. The surface that limits the region is defined by the operating constraints, made up mainly by the controlled variables. Neverthe- less, the operator can also make active constraints related to upper and lower boundaries ofthe manipulated variables, making the operating region even more restricted. Obviously, this practice is not recommended, because controller freedom is reduced and the solution eventually achieved is not the best. Fig. 2 illustrates an operating region for an FCC unit converter where two variables are considered: con\ersion (controlled variable) and teed flowrate (manipulated variable). A solution ofthe MVC linear or quadratic optimizxr corresponds to a constraint vertex in Fig. 2. For instance, if the optimal operating points are represented by point A (maximum conversion) or B (maximum feed load) the MVC will fmd the best solution, pro\'ided the convenient coefficients; ofthe manipulated variables (feed flowrate and reactor temperature) are given to the linear or quadratic optimization objective function. If the largest coefficient corresponds to reaction temperature, the optimum solution will correspond to point A, otherwise the operating optimum will be point B. If the operating economical optimum alternates between points A and B, the relative values of the aforesaid coefficients need to be changed tor the right solution to be found by the optimizer. Overall, the optimizer has tuning parameters that are the coefficients oi the manipulated variables in the linear or quadratic programming {Wi m Eq. 1) that must be changed according to the energy cost or product prices. 62 JUNE 2005 HYDROCARBON PROCESSING SPECIALREPORT PROCESS AND PLANT OPTIMIZATION Gas compressor constraint Max catalyst circulation rate Max feed temperature Feed pump constraint Min. conversion Feed rate Operating region of the FCC converter with two variables. If the operating optimum is unconstrained (for instance, an interior point of the shadowed area on Fig. 2), ir cannot be found by the MVC quadratic optimizer. Multivariable controller optimization algorithm limitation. The hmitation of the MVC consists of its inability to find an optimum sokition when the optimum is inside the operating region, since the controller quadratic programming algorithm only fmds solutions on the boundaries defined by the process constraints. As an example, consider the diesel production of an atmo- spheric fractionator (Figs. 3 and 4). Fig. 3 shows that an economical balance exists between the increase of diesel production through its heavy fraction and the energy consumption in the furnace. For the same feed flowrate and diesel ASTM 85% distillation specification, an increase of the coil outlet temperature (COT) corresponds to an augment of the o\erflash that then improves fractionarion, allowing a larger diesel draw. A limit to this procedure is when a diese! production increase does not compensate economically for the additional energy consumption. Again in Fig. 3, we see the optimum operating point is a func- tion of an economical balance between the energy cost and diesel price. For instance, tor a larger energy cost, the optimum operat- ing point consists of a smaller COT and consequently smaller amount of ovetflash. Fig. 4 illustrates the diesel production increase as a function of the fractionation constrained by the maximum ASTM 85% specification. Starting from an operating point with low overflash flowrate, SV', a COT increase allows the process to reach an optimum overflash flowrate, 5V'"', which is the best economical balance between energy consumption and diesel production. Operating with excess overflash. SV*, the economical gain with the additional diesel production does not compensate for the extra energy consumption cost. If the optimization task was attributed to the MVC, it would lead the proce.ss to the maximum COT, provided no constraints were violated, even if maximum COT was not the most economi- cally advantageous operating condition. In diesel optimization by MVC, the process is operated with fixed overflash flowrate setpoint (calculated offline by the pro- Energy consumption/feed rate FI G. 3 An economical balance exists between increasing diesel production and energy consumption. cess analysis team), in spite of changes in petroleum quality and fractionator operating conditions. Real-time optimization. Optimization through MVC is hindered by the simplicity of the process model, which is a set of linear equations. The objective function is related to maxi- mum or minimum values of the MVC manipulated variables. A quadratic or linear programming algorithm achieves the optimal solution. According to Marlin and Forbes,^ due to the reliability and relative simplicity of MVC technology, the preferred location tor economic optimization would be the controller, provided it performed well. On the other hand, the RTO model consists of a ser of nonlin- ear equations that represent as close as possible the system steady state. The objective function consists of the system economic model that translates into its profitability. A nonlinear program- ming algorithm achieves the optimal solution. Optimization makes the connection between production plan- ning and scheduling tasks and those evaluated by MVC. Fig. 5 illustrates the traditional RTO structure of a system with multiple MVCs, with the interrelationship of its components.^ Its main components are: Process steady-state model. The mathematical model must represent the system over a wide range of operating conditions with a high degree of accuracy in such a way that the maximum profitabilit)' predicted by the objective function be effectively the maximum potential profit of the real process. Also, constraints of the real process must not be violated when the optimizer solution is implemented. According to Friedman,'^ effective optimization is still limited by availability of good models. Data reconciliation and model parameter updating. Rig- orous models require a large amount of measured information that contains uncertainties. This subject is dealt with by the data reconciliation procedure. The first data reconciliation procedure consists of gross error detection. In this step, invalid measure- ments due to instrumentation malkmctioning are identified and treated. Afterward, the small differences in the mass and energy balances are spread among measurements throughout the pro- cess model, taking into consideration the statistical uncertainties of the instruments and redundant measurements. Rigorous models are expected to accurately reproduce the real 64 JUNE 2005 HYDROCARBON PROCESSING SPECIALREPORT PROCESS AND PLANT OPTIMIZATION process over a wide range of operation conditions. Parameters of such models are updated to compensate for nonmcasured distur- bances and changes in process performance caused by factors such as catalyst deactivation, heat exchanger fouling, and furnace and columns efficiency reduction. By analogy to a traditional PID con- troller, model parameter updating using plant data corresponds to process Feedback. Optimization algorithm. After the model is properly fit- ted ro real plant data, optimization is performed. To attain the optimal solution of the optimization problem, a nonlinear pro- gramming (NLP) algorithm is used. Such a solution corresponds to the maximum profitability of the unit inside the operating region limited by process constraints. The NLP algorithm usu- ally employed in industrial applications is successive quadratic programming (SQP).^ The real-time optimization problem is defmed as: Diesel ASTM 85% SV Atmosphetic residue Delta die5el (economic benefit) V ol., % F I G . 4 Diesel production is a function of the fractionation constrained by the maximum ASTM 85% specification. subject to the constraints: <<; (9) (10) (11) (12) (13) where: - Constraints limits - Parameters - Instrumentation data Scheduling/planning - Economic data - Constraints: feed/products Data base/information system Optimizer Targets Process model NLP solver Multivariable controller 1 Multivariable controller 2 Setpoints Setpoints Setpoints DCS - regulatory control Traditional real-time optimization structure. i/; = steady-state disturbances vector ^,,y = economic objective function h^ = economic model constraints hp = nonlinear model constraints X; = vector of constraints on the steady state. Integrating the real-time optimizer with the MVC. RTO determines the optimum operating point to which the process must be driven. This solution cannot be implemented directly in the DCS due to dynamic constraints. High-frequency disturbances would destabilize the process and move it outside its constrained region. This can happen because the low-fre- quency optimizer is not able to deal with such disturbances. By its characteristics, MVC is the adequate tool to move the process in a robust way to such an operating optimum. The optimizer solution is normally made a\'ailable to the MVC in the form of optimum steady-state targets of its manipulated and/or controlled variables (Fig. 6). Fig. 6 illustrates the classical structure of RTO integrated to the MVC, available in most industrial applications. This strategy displays two main functions: a) Optimization of the steady state, which is accomplished by the RTO in a relatively low frequency (superior layer). b) Implementing the optimal solution achieved in (a) above by the MVC (interme- diate and lower layers), which is responsible by the dynamic driving of the process from the present state to the optimum operating point calculated by the RTO. MVC execution is divided in two parts: Dynamic control, which is executed in a relatively high Frequency and is responsible for keeping che process inside the envelope established by its constraints and for moving the unit to its optimum operating point. Linear optimization, which is executed in the same frequency as the dynamic con- trol, whose function is to send the RTO solution to the MVC, but making small adjustments to it due to disturbances enter- ing the process inside the RTO execution interval. Thus, each task of the control and opti- mization strategy is distinct and displays its own algorithm, which is executed sequen- tially at the same MVC run. Integrating RTO with MVC is achieved through the objective function of the MVC Data reconciliation model updating Multivariable controller N 66 JUNE 2005 HYDROCARBON PROCESSING SPECIALREPORT PROCESS AND PUNT OPTIMIZATION RTO {4h) - steady-state rigorous model - economic model of process - nonlinear programming {NLP) I Optimum values: Manipulated variables, u'"' Controlled variables, y'"' Linear optimizer (1 min} - steady-state linear model - QP or LP algorithm Optimum values: u'', y'' Controller (1 min) -dynamic linear model - predictive multivariable algorithm Setpoints (manipulated), u Variables y.u DCS - regulatory control FI C. 6 The optimizer solution is normally made available to the multivariable controller in the form of optimum steady-state targets of its manipulated and/or controlled variables. linear optimization layer. In this way, when the RTO is active, Eq. 1 is modified to: mm U, JiCV (14) subject to constraints represented by Eqs. 2-5, where: u " = vector of optimal values of rhe manipulated variables determined by the RTO. The first term of Eq. 14, which represents implementation of the optimizer solution, does not possess the tuning parameter present in Eq. 1. In this case, the information embedded in the vector of economic weights, W,, is informed through the product price.s in the RTO economic function. Case studi es. Next, some variables of a crude distillation unit and a. fluid catalytic cracking (FCC) converter are analyzed. In the first case studied, it is shown that MVC optimization must be complemented by RTO. In the other cases, the optimization task is exclusively made by RTO. Cases that can be optimized exclusively by MVC, i.e., their optimum is on the constraints, are not analyzed. Maximizing jet fuel in a petroleum fractionator and pressure balance on an FCC converter when compressors/blowers are operating in their full capacity are examples of such cases. Crude atmospheric distillation unit: COT and overflash flowrate. COT is manipulated to optimize diesel production, generally constrained by maximum diesel ASTM 85% distillation temperature. Fig. 7 illustrates diesel production and overflash control once furnace outlet temperature has been fixed. Overflash is adjusted by the setpoint of the external reflux. The heavy diesel Feed TTT Overflash (optimizer Overflash is adjusted by the setpoint of the external reflux. The heavy diesel production is determiend by a mass balance over that particular region of the column. production is determined by a mass balance over that particular region of the column. The MVC of the fractionator column increases the COT until the diesel ASTM 85% distillation temperature becomes active. Fig. 8 illustrates the influence of furnace outlet temperature and overflash flowrate on the diesel quality. When operating with a deficient overflash, SV~, the MVC increases the COT until operation point 1 is attained, where Vi is the diesel flowrate and the diesel ASTM 85% distillation temperature constraint becomes active. As the overflash flowrate is optimized, SV'", increasing the controller flowrate setpoint without increasing the COT, a new operating condition represented by point 2 is achieved in which the mass and energy balances of the heavy diese! region reduce the diesel flowrate to Vj. In this condition, there is a giveaway in the ASTM 85% diese! speciflcation. Starting from point 2, the MVC increases the COT until the diesel ASTM 85% distillation temperature constraint becomes active again at point 3. In this condition, diesel flowrate, V'3, is increased by A V' in relation to initial point 1. Operating with excessive overflash, SV'', the operating point represented by 4, the economic gain with the additional diesel production, A^'*, does not compensate for the additional energy expenditure. In this case, we verify that the MVC optimizers benefltsby maximizing the COTdepends on the overflash flowrate, whose optimum value can be determined by the RTO since it depends on the feed quality and many other variables of the atmospheric column. Determining the optimum overflash flowrate, as shown, is not part of the scope of the MVC, thus the fractionation problem requires an RTO based on a rigorous model. Distribution of pumparounds in a fractionation column. Distribution of energy withdrawal through the pumparounds {Fig. 9) is exclusively an RTO problem. An economic balance exists between the energy recovery by the pumparounds and the internal reflux along the column that are responsible for the products fractionation. In the case shown in Fig. 9, the distribution of pumparounds 68 JUNE 2005 HYDROCARBON PROCESSING SPECIALREPORT PROCESS AN D PLANT OPTIMIZATION umparounds distribution: - Frattitniation x energy recovery Diesel ASTM 85% Optimum stripping steam: - Kerosene flash point - Bottom light components Heavy naphtha _ T^ Steam Diesel Delta diesel ( economic RTO benefit) When operating with a deficient overflash, the multivariable control increases the COT until operation point 1 is attained. must consider the following internal refluxes: i ] and i g, responsible for the fractionation, respectively, of the kerosine light and hea\y ends L^j, responsible for the fractionation between diesel and atmospheric residue I46, that corresponds to overflash flowrate. www, GEAJ ET. CO M More than just vacuum We provide optimal, process in- tegrated solutions for any type of vacuum system, using our well known jet pumps together with other types of vacuum pumps. Our jet vacuum pumps are used in oil refineries, steel degassing, deodorizing and dehumidifying of vegetable oils/fatty acids, sea water desalination and vjriou.s applications in the chemical, phar- maceutical and food industries. The opiimum design of a jet vacuum pitmp requires a lot of experience. We have it. GEA Jet Pumps GEA Jet Pumps GmbH Einsteinstrasse 9-15 76275 Ettlingen Germany Tel.: +49 7243 705-0 E-mail: i nfogeaj et.de Internet; www,geaj et.com Select 95 at www.HydrocarbonProcessing.com/RS Atmospheric residue Distribution of energy withdrawal through the pumparounds is exclusively a real-time optimization problem. As consequence of the strong system nonlinearities, the col- umn duty distribution is not conveniently handled by MVC, which operates hased on linear models. In cases where the MVC manipulates pumparound flowrates, their setpoint values remain almost permanently on their operating limits, not being effectively manipulated MVC variables. Additionally, heat load distribution must consider pressure drops in the pumparounds regions to optimize pressure in the vaporization zone. This task is exclusive of an optimizer with a rigorous model because MVC does not take into account the column tray hydraulics for flooding detection. Stripping steam flowrate. Kerosinc .stripping steam is meant to be the fine-tuning factor of tbe product flash point property; mean- while, stripping steam in the bottom of the fractionator removes the light components of tbe atmospheric residue. Only through a rigor- ous process model is it possible to determine the optimum stripping steam flowrates. Tbe optimizer can determine an optimum product/ steam ratio, which depends on crude quality, and the steam flowrate can be controlled by regulatory control in the DCS. Stripping steam manipulation cannot he bandied by MVC because the process model (obtained from plant tests) depends on the steam rate at the very moment of the test. For instance, if the steam flowrate is above a certain value, the controlled prop-^ erty is not sensible to it. Therefore, this is a system with strong nonlinearities in whicb tbe economic gain may change its signal. For instance, sometimes it is economically convenient to increase the steam injection; other times the optimum is obtained by the opposite procedure. Preheat trains and furnace feed allocation. One of the open questions in optimizing distillation unit operation consists of allocating the crude feed to preheat trains and furnaces. When active constraints exist, the allocation problem can be solved with MVC alone by keeping tbe process variables equally away from the constraints. 70 PROCESS AND PLANT OPTIMIZATION On the other hand, when there is no active constraint, feed allocation becomes a rigorous model optimization with an eco- nomic objective function that minimizes energy consumption. Column reboiler duty. In a distillation column with top and bottom products (for instance, stabilizer and crude pre- flash), normally the effective manipulated variables are reflux flowrate and reboiler duty, and the controlled variable is a property oft he top product. In such a situation, several pairs of reflux flowrate and reboiler duty values satisfy the same top product specification. The difference among those pairs of manipulated variables is translated into larger or smaller distillate amounts. It becomes clear that a point of economic optimum must exist. Operating outside this point causes distillate product loss or excess energy consumption. This optimum point can be determined by a rigor- ous model and an RTO, but cannot be determined by MVC since it does not consider the trade-off between energy consumption and product fractionation. FCC converter: reaction temperature. To keep a converter stable, MVC normally constrains the operating reaction tem- perature to a value in the range of 1C to 2C. This is practically equivalent to keeping such a variable at a fixed setpoint. This operating procedure is adopted because the MVC optimizer tends to augment the conversion by increasing reaction temperature up to the compressor limit. Nevertheless, the operating optimum can occur before such a constraint is activated. In this case, operat- ing on the constraint (compressor limit) causes overcracking, result- ing in inadequate yield profile or high olefin content in the cracked naphtha. Hence, determining the optimum reaction temperature is an optimization problem with a rigorous model. In practical operation, the values that bound the narrow reac- tion temperature operating range are calculated "offline" hy the process analysis engineering as a function of an economic ohjec- tive and sent to the MVC to be sought. Regenerator dense phase temperature. In the total combustion FCC converter, the regenerator dense phase tempera- ture is a consequence ofthe burned coke on the spent catalyst. The regenerator control is performed through the excess of oxygen. O scar R ot ava is a senior process control engineer for Petrobras. He graduated as a chemical engineer at the Federal University of Rio de Janeiro and joined Petrobras, Brazil's largest oil company, the same year and worked in the training department, lecturing on fluid mechanics (pipe and pump design), thermodynamics and process control. Presently Dr. Rotava works with the optimization group, commission- ing multivanabie process control systems on distillation and FCC units in several of the company's refineries. He is in charge of the corporate mass balance implementation program. Dr. Rotava got an MSc degree in chemical engineering from the Federal University of Rio de Janeiro and a PhD degree in process control from the Imperial College, London. A nt oni o C arl os Z ani n is a senior process control engineer for Petrobras automation group. He received his BSc in chemical engineering from the Federal University of Rio Grande do Sul. Dr, Zanin holds an MSc (dissertation in predictive multivariabie control) and PhD (thesis in real-time optimization] degrees in chemical engineering from the University of Sao Paulo. He worked in developing Petrobras's multivariable control technology. Presently, Dr, Zanin is responsible for developing and implementing inferential property algorithms, advanced control and real-time optimization on crude and FCC units at the company's refineries. In the partial comhustion FCC converter, the regenerator dense phase temperature is related to the CO/ CO2 ratio in fuel gases and to coke content in the regenerated catalyst. In the latter case, regenerator temperature optimization must consider the catalyst/oil ratio and the coke content on the regen- erated catalyst, which affect the conversion in opposite ways. For instance, a reduction in the regenerator temperature increases the catalyst/oil ratio and, consequently, the conversion. On the other hand, the coke content on the regenerated catalyst Increases, too. Therefore, the combined action can decrease conversion. Balanc- ing these effects can only he accomplished through an optimizer with a rigorous model. HP LITERATURE CITED ' Cutler, C. R. and R. T. Perry, "Real time optimization with multivariable control is tequired to maximize profits," Computers and Chemical Engineering, V. 7, n. 5. pp, 663-667, 1983. ^ Marlin, E. T. and J. R Fotbes, "Selecting the proper location for economic optimization: multivariabie control or RTO," NPRA Computer Conference, National Petroleum Refiners Association, paper CC-95-125, Nashville, Nov. 6-8, 1995. ^ Hardin, M. B., R. Sharun, A. Joshi and J. D. Jones, "Rigorous crude unit optimization," NPRA Computer Conference, National Petroleum Refiners Association, paper CC-95-122, Nashville, Nov. (i-8, 1995. * Friedman, Y. Z., "Closed-loop optimization updateWe are a step closer to fulfilling the dream," Hydrocarbon Processing, HPIn Control, January 2000. '' White, D. C, "Online optimization; what, where and estimating ROK" Hydro- carbon Ib-ocessing, pp. 43-51, June 1997. PcTROLAB www.petrolab.com Select 96 at www.HydrocarbonProcessJng.com/RS 71