REV. CHIM. (Bucureºti) ♦ 58 ♦ Nr. 4 ♦ 2007406

Comparison between Different Control Approaches of the UOP Fluid

Catalytic Cracking Unit

VASILE MIRCEA CRISTEA1 AND PAUL ªERBAN AGACHI“Babeº-Bolyai” University of Cluj-Napoca, Faculty of Chemistry and Chemical Engineering, 11 Arany Janos, 400028 Cluj-Napoca,Romania, tel: +40 264 593833

Different traditional and advanced control strategies have been tested and their performance investigated.The considered control approaches are: conventional decentralized PID control, Model Predictive Control(MPC) based on analytical model, MPC based on Artificial Neural Networks (ANN) model and Fuzzy LogicControl (FLC). Comparative control results are shown for multivariable control structures, aimed to controlthe main variables of the unit in the presence of typical disturbances. The results reveal incentives of themodel based control approaches (with analytical or ANN models) over the classical decentralized PID controlscheme, but also the benefits of the Fuzzy Logic Control.

Keywords: PID, Model Predictive Control, Artificial Neural Networks, Fuzzy Logic Control

* email: mcristea@chem.ubbcluj.ro

Importance of the catalytic cracking process has beenconserved at high level during its continuous development.Its products, such as: LPG, fuel gas, gasoline, diesel,aviation kerosene, slurry oil, and mostly the high-octanegasoline, are of continuously increasing demand on themarket, resulting in major benefits of the overall oil refinery.

Modelling and control of the FCC process have beengrowing in importance during the last decades as they maybring significant profit. Their task is to cope with challengingtasks such as: complex raw material characteristics, non-linearity and interactions between process variables. Alarge amount of effort has been devoted both in academiaand industry to study different aspects of FCCU such as:chemistry of the catalytic cracking, modelling [1], nonlineardynamic behaviour [2] with steady state multiplicity andchaotic characteristics, dynamic simulation [3], on lineoptimisation and control [4].

A special interest is devoted to find the way optimaloperation may be achieved by advanced control. FCCU isdifficult to control due to its multivariable, strong interactingand nonlinear features leading to very complex dynamics.Usually, the optimum operating conditions are met whereconstraints are also active. As a consequence, it isimportant to control FCCU as close as possible to theoperating and equipment constraints but without violatingthem. Model Predictive Control proves to be a goodcandidate for performing such tasks but it requires a robustmodel of the unit. Development of an analyticalmathematical model implies several simplifyingassumptions. These assumptions usually refer to the needof lumping the individual components of the feed andproducts into groups, the simplification of complicatedprocesses occurring on the catalytic surface during thecracking process or during coke removal from the spentcatalyst and the complex heat and mass transferphenomena describing the fluidised bed [5]. All theseassumptions limit the capability of developing accuratedynamic simulators based on first principle models.

Statistical models developed by means of ANN and usingprocess data are promising alternatives to the traditionalfirst principle models [6-8]. They may be also used in a

mixed neural networks and first principle modellingapproach. Built on the basis of a consistent set of data,ANN models are able to capture the complexity of theintrinsic processes featuring the global process behaviourand motivate the ANN based modelling approach.

Description of the UOP FCCU modelsThe FCCU of UOP (Universal Oil Products) type, for

which first principle and ANN based modelling and FLCcontrol investigations have been performed, is presentedin figure 1.

A mechanistic complex dynamic first principle modelhas been previously developed, based on construction andoperating data from an industrial UOP FCCU working inpartial combustion mode [4]. Models describe the reactor-regenerator assembly consisting in its main components:feed and preheat system, reactor, regenerator, air blower,wet gas compressor and catalyst circulation lines. The mainfractionator is included in the models only by its buffervessel effect on the flow of gaseous products produced inthe reactor. The FCCU is operating in partial combustionmode.

The Weekman’s triangular kinetic model has beenconsidered to describe the phenomena taking place in thereactor. Reactor is divided in two sections: riser andstripper. The riser model is built on the assumptions of idealplug flow and very short transient time compared toregenerator’s time constants. Mass balance equationsreveal the gasoline and coke+gases production. Equationsalso account for the amount of coke deposited on catalystand for the cracking temperature dynamics. A CSTR modelwas used for the stripper in order to evaluate thetemperature in the stripper and the fraction of coke onspent catalyst.

The regenerator model has a higher complexity due itsoverall rate determining role for the dynamic behaviour ofthe whole process. It consists in two zones: a dense bedzone and a zone of entrained catalyst. The dense bed zoneconsists of two phases: a bubble phase of gaseousreactants and products moving up the bed in plug flow

REV. CHIM. (Bucureºti) ♦ 58 ♦ Nr. 4 ♦ 2007

and a perfectly mixed dense phase containing gases andsolid catalyst. The dense phase is assumed perfectly mixedand the temperature is assumed uniform in the bed. Thegaseous phase is considered in equilibrium with the densephase. Because catalyst entrainment is also present in thezone above the dense bed the model describes it. Theregenerator model consists in mass and heat balanceequations for O2, CO, CO2 and coke, but also in heat balanceequations for the solid and gaseous phase. These balanceequations are complemented with equations describingentrained catalyst in the zone above dense bed, catalystflow and pressure in the regenerator.

Due to the small time constants, for the catalyst flow inthe spent and regenerated catalyst circulation lines it isassumed the steady state behaviour. Spent andregenerated catalyst circulation considers a single-phaseflow, based on force balance equations.

Since the modelling is the most time consuming part ofthe MPC design and the choice of proper modellingenvironment is crucial, one major challenge related toindustrial MPC applications is the efficient developmentand identification of the control-relevant model [9]. Findingthe proper balance between the accuracy of the modeland computational load is usually very challenging [10].Typically, the model of the process used by the MPCalgorithm is of first principle type. As the complexity of thefirst principle model increases, the computationalrequirement for solving the MPC optimisation problem isbecoming more and more restrictive, especially in thenonlinear case. For such situations the real timeimplementation may become unfeasible. As the ANNmodels are generally demanding less computing power,the use of an ANN based model may bring the desiredimprovement in decreasing the computation effort, thusincreasing speed of calculations.

An ANN model of the FCCU has been developed forusing it as inherent model used by the MPC algorithm. AsANN inputs has been considered the set of states/outputsof the first principle model together with the manipulatedvariables, all considered at the current moment of time t.The ANN outputs (targets) have been selected as consistingin the same set of state/outputs variables but consideredat the next moment of time t+Ät. Slide valve position onboth spent and regenerated catalyst circulating lines havebeen regarded as most effective manipulated variables.

The ANN has been trained to predict the values of thechange in the state/output variables, from one sample timeto the next one, based on the current values of the states/outputs and manipulated variables. A two-layer feed-forward/back-propagation training ANN was used forcomputing the network biases and weights.

Both the first principle and ANN developed modelssucceed to reflect the dynamic behaviour of the FCCU.They will be further used for building the dynamic simulatorand for the internal models used by the MPC algorithm.

Results and DiscussionsControl Approaches

Different control strategies have been applied for FCCUcontrol during the development of the modern controltheory. Starting with the classical control approach thatuses the PID controllers, the FCCU control has been laterused as a challenging benchmark for model based controltechniques using both first principle and statistical models.Development of first principal models involves severalsimplifying assumptions: lumping the individualcomponents of the feed and products into groups, thesimplification of complicated processes occurring on thecatalytic surface during the cracking process or during cokeremoval from the spent catalyst, associated to the complexheat and mass transfer phenomena occurring in thefluidised bed. All these assumptions limit the capability ofdeveloping accurate dynamic simulators based on firstprinciple models but challenge the modelling approachbased on the means offered by ANN. Other artificialintelligence instruments, such as fuzzy logic and geneticalgorithms may bring improved performance to the controllaw.

Based on analysis of current industrial FCCU operationand literature survey, a set of process variables has beenselected as having first role importance in efficient and safeoperation of the unit. The controlled variables consideredin the study are: inventory of catalyst in the regeneratorWr, regenerator temperature Treg, reactor temperature Trand the oxygen concentration in stack gas xO2sg. Themanipulated variables are: spent catalyst valve position(flowrate) svsc, regenerated catalyst slide valve position(flowrate) svrgc, air vent flowrate V7, and stack gas flowrateV14. The coking characteristics (coking rate constant) ofthe feed oil KC (+3.2% step increase) and main fractionator

Fig. 1. Representation of the UOP Fluid Catalytic Cracking Unit

407

REV. CHIM. (Bucureºti) ♦ 58 ♦ Nr. 4 ♦ 2007

pressure drop Äpfrac (+10% step increase) have beenconsidered as representative disturbances.

Results for FCCU PID and MPC controlPairing of controlled and manipulated variables used

for the PID decentralized control has been suggested bythe Relative Gain Array study [4]. They are: Wr-svsc, Treg-svrgc, Tr-V7 and xO2sg-V14. Compared to previousinvestigations of the control structure design based on RGA,the present work considers a larger dimension of thecontrol scheme (i.e. 4´4 controlled´manipulatedvariables), revealing incentives of the proposed controlapproach [4-5]. Anti-windup PID digital controllers havebeen used. MPC has been applied for the control of thesame process variables. The tuning has been made in away to achieve good control performance for both testdisturbances taken into consideration. Results obtained forthe case of the coking rate KC disturbance are presentedin figure 2.

Results reveal superior behaviour for the case of thenonlinear MPC, based on scheduling linearization, bothwith respect to overshoot and response time. The MPCtuning was performed by choosing the output and inputweights based on the maximum allowable deviation of thecorresponding variables, followed by a trial and errorrefining step. Following the performed simulations it may

be concluded that, as the number of controlled variable ishigher and the interactions between them is considerable,the MPC multivariable control strategy is more efficientcompared to the decentralized PID approach. The MPCcontrol proves to be the preferred control approach, asFCCU control also implies satisfying safety, equipment andoperating constraints.

Results for ANN based MPC controlControl of the catalyst inventory in the reactor-stripper

Wr and temperature of the regenerator bed Treg, has beeninvestigated in the presence of the coking rate constant KCdisturbance, with regenerated svrgc and spent svsc catalystslide valve position as manipulated variables. Simulationresults are presented in figure 3. They show the twocontrolled variables for the case of the Kc disturbanceapplied at moment t=50000 s and starting with initialconditions of the controlled variables different from thesetpoint values set for time t>0.

Simulation results of the MPC based on ANN modelshow a good setpoint following performance alsodemonstrating zero offset. The disturbance rejection abilityis efficient as short settling time and low overshoot proveit. The ANN model based control approach also features asignificant reduction of the computation time. The

Fig. 2. Comparative resultsfor PID (dashed line) and MPC(solid line) control of: catalyst

inventory in the regenerator Wr,regenerator temperature Treg,reactor temperature Tr and

oxygen concentration in stackgas xO2sg, in the presence of the

coking rate KC disturbance.

Fig. 3. ANN based MPC of thecatalyst inventory Wr in the

reactor and regeneratortemperature Treg, in the presence

of the Kc disturbance

408

REV. CHIM. (Bucureºti) ♦ 58 ♦ Nr. 4 ♦ 2007

Fig. 4. FLC of Wr, Treg, Tr and xO2sg in the presence of the Kc disturbance

reduction of the computation time is about one order ofmagnitude when the simulation is performed using theANN model, compared to the use of the first principlemodel.

Results of FCCU fuzzy logic controlThe investigated FLC approach considered triangular

and trapezoidal membership functions in order to fuzzifyboth manipulated and controlled variables. The fuzzy rulesemerged from the traditional operation of the FCCU.Mamdani’s inference method and centroid defuzzificationwere used together with a set of cross-connected rulesaimed to perform decoupling and efficient control. Integralaction has been also added to the control law in order toreduce steady state offset. Results of the fuzzy logic controlfor the FCCU main variables are presented in figure 4. Theyshow very short settling time and reduced overshoot, butalso small steady state offset.

ConclusionsSimulation results reveal the incentives of the model-

based control over the traditional decentralized PIDapproach. Moreover, the ANN model based MPC offers theopportunity for achieving reliable control using statisticalmodels built on the basis of process data, but also forconsiderably reducing the computational effort. The realtime application may substantially benefit on the one orderof magnitude computation time diminution. Despite itssimplicity, the FLC proves to have the shortest settling timeand overshoot although integral action should be addedin order to reduce the steady state offset. Nevertheless,the capability of MPC for performing optimal control in thepresence of constraints makes it most relevant. Combiningincentives of the investigated control approaches newcontrol systems with increased performance may bedeveloped for FCCU safe and efficient operation.

NomenclatureANN- Artificial Neural NetworksCSRT- Continuous Stirred Tank ReactorFCCU- Fluid Catalytic Cracking Unit

FLC- Fuzzy Logic ControlF3- feed flowrate [kg/s]F4- slurry recycle flowrate [kg/s]Kc- reaction rate constant for coke production [s-1]LPG- Liquefied Petroleum GasMPC- Model Predictive ControlRGA- Relative Gain Arraysvsc- spent catalyst slide valve position [0-1]svrgc- regenerated catalyst slide valve position [0-1]t- time [s]Tr- reactor temperature [°C]Treg - temperature of regenerator bed [°C]UOP- Universal Oil ProductsV7 - position of the air vent valve [0-1]V11 - position of the wet gas compressor suction valve [0-1]V14 - position of the stack gas valve [0-1]Wr- inventory of catalyst in the reactor-stripper [kg]xO2sg- oxygen concentration in the stack gas [0-1]∆t- sampling time [s]∆P

frac- pressure drop across reactor-main fractionator [Pa]

References1. MCFARLANE, R.C., REINEMAN, R.C., BARTEE, J.F., GEORGAKIS, C.,Comp. Chem. Engng., 17, nr.3, 1993, p. 2752. KARLA, L., GEORGAKIS, C., Ind. Eng. Chem. Res., 33, nr.12, 1994,p. 30633. ARBEL, A., HUANG, Z., RINARD, I. H., SHINNAR, R., Ind. Eng. Chem.Res., 34, nr.4, 1995, p. 2284. CRISTEA, M.V., AGACHI, P.S., MARINOIU, V., Chem. Eng. Proc., 42,nr.2, 2003, p. 675. ARBEL, A., RINARD, I.H., SHINNAR, R., Ind. Eng. Chem. Res., 35,nr.7, 1996, p. 22156. VIEIRA, W.G., SANTOS, V.M.L., CARVALHO, F.R., PEREIRA, J.A.F.R.,FILETI, A.M.F., Chem. Eng. Proc., 44, nr.8, 2005, p. 8557. ALARADI, A.A, ROHANI, S., Comp. Chem. Eng., 26, nr.3, 2002, p. 4018. YU, D.L., GOMM, J.B., Cont. Eng. Pract., 11, nr.11, 2003, p. 13159. JIA, C., ROHANI S., JUTAN A., Chem. Eng. Proc., 42, nr.4, 2003, p. 31110. HASSAPIS, G., Microproc. Microsys., 27, nr.7, 2003, p. 327

Manuscript received: 10.04.2007

409