677

4.7 DCS: Control and Simulation Advances

T. L. BLEVINS, M. NIXON (2005)

Advanced Control A. AutotuningSoftware Packages That B. Fuzzy LogicAre Embedded in Modern C. Model Predictive ControlDCS Systems by Emerson: D. Neural Network–Based Variable Estimation

E. Performance Monitoring F. Process and Control Simulation

Partial List of Software Adaptive Resources (A, C) (www.adaptiveresources.com)Suppliers: Allen-Bradley (A, B) (www.ab.com)

Aspentech (C, D, E, F) (www.aspentech.com)ControlSoft (A) (www.controlsoftinc.com)Emerson (A, B, C, D, E, F) (EasyDeltaV.com)Expertune (A, E) (www.expertune.com)Gensym Corporation (D, E) (www.Gensym.com)Honeywell (A, B, C, D, E, F) (www.honeywell.com)Invensys (A, B, C, D, F) (www.invensys.com)Matrikon (D, E) (www.matrikon.com)Neuralware Inc. (D) (www.neuralware.com)Pavilion Technologies (C, D) (www.pavtech.com)Siemens (A, B, C, D) (www.siemens.com)Universal Dynamics Technologies (A, C) (www.unadyn.com)Yokogawa (A, B, C) (www.yokogawa.com/us)

This section discusses some of the advanced software pack-ages that are embedded in modern distributed control systems(DCS). They serve control, estimation, monitoring, simula-tion, and tuning functions. For the advanced control strategiesthat are not discussed in this section, refer to the followingsections in this volume:

•••

For in-depth discussions of auto-tuning, fuzzy logic, andmodel predictive control (MPC), which are also discussed inthis section, the reader is referred to the following sections:

••

Patents”

••

zation”••••••

INTRODUCTION

The computational requirements of software products foradvanced control have traditionally limited their implemen-tation. For that reason, their associated software resided eitherin a host computer or in dedicated control units that were

LT221

LIC221

Fuzzylogic

Surge tank level controlusing fuzzy logic

Flow sheet symbol

© 2006 by Béla Lipták

Section 2.20: “Optimizing Control”Section 2.33: “State Space Control”Section 2.34: “Statistical Process Control”

Section 2.13: “Model-Based Control”Section 2.14: “Model-Based Predictive Control

Section 2.16: “Modeling and Simulation of Processes”Section 2.17: “Model Predictive Control and Optimi-

Section 2.18: “Neural Networks for Process Modeling”Section 2.19: “Nonlinear and Adaptive Control”Section 2.27: “Sampled Data Control”Section 2.29: “Self-Tuning Controllers”Section 2.31: “Fuzzy Logic Control”Section 2.38: “Tuning by Computer for Optimizing”

678 Control Room Equipment

external to the basic DCS package and were “layered” ontothe plant’s control system. This situation is changing, as someof the more advanced software packages can be embeddedin the basic DCS system.

Recent advances in the processing power available withinsome DCS control systems have allowed them to includesome advanced control capability. These embedded softwarepackages include:

• Performance monitoring• Loop tuning• Fuzzy logic control• Model predictive control• Neural network-based variable estimation• Control simulation

Most of the successful advanced control applications, suchas MPC, that are in use today are used on high-throughputprocesses, where the savings justify the high cost of devel-oping, installing, and maintaining these layered control appli-cations. This new generation of advanced control is oftendesigned so that the average process engineer can use thesefeatures with the same level of confidence as he or she usedthe more traditional controls. Fuzzy logic control, modelpredictive control, and property estimation may be imple-mented as a function block that can be assigned to executein the standard redundant controllers.

This operating environment has improved security andperformance in comparison to the traditional approach ofleaving the advanced controls outside of the basic controlsystem, which is referred to as “layering.” For this embeddedimplementation, the engineering tools for the configurationand troubleshooting of traditional control may also be usedwith these advanced software blocks.

Also, the standard dynamos of these blocks support oper-ator displays that include advanced controls without the needto map parameters. Thus, the engineer and the operator canhave consistent interfaces to both the advanced and the tra-ditional control applications.

Embedding advanced control applications has many otheradvantages as well. Probably the most important is that theadvanced control application software runs inside the embed-ded controllers where the applications can fully participate incontroller redundancy, alarming, configuration downloads,and online upgrades. This also means that all parameter ref-erences are managed by the runtime and the configurationsystem, and such terms as “Mode” and “Status” have meaningsconsistent with the terminology of other software packages inthe system. To the operator and other operations personnel,the advanced control applications are largely invisible — theyare just other function blocks and parameters.

PERFORMANCE MONITORING

The maximum production rate and efficiency of a plant aredetermined by the process design and the capacity of the

equipment used. However, many plants never reach theproduction capacity that is inherent in the plant design andequipment. In addition, there usually is little time to study theprocess operation to determine whether it could be improvedbecause attention is concentrated on maintenance and onabnormal conditions that can become major sources of processdisturbances.

To improve on this situation, some distributed controlsystems are now designed to support automatic and contin-uous monitoring and detection of abnormal operation ofcontrol and field devices. When control is based on the Foun-dation fieldbus architecture, the status associated with eachfunction block output gives a direct indication of the qualityof the measurement of control signal. Also, the actual andnormal mode attributes of these blocks may be used to deter-mine whether a function block is operating in its designedmode. Thus, by continuously monitoring the status and modesupported by fieldbus function blocks, it is possible to auto-matically determine the following abnormal conditions incontrol and input-output (I/O) function blocks:

• A condition exists that could be reducing the accuracyof the measurement.

• Control effectiveness is limited by a downstream con-dition.

• The block is not running in its designed mode ofoperation.

Based on the inherent accuracy of the sensor and its maxi-mum error, it is possible to compute a variability index thatcompares the actual performance to the best achievable.These calculations can be built into the control and I/O func-tion blocks to support control analysis. By this approach, thecontrol performance of even the fastest processes can beevaluated.

Within the control system, the status and mode values areautomatically processed to determine the percent of time whenan abnormal condition existed during the current and previoushour, shift, or day. If this percentage exceeds a preset limit, thenthe condition is flagged as abnormal and the module containingthe function block will show up in the monitoring interface. Anexample of an abnormal condition monitoring interface isshown in Figure 4.7a.

CONTROLLER TUNING

Loop tuning can be done either when needed (on demand) orcontinuously by adaptive tuning. On-demand tuning is initiatedby the operating personnel. Continuous adaptive tuning is per-formed automatically after set-point changes or significant dis-turbances, or when noise is experienced. On-demand tuning isoften used during startup to obtain approximate tuning con-stants during the initial commissioning of a control system.Adaptive tuning corrects the tuning in response to changes inprocess dynamics.

© 2006 by Béla Lipták

4.7 DCS: Control and Simulation Advances 679

On-Demand Tuning

One of the most successful techniques for providing on-demand tuning is based on using the relay oscillation tuningmethod to identify the process ultimate gain and ultimateperiod. (For a detailed discussion of tuning by computer, refer

of ultimate process gain and ultimate time period, seeFigures 2.38c and 2.38i in that section.)

This “relay oscillation” tuning method can be used onboth self-regulating and integrating processes but is not suitedfor pure dead time processes, which require sample and holdcontrol. Once the ultimate gain and ultimate time period ofa loop are known, the tuning rules (see Sections 2.35 to 2.38)may be automatically applied to obtain the initial controllersettings.

The relay oscillation technique of tuning is performed inan open loop under two-state control. Each time the con-trolled variable (process output) crosses the set point or initialvalue of the controlled variable, the relay is switched. Relayaction causes the loop to oscillate at its ultimate period, Tu.The ultimate gain, Ku, is determined as the ratio between theamplitude of the two-state controller output and the ampli-tude of the oscillations in the controlled variable.

Upon completion of the relay oscillation test, the auto-tuner calculates the ultimate gain, ultimate period, dead time,time constant, static gain, and integrating gain of the process.The user can input the type of process (only self-regulating orintegrating), the type of tuning rules to use, and perhaps anadditional tuning factor such as the closed-loop time constant.The tuning rules are then applied and the calculated settingsare provided for the controller, as illustrated in Figure 4.7b.

Adaptive Tuning

Adaptive tuning can be used to tune both feedback and feed-forward controllers. There are two ways to design an adaptivePID controller: direct and indirect or identifier based. Anidentifier-based approach is advantageous when modelswitching and parameter interpolation are used. In this casea set of models is available together with a switching strategyamong them, and parameter interpolation is based on theintegrated squared error assigned to every value of the modelparameter.

The adaptation cycle continues through a declared num-ber of scans or until there is enough excitation on the inputs.As soon as a model has been updated, controller redesigntakes place based on the updated model parameters. Adap-tation can proceed for the whole model or separately for thefeedback or feedforward portions of the model. The externalexcitations can be injected into the feedback loop automati-cally. The applied excitations can be a small change of theset point or controller output in either the Manual or in theAutomatic modes.

Since the process model obtained is first-order plus deadtime, any tuning rules can be used, including applied, typi-cally lambda or IMC tuning. A general adaptive PID control-ler structure with model parameter interpolation is shown inFigure 4.7c.

FUZZY LOGIC CONTROL

Many industrial processes operated by humans cannot beautomated using conventional control techniques since the

FIG. 4.7aOperator’s display for the monitoring of abnormal conditions.

© 2006 by Béla Lipták

to Section 2.38 and for an explanation of the determination

680 Control Room Equipment

performance of these controllers is often inferior to that ofthe operators. One of the reasons is that linear controllers,which are commonly used in conventional control, are notappropriate for the control of nonlinear processes.

Another reason is that humans aggregate various kindsof information and combine control strategies, which cannot

be integrated into a single analytic control law. The underly-ing principle of knowledge-based (expert) control is to cap-ture and implement experience and knowledge available fromexperts (e.g., process operators).

One would usually consider fuzzy logic control only afterall attempts at controlling the process by optimized PID

FIG. 4.7bOperator’s display after “on-demand” tuning been performed. The lower line in the graph shows the oscillating step changes that aremade in the manipulated variable (usually the control valve opening), and the upper line shows the response of the controlled variable.

FIG. 4.7cThe configuration of the blocks of software in an adaptive PID controller, which includes both feedback and feedforward parts.

SP

Excitationgenerator

Controllerredesign

PID controllerw/dyn compfeed forward

Modelevaluation

Set ofmodels

Modelinterpolation

Manipulate

Process

Measureddisturbance

Supervisor

Adaptive control block

© 2006 by Béla Lipták

4.7 DCS: Control and Simulation Advances 681

controllers have failed. Because fuzzy logic control requiresexhaustive amounts of process data, the other essentialrequirement is that the process be somehow manually con-trollable. For a detailed discussion of fuzzy logic control,

The concept of fuzzy logic originated in 1965 when LoftiZadeh of the University of California at Berkeley proposed whathe called “fuzzy set theory.” Theoretically, fuzzy logic-basedcontrollers can provide significant performance improvementsover conventional PID-based controllers. The nonlinearity intro-duced in the fuzzy PID controller can reduce overshoot signifi-cantly without compromising disturbance rejection capability,but fuzzy logic has not been widely adopted in the processindustries.

One reason is that the tuning is difficult since no standardguidelines exist for the establishment of membership functionscaling. To address this issue, many manufacturers providepreengineered versions of fuzzy control that require little userinput to commission. In addition, manual tuning may besubstituted by relay-based automatic tuning. Such controllerscan be used as nonlinear single-loop replacements for con-ventional PID controllers.

Fuzzy logic controllers combine the concept of member-

tions and detailed explanations). The rules are used to determinea controller output based on the input membership functions.Tuning is accomplished through adjustment of the scaling asso-ciated with the membership functions. Figure 4.7d illustratesthe software block components of a fuzzy logic controller thathas already been defined.

The fuzzy logic controller in many cases providesincreased performance and improved response with little or noovershoot relative to PID control. According to some sources,

the improvement can range from 20% on the basis of integralof absolute error (IAE) to almost 50% for integral of absoluteerror multiplied by time.

MODEL PREDICTIVE CONTROL

In internal model controllers (IMC), MPC, and pole-cancel-lation, lambda, and Dahlin controllers, the model used by the

ing to some users, the MPC model relates to the inverse ofthe process model. Therefore, when the process dynamicsare unchanged (set point changes), these controllers are effec-tive; if the model is exactly known, they are better than PID.Therefore, people use them, when the process cannot tolerateset-point overshoots. Their main limitation is their sluggishresponse to changes in process dynamics (process loadchanges). In some applications it is the response to loadupsets that is critical.)

Since the introduction of M PC in the 1980s until 2004,over 5000 plant sites have utilized this technology. ModernDCS software packages include embedded M PC controlfor small applications (models of eight inputs by eight out-puts in size) as well as larger applications, which traditionallywere handled externally to the DCS (layered MPC). Fastexecution rates and ease of use in commissioning allow thenew embedded M PC control to be used where previouslyonly traditional control techniques based on PID control hadbeen applied. Improvements in the user interface and in thetools for commissioning and operating the MPC allow someexperienced process or control engineers to apply this tech-nology without the assistance of outside experts.

FIG. 4.7dApplication of preengineered fuzzy control.

© 2006 by Béla Lipták

controller is based on the controlled process (see Sections2.13, 2.14, 2.17, and 2.19 for details). (Editor’s note: Accord-

ship functions with rule inference (see Section 2.31 for defini-

refer to Section 2.31.

682 Control Room Equipment

In some simple M PC applications, the controller con-figuration might consist of only a few parameters. Examplesof such parameters include:

• Controlled variable — Set-point limits• Manipulated variable — Maximum rate of change,

controller output limits• Constraint parameter — Constraint high and low limits

The MPC controller is based on a dynamic model of theprocess. Traditionally, the dynamic model of the process isdetermined by manual manipulation of the control valveopenings (process inputs). When the model is small enoughfor embedded M PC control, automated testing of the processis provided as part of the MPC engineering tools.

Once data on the dynamics of the process is collected, theprocess model and controller the MPC, can be generated auto-matically. Both the step response of the MPC model and ofthe actual process can be displayed. An example of the over-view step response model and a detail of an individual outputresponse for a particular input are shown in Figure 4.7e.

MPC identification tools compare the measured processvariables with those calculated based on the model, if themanipulated variables to both the model and the actual pro-cess are the same. This comparison may be displayed indifferent formats to allow the user to determine how goodthe model is, i.e., how closely the model dynamics matchthose of the actual process.

A simulated environment is used to evaluate the controlresponse to both set-point and load changes before the MPCcontrol is commissioned. Using the model of the process forsimulation, it is possible to see the actual process responsein advance of real time. Similarly, on very fast processes, theability to simulate the control and process response slowerthan real time can also be valuable. The ability to adjust thespeed of execution allows the control response to be quicklyverified for a variety of operating conditions. Figure 4.7fillustrates the testing of the M PC model by viewing thesimulated controlled variable response of the M PC model toa change in the manipulated variable.

Once the M PC control has been tested offline using sim-ulation, then the control can be commissioned.

(Editor’s note: If the process model accurately reflectsthe process, the control response to set-point changes is likelyto be good. The response to load disturbances is usuallypoorer if the load disturbances result in substantial changesin process dynamics; in such cases, adaptive control shouldbe considered as a possibly better option.)

NEURAL NETWORK APPLICATIONS

Artificial neural networks (ANNs) can learn complex func-tional relations by generalizing from a limited amount oftraining data. Hence they can thus serve as black-box models

FIG. 4.7eThe step response of an accurate MPC model and of the actual process can be very similar if the process load is the same as it was at thetime when the model was developed.

© 2006 by Béla Lipták

4.7 DCS: Control and Simulation Advances 683

of nonlinear, multivariable static and dynamic systems andcan be trained by the input–output data of these systems.Applications that are suitable for ANN solutions and the

ANN is similar to fuzzy logic in that the mathematicalmodel that relates the inputs and the outputs of the modeldoes not need to be known. To use ANN, it is sufficient toknow the behavior or response of the model. The main dif-ference between fuzzy logic for property estimation and ANNis that the gains and functions cannot be modified in ANN;they can only be “trained” with data.

The use of estimated process variables can improve con-trol and monitoring when critical measurements are onlyavailable through laboratory analysis. The intermittent labo-ratory analysis results can be used as inputs to provide onlineinferred or estimated variable values when actual analyticalreadings are not available. Neural networks are often usedfor process variable estimation even if the inferring relation-ships are nonlinear. ANN provides a nonlinear model of aprocess, which uses those input variables that indirectly influ-ence the composition that is intermittently detected by labo-ratory analysis.

After a neural network has been commissioned, the offlinelaboratory analysis is still required to correct for unmeasureddisturbances and changes in the process. Without this correc-tion, the prediction provided by the model of the neural net-work will drift away from the true measurement value as timepasses. For example, the laboratory analysis may be used as

an input to the ANN model to obtain automatic adaptation ofits prediction in response to changes in process.

Modern tools embedded in DCS control systems havereduced the complexity of designing and training neuralnetworks. ANN can be integrated as a function block intoa distributed control system and that block can be config-ured similary to other function blocks. The tools providedto train the neural network take into account the dynamicnature of the processes for which a neural network isdeveloped.

Once a potential application for a neural network is iden-tified, the first step in the ANN development is to configurethe function block. For example, as a first step in configuringthe ANN block, the user may browse the past history of theprocess to identify the variables that tend to influence com-position and use them as function block parameters. As actuallaboratory analysis takes place, the results are provided tothe ANN model for updating the model of that analyticalvariable that is being estimated by ANN (Figure 4.7g).

Once the neural network block is configured and down-loaded, all measurements used in the ANN block are assignedto a historian for collection. After a sufficient amount of processdata have been collected, the operator can graphically view thehistoric data and make corrections by adding/removing inputs.Once the input data have been screened, the ANN model canbe generated. During this generation process, the dynamics ofthe input delays and output sensitivity to every input are calcu-lated. To help the user determine the accuracy of the neural

FIG. 4.7fThe MPC model of the process is tested by introducing a change in the manipulated variable and viewing the resulting response of thecontrolled variable of the simulated process.

© 2006 by Béla Lipták

detailed theory of ANN are discussed in Section 2.18.

684 Control Room Equipment

network, the value calculated by the neural network may becompared to the lab analysis, as illustrated in Figure 4.7h.

PROCESS AND CONTROL SIMULATION

Process simulation packages are used to train operators priorto plant startup; they are also used to examine the dynamicresponse and performance of a process during operation, andfinally, they are used to prepare the operators for complexand critical shutdown procedures.

The cost of developing the simulation software rises expo-nentially with the required accuracy of the process model. Yet,it is desirable that the model reflect not only the processdynamics but also the operation of the actually implemented

control system and that it incorporate the simulation of pro-cess upsets to train the operators in a real-time setting. It isalso important that the keyboard, strokes, sequences, mouse,and touch screen used during simulation be identical to theactual ones.

Some commercial systems allow all features of the pro-cess control system to be executed on a single PC platformor distributed between multiple computers. This high-fidelityprocess control simulation capability may be used to checkout control system logic and operator interface as well as totrain operators on the continuous and discrete control, diag-nostic, and alarming under a variety of conditions.

The software and the associated configuration of a con-trol system may be designed to be used in multiple operatingenvironments without change. When this approach is taken,

FIG. 4.7gNeural networks can be used to estimate analytical laboratory measurements on the basis of the past history of such readings by sendingthe associated laboratory readings to the ANN block. Later composition measurements from the laboratory are used to update the ANN model.

References amaximum of 30process measurementsfor analysis

FIG. 4.7hComparison between the actual laboratory readings and the ANN-predicted values of composition.

© 2006 by Béla Lipták

4.7 DCS: Control and Simulation Advances 685

it is possible to distribute control system functionality in avariety of ways without needing to reengineer the softwareor reconfigure the associated applications.

All features of the control system can be combined on asingle platform or distributed between multiple PCs to sup-port configuration and dynamic checkout of the controls andinterface associated with a process control system, as illus-trated in Figure 4.7i.

When a control system utilizes the Fieldbus Foundationfunction block architecture, a SIMULATE parameter is avail-able in each I/O function block. Through this parameter, avalue and status may be provided to I/O function blocks tosimulate the field measurement or actuator in the controlsystem.

This capability may be used to check out control logicand operator displays. Also, where the control systems support

FIG. 4.7iOperator training can utilize single or multiple PCs loaded with the same software package representing the process and its control system.(Courtesy Emerson Process Management.)

FIG. 4.7jDisplay interface for simulation coordination.

© 2006 by Béla Lipták

686 Control Room Equipment

the object linking and embedding (OLE) interface for pro-cessing control (OPC), the process simulation package maywrite the simulated value and status. This capability, com-bined with the ability to redistribute control functionalitywithout reconfiguration, provides a strong foundation for theintegration of high-fidelity simulation with control systemsimulation in a single or multi-PC environment.

Depending on the modeling rigors and computingresources, the process simulation can potentially run fasteror slower than real time. To allow the control simulation tomatch the process simulation execution, the control systemexecutes function blocks faster or slower. In addition, atrainer or an application may coordinate the execution ofcontrol and process simulation. An example interface for suchcoordination is shown in Figure 4.7j.

CONCLUSION

Modern distributed control systems can provide embeddedadvanced control solutions. The DCS system supports theadvanced control capability similarly to the support providedfor the traditional control tools. Among the major advancesin DCS system designs are the easy-to-use engineering andcommissioning tools.

Reference

1. Ruel, M., Instrument Engineers’ Handbook, Vol. 3, Section 5.6, Jointlypublished by CRC Press and ISA, 2002.

Bibliography

Astrom, K. J., and Hagglund, T., PID Controllers: Theory, Design, andTuning, Research Triangle Park, NC: ISA, 1995.

Blevins, T., McMillan, G., and Wheatley, R., “The Benefits of CombiningHigh Fidelity Process and Control System Simulation,” in Proceedingsof ISA Expr/2000, Chicago, October 2000.

Blevins, T., and Ochoa, D., “Statistical Analysis Rides the Bus,” Intech, pp.48–59, March 2000.

Bonavita, N., “Benchmarking of Advanced Technologies for Process Con-trol: an Industrial Perspective,” Proceedings of the American ControlConference, Vol. 6, pp. 4328–4329, 2001.

Bonavita, N., Martini, R., and Grosso, T., “A step-by-step approach toadvanced process control,” Hydrocarbon Processing, Vol. 82(10), pp.69–74, 2003.

Canney, W. M., “Special Report: The Future of Advanced Process ControlPromises More Benefits and Sustained Value,” Oil and Gas Journal,Vol. 101(16), pp. 48–55, 2003.

Hokanson, D., Mun, H.C., and Rebeck, J., “Integrating Advanced ProcessControl on a Grassroots Project,” Hydrocarbon Processing, Vol. 78(6),p. 63, 1999.

Kuchle, D., “Multivariable Control on a Single-Loop Budget,” Control Solu-tions, December 2001.

Narendra, K. S., and Balakrishnan, J., “Adaptive Control Using MultipleModels,” IEEE Transactions on Automatic Control, Vol. 42, No. 2,pp.177–187, February 1997.

Macias-Hernandez, J. J., and Feliu, J. A., “Dynamic Simulation and OPC,”Hydrocarbon Engineering, Vol. 8(9), pp. 53–58, 2003.

Mehta, A., and Tzovla, V., “Automating the Approach to Development andOnline Operation of Soft Sensors,” Intech, pp. 30–33, September 2002.

Qin, S. J., “Auto-Tuned Fuzzy Logic Control,” American Control Confer-ence, Baltimore, 1994.

Wang, H., “Introduction: Advanced Process Modeling and Control in Paper-Making,” Measurement and Control, Vol. 32(8), p. 228, 1999.

Xionglin, L., “Problems and Countermeasures for Engineering Applicationof Advanced Process Control,” Petroleum Refinery Engineering, Vol.101(16), pp. 48–55, 2003.

© 2006 by Béla Lipták