You are on page 1of 9

ISA Transactions 50 (2011) 2–10

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Editorial — Choosing advanced control✩

1. Introduction Attributes of processes within the chemical process industries
(CPI) pose particular challenges and a unique relative importance
1.1. Bottom line to issues that are different from those encountered in other control
applications such as robotics, aerospace, and communications.
Use an appropriate tool for the job. Start with process Furthermore, differences in economic and risk considerations and
knowledge. This is essential to making right decisions. in standardization of process and equipment design, between
the CPI and other control applications, restrict the permissible
1.2. Scope investment of sensing and control modeling. As a result, only
a subset of control technologies has found acceptance within
ARC (Advanced Regulatory Control) seems to have a common the CPI. There are many other advanced control approaches
meaning. It refers to what used to be called advanced process used outside of the CPI. These include sliding mode control (an
control: gain scheduling, ratio, cascade, feedforward, decouplers, adaptive approach), fractional order (a nonlinear approach), and
override, and related and ancillary techniques such as anti-windup, state–space (or modern) control and linear quadratic regulator
bumpless transfer, PID modifications, and tuning techniques. (multivariable linear model-based predictive approaches). The
The ARC techniques were known prior to computers and the focus of this editorial is on technology appropriate to the CPI.
modern era of state–space and model-predictive control. Most of
these techniques were known during the time of analog control. 2. Challenges in process control
However, the problems of individual component cost, component
reliability and consistent performance precluded the use of most Understanding the issues that face CPI applications is essential
of them, except for occasional cascade loops. Perhaps the heyday to choosing a right control strategy. Here is a listing of CPI
of these ARC techniques was after the introduction of computer- application attributes that cause difficulty for control.
based distributed control systems (DCS) and prior to the wide- • Nonlinear Process—The process gain changes with manipulated
spread use of APC. variable (MV) (controller output, process input), to process
By contrast, APC (Advanced Process Control) has many mean- variable (PV) (process response), and to controlled variable (CV)
ings. Within the model predictive control (MPC) community, APC (the process response that is controlled, controller input). If the
means MPC. Since the ‘‘big ticket’’ APC item within the chemi- sensor or the final element introduces the nonlinearity, often
cal process industry (CPI) is MPC, APC and MPC are synonymous signal characterization or control element characterization (an
to many of us. However, many control experts recognize that the inverse of the nonlinearity) can linearize the control loop.
modern computer era also brings us other advanced nonlinear • Non-Stationary Process—Process attributes (such as time-
and adaptive controllers, automation of supervisory real-time eco- constants, gain, interactions, and dead time) change in time
nomic optimization of controller set points, computer perception due to product grade, piping arrangement, unit switching for
and monitoring of status and health to trigger corrective action, maintenance or production, operational stages within units,
control of inferential variables, and computer-based planning and and fouling or other degradation of equipment or sensors.
scheduling. • Ill-Behaved Dynamics—This includes large dead time relative
Since this editorial is primarily about a comparison of feedback to other time-constants or sampling interval, integrating, open
control strategies, it includes nonlinear and adaptive algorithms loop-unstable, inverse acting, or disparate settling times for
along with MPC in the definition of APC. different variables.
With this focus on feedback control, this article does not • Multi-Variable—Each of several MVs affects each of several
explicitly address essential and control-relevant methods for CVs.
inferential process sensing, supervisory process health analysis, – Interaction—Requiring a coordinated plan for the several MV
fault detection and abnormal situation support, device function moves to temper interactions.
and reliability. But the importance of such techniques is frequently – Degree of Freedom—There may be more MVs than CVs, an
acknowledged in the paper. extra degree of freedom (DoF) situation that requires an
economic optimization of the best MV combination that
meets the CV objectives. Alternately, there may be fewer
✩ This editorial is based on a presentation ‘‘Advanced Classical or Model- MVs than CVs, a negative DoF that requires optimization
Predictive Control?’’ at the ISA 2010 Automation Week Conference, Oct 4–7, to best limit deviations from set points (SP). The DoF
Houston, TX, USA. situation changes in time as production, product mix, raw
0019-0578/$ – see front matter © 2010 ISA. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.isatra.2010.10.004

or due to process turbulence impacting flow rate • Constraints—This includes product specifications. cation unique. .). where increasing constitutes the single loop. every robot. reduce DoF. what appears to not permissible). due to the eventual expression of control action taken now. process and control method complexity. ad. . environmental up- for the DoF > 0 optimization change with economic sets (weather and raw material variability). choking. . rather than updating optimizer perturbations to what might be an average signal. measurement. catalyst degradation. the constraint stream controller cycling. models can be either short lived. up-stream process factors affecting the business (waste penalty. alarm/priority overload. car. PV is safety each overrides concern about measurement accuracy. the controller is not constrained. and flexibility internals. and environmental factors • Disturbances—These can take many forms: Accidental human (disturbances) shift active constraints. vacuum). Constraints may be encountered now or in the future. surprising the control system with a CV devi- political factors (What is management’s concern today? How ation to fix. limiting process information available to the control etc. series has the same behavior (usually). . In this case. etc. . require (volume flow rate instead of mass flow rate) because it is maintenance. CV or any be independent noise begins to express autocorrelation and PV. bias or systematic error. Orifice flow meters. If their persistence is less than DoF < 0 optimization could change with the same factors. they could also reflect scheduling issues. maintenance procedures. Such ‘‘rational’’ By contrast. aging. training and primary value (pressure drop can infer viscosity which infers education required to implement and maintain the controllers polymer average molecular weight). calibration drift. and gets the flow rate set control in fired heaters/boilers where the fuel follows air when point from the primary temperature controller. water treatment. (any form of upset). temperature. rather than to goodness of control. ion exchange units. operational measurement. fired boil. and confounding the measurement with inferential plant has a unique design with unique behavior. • Cause and Effect Relations—Depending on one’s viewpoint. Such sources of noise are due to the process- limits on equipment (cavitation. ownership history. or because it is easy to ‘‘prove’’ to regulatory adaptive. the • Models—Since processes are nonlinear. etc. There are not enough sensors • Plant Staff Experience (of the Operator and Process Engineer)— to measure everything. If constraints strategy. Active constraints product temperature. • Individuality—No two CPI processes are the same (usually). vibration. . . and replacements cause – Measurement Error—Due to sensor failure. designer usually employs the minimum number of sensors capacity. pump capacity. instrument air failure. fumes. raw material changes. camera. Constraints may be on the MV. providing instruments and final control elements prone unit-specific control system. safety. the process pressure in down-stream utility lines. the sampling interval they appear as noise. essential for control and analysis. equipment failures. If the sampling frequency is increased. and they often measure a related but not then easy and understandable process overview. such as: making CVs have a nonlinear and interacting response ers. The economic weights perturbations or errors. Admittedly. In this case a model that relates current For example. inventory storage (tank level). There is a fine line between disturbance and permissible) or hard (violation is either impossible or absolutely noise. Alternately. or are not measured. on a primitive heat exchanger control. explosive limits). cavitation. refrigeration units. etc. Constraints may require one variable to always have a However. worse. flange taps transmission. model development is expensive. physical exertion). And. to MVs. eliminating that disturbance from the inactive. sensors are often located in a spot Several human issues include: Human machine Interface (HMI) that leads to delay or lag. – Final Control Element Problems—Valve sticktion. Since processes composition control. product price. limits and targets are normally changed. multivariable. fouling. Constraints may be contradictory. have a 3%–7% measurement uncertainty. noise may be the result of position. . occupational health transmitted signals. . This makes each control appli. uncertainty. the signal MV action to the future values of the process variable is from the product temperature controller to the steam valve required. In the cascade increasing and air follows fuel when decreasing). . broken mixer impeller. • Initial Capital Cost—The desire to minimize investment cost of- ten drives the design and equipment decisions. are often used because desire for maintenance convenience and – Calibration—Discrimination error dominates a signal. in either a cascade or ratio arrangement. nearly every CPI to faults. processes to evolve individually. some processes have standardized designs (air separations. However in practice. Disturbances can violation and SP violation concern values needed for the have short or long persistence. valve sensor combination.). Editorial / ISA Transactions 50 (2011) 2–10 3 material. The constraints Gaussian distributed that it can be accepted as normal for may be either soft (some violation for a short period is analysis methods. steam • Sensors—Seeking to minimize capital investment. Often sensors measure the wrong thing are non-stationary. and numerical value of flow rate (reflux perhaps) is irrelevant to unique. mechanical vibration or stray electromagnetic corruption of safety issues (pressure. for – Control System—High traffic leads to missed or delayed example. nally built from a standard design. bodies by convenient calibration. . process maintenance. a or due to the effect of measured or unmeasured disturbances substantial shift can be made in the input–output perspective. appears like a disturbance with some persistence. CPU overload. But by contrast. upgrades. beyond calibrated range. delays. If constraints are ‘‘immediately’’ corrected. aging equipment. variables. the steam proportion that is larger than another (such as cross-limiting flow controller operates the valve. or airplane of the same economic decisions often result in control-related undesirables. energy cost. or must be periodically updated or made convenient. any flow rate fluctuation due to steam pressure is are ‘‘active’’ they are violated or about to be. up- cost. . and standardized process units would be shipped with a system. models). valve sticktion. surging. Most disturbances Additionally. . As long as the flow meter is internally consistent. and associated aspects (HMI. and noise. • Faults—These take many forms: ditions. Often noise is close enough to being issues (noise. or parts of be due to mixing fluctuations causing variation on composition the application are turned off. . Even if origi. collapsed and located with a priority given to cost. blockage. steam pressure is a the MV relieves one constraint violation but makes another significant unmeasurable disturbance to product temperature. The sensors are often selected – Process—Bypass.). and providing little information about the impend- soon is the labor contract to be re-negotiated? How long ing magnitude or trend in the CV deviation. . Noise could weights. low-cost sensors often have high understandability. has it been since a community or regulatory complaint? • Noise—This comprises seemingly random and independent etc.

This of the problems listed above. It includes anti-windup MV. compensation. Where the model dead time is a rea. I. The on–off switch could be also provide simple structures. GMC reduces to a PI controller sending a signal controller tuning values to meet dynamic performance that represents a biased set point to an output function choices as the process changes. basic modeling) process model can also may be adjustable). and any of these can initiated by a CV crossing a threshold and tempered by a be implemented by an engineer with MS-level skills. indication that a change will affect the primary CV. in how the wild and secondary controlled variables affect nal learning ability that adapts to the dynamic process data. or a gradual return to the prior • Infrastructure—The choice of control scheme and devices must set point in the AUTO mode for bumpless transfer. controller to determine the set point for the secondary CV. etc. These are organized from There are several control approaches that use process- simple to complex. set point softening. This is simple (GMC) [4] and Process-Model Based Control (PMBC) [5] to understand and implement. and are no MV to control the secondary CV.3] single-output (SISO). use the jective. PFC uses a coincidence point be apportioned as a percent of a fixed cycle time. velocity mode. Internal Model Control (IMC). the users must assess knowledge can be used to determine controller gain. economic. There is only one tuning parameter. the process. etc. The essential minimal control following all reduce to PID or one of the subsets (P. other sampled-data model-based controllers. software. and use the primary CV different from a PID. safety. GMC with steady-state models uses a within each cycle time. generally single-input. as in time. fraction beyond the set point as a target to calculate (b) PID—This includes P-only. (i) Gain Scheduled PID—Either from explicit mathemati- cal models or from experience it is relatively simple to 3. The (a) On–Off—Sometimes referred to as bang–bang control. procedures.4 Editorial / ISA Transactions 50 (2011) 2–10 • Return on Investment—Control system and algorithms need (f) SISO Nonlinear Controllers—With a linear model. Feedback correction could be used to bias the set and bumpless transfer features.). site. there is a higher level the process match the reference trajectory) for the controller.) In ideal cases. relief valve. PMBC and GMC with dynamic models use a This is the classical controller that was substantially com. and many variations (rate point. be compatible with the legacy of field instrumentation. room heater. Accordingly. I-only. bumpless transfer. parallel gains. And. An engineer that is able to generate a first-principles derivative (PID) tuning parameters (switching dead band (simplified. and then uses can be viewed as feedforward. Control schemes understand how the process gain. refrigeration. (a time in the future when it is desired to have proportioning control. the on–off periods may write the controller code. is easily implemented a digital controller. each needs to initialize model economic. Generic Model Control compressor. air controller software. that uses the nonlinear model to calculate the MV GMC (d) SISO Linear Model-Based Controllers—Targeted to com. auto-tune variation. The controller tuning response to choose a control scheme that is right for that confluence of issues the nonlinear or non-stationary process can be placed and concerns. the CV then include dynamic compensation. model-based. Listed below are a wide range of control schemes that have been (ii) Relatively Simple Process-Model Based Controllers— accepted in the CPI by industrial practice. However. In this case. which problems are most important within their process context integral and derivative time by conventional reaction- (this includes technical. and in the broad categories of Basic Control. for example). then one PV times the ratio becomes the set point control the model rather than having to wait to see what for the second PV. controller. As with SISO PID. models which are relatively simple for an individual ARC and APC. control each needs to limit the output between 0 and 100% (or −6% systems. When a wild flow changes. simplifying on-line tuning. and curve tuning rules. These also must have MAN–AUTO bumpless transfer variables. is obvious to the process owner. safety. dead band. the ratio the process does. Relative advantages are described for each. this has the additional benefit of providing (c) Self-Tuning PID—These products use a range of strategies information about process health. keeping the model locally true to band. etc. evolutionary optimization. (IMC forms (a) Cascade—One controller sends a set point to a lower-level the basis for lambda-tuning of PID control. point. transport delay.) to periodically adjust the PID model. desired rate of change toward the set point to calculate pleted and analyzed by the 1940s. but without dynamic the models for control action. Ratio control detects a change and takes and output limit features as do PID. Ratio control tracks changes in process dynamics and gain. and training materials for the and 106%. In this case. Alternately. etc. are of the few that are also marketed as a product. such as PID. in a look-up table or calculated by equations. Ratio could be based on energy or other composite generating the model is a step more complicated for the user. the MV. and political issues). PI. How does one justify the PI. and Active Disturbance Rejection Control (ADRC) [2.). and It seems that no control scheme (strategy) solves every one time-constant change with operating conditions. Predictive Functional Control (PFC) [1] (1) Basic Control—First level controllers. the to be economically justified. these algorithms generate PI or PID rules. waste or risk reduction benefit and pay-out values and have other auxiliary operations such as set point time for each additional level of control complexity? tracking in the MAN mode. In PMBC feedback correction is used to adjust before reset. there are no proportional-integral. Alternatively. to implement. that is setting the percent on-time MV calculation. . proportional model parameters. pensate for dead time is the Smith predictor. usu- times the wild flow determines the controlled flow rate set ally for CV damping. and user must define the process model and embed it in the representing the signal to a solenoid valve. and (MISO). Using a steady-state (expert systems. When an intermediate PV can provide early such as a low-order process model and a simple control ob. and for (2) Advanced Regulatory Control multi-input–single-output more complex dynamics. sonable approximation to the process dead time these con- (b) Ratio—One controller determines a desired ratio between trollers provide excellent results because they essentially PVs. immediate action. If there is a large dynamic difference (e) Adaptive Controllers—These usually have models or inter. and all combinations.

Vendors use known and relatively stationary. However. there are other technologies that are sometimes nonlinear PI control is desired. and MPC–NN MV sequence that will avoid or minimize constraint (a neural network based model within an MPC structure) control violation. feedforward can handle that. from up-stream events. Heuristics or use linear (stationary) dynamic representations and expert rules are embedded in a logic system that uses either nonlinear static gains or multiple linear models. Tempering control MV so that the change in the other MV does not upset the action is by either CV damping (CV reference trajec- one CV. In any case it is especially important that DoF < 0 calls for balancing CV objectives. nonlinear. and finally. then classic gain scheduling considered as within the realm of APC. tuning or gain-scheduling in the PID. must calculate an optimal MV sequence at each stage. a vector of CV values after an MV (e) Decouplers—In a multivariable interactive process with impulse. classic override by either high or MPC typically uses a finite impulse response (FIR) low select blocks is inadequate. carbon bed absorption capacity degrades. interactive. Most commercial products set of controllers uses human knowledge. . and constraints. with a vector length (time duration) equal SISO control loops. when it changes. and ARC algorithms on a commercial-scale the impact of past MV and disturbance influences plasma etch reactor and a lab-scale distillation column. interactive. change. disturbances. Joshi on the future of the process. IMC. tory) or MV damping (a penalty for large MV moves). Although the test processes expressed many control problems if the process model is non-stationary the controller (ill-behaved dynamics. external reset feedback. Dynamic Programming can calculate a MPC on lab-scale distillation. niques. in situations where specification. the nonlinearity. that MV action upsets to the time required for the CV to return to within a another CV. (3) Advanced Process Control—generally multi-input–multi- output (MIMO). able aspects. [7] evaluated ARC. and use between several linear models as operating conditions the information to adjust the controller. PMBC. Nonlinear MPC can accommodate for trol. This can adjust for nonlinearity. and grouped NNs to individually predict select (i) Fuzzy Logic (FL) or Expert System (ES) Controllers—This future process values. They are not ‘‘control’’ per se. proportional band. the boss is watching. Subawalla et al. measurement noise. the least squares algorithms for on a pilot-scale fluid flow and heat exchange inverse is relatively easy to obtain. and acts as a measurable disturbance to the noisy vicinity of its original value. 4. There are a variety of approaches. expert estimates. However. (h) Adaptive and Self-Tuning Controllers—These algorithms first-principle models. Collectively the processes express best path. economic conditions. Signal transfer. Either case imposes an inconsequential computational varying dynamics. PMBC. Editorial / ISA Transactions 50 (2011) 2–10 5 (c) Feedforward—A disturbance is observed. FLC. human understanding to calculate the MV. useful range. safety in- control of parallel streams. disparate dynamics. Model Pre- the non-selected controller not windup. in situations could make an auxiliary PV violate a limit. may be impractical in situations of (d) Override—Control intended to keep one CV at the SP very high number of interacting MVs. MPC packages may include some are all of the following features: 4. Override could be the result of safety. and choose a future et al. noise. (f) Signal characterization—If a nonlinearity is fairly well. or automation of however. It appears that FLC and ES are better expert systems. there are even characterization will often eliminate the need for self. they all were of low dimension (2–4 MVs) . some companies are offering true nonlinear compensation for measurable disturbances or coordination control approaches. ). flicting constraints. and substantial dead time. model of the process. solve that problem. and have developed techniques to handle multivari. Feedback of process- other controller. These include would be simpler. burden for on-line. IMC. such as coordinated mal situation management. however. reset. When there are dictive Control handles these in a unified framework. which can be placed in a look-up table. including the use of neural network static models. DoF > 0 calls for economic optimization or where maintenance. etc. real-time calculations. . classic problems associated with MIMO. more options within the concept of MPC. in situations where there are con- CV controller. and often includes issues such as dynamic However. (ii) Linear MIMO Control—If interactive. Ou and Rhinehart [8] evaluated grouped neural network but static. [6] evaluated MPC (a commercial product Model-based Constraint-handling controllers forecast [DMC]⃝ r ).1. artificial neural nets. These tech- composite of several measurements. optimization. nonlinear that linear control only has a limited nies specialize in implementing process-model based con. bumpless thus linearizing the control loop as a whole. Optimization handles the DoF ̸= 0 case. Compared with the number of function (the inverse of the process nonlinearity) can options available with PID (velocity mode. Comparisons of ARC and APC (a) Model Predictive Control—This is the most often consid- ered component of APC. model evolution. one controller changes its MV to correct for a disturbance to its CV. Most vendor products show an (b) Other—Depending upon the corporate organization and example FLC as a nonlinear velocity-mode PI. If nonlinear unit. decouplers can a time-scheduled compensator (typically lead–lag-delay. But if culture. switching observe the MV and CV and possibly disturbances. Dynamic compensators can decouple the model-mismatch (residual) corrects all future model interaction by adding a time-compensated change on one predictions by the current residual. Override can might be either a single measurement or a calculated handle constraints. alarm management. set point softening. etc. conflicting constraints. when they are hit. nonlinear). If the model is linear. proaches used outside of the CPI. The disturbance bances. In this case where avoidance of future constraints is an impor- control of the MV needs to be taken over by the auxiliary tant consideration. If subject to measurable distur- gain) is added or multiplied to an MV. Experimental (i) SISO or MISO constraint-handling control—Action now could lead to violating a constraint in the future. (iii) Nonlinear MIMO Control—Some processes are so (g) Tailored Process-Model Based Controllers—Several compa. then a characterizing a variety of optimizers. MPC. GMC. abnor- suited for higher level applications. rate before be employed to overcome the process nonlinearity. . to change cycling periods as strumented systems. There are also many other advanced control ap- other supervisory activities of the process operator.

There are many other benefits to consider. covered all major sectors of the CPI. rate). dynamically coordinated. • Down time reduction (18%). • Where a single problem dominates. of the economic benefit. with relatively simple constraint and degree of freedom (DoF) Controller models based on process first-principles can be aspects. constraint avoiding manner. it is technically equivalent to any other controller. Respondents cited a range of 35%–85% for CV variability reduction gineering aspects such as trouble shooting. They ments. 38 users. dead time compensation. process improvement. • Computational aspects (computer speed. set point tracking and disturbance rejection. and which are listed in the following. use the simplest control • Better use of raw materials (15%). knowledge development). memory require. process diagnosis and health aspect as important. approach designed to handle that problem. development. supervisory economic modeling. ancillary routines. revealing MVs. is a critical element 5. supervision/intervention. 4. politics). Evaluation • Safety increase (9%). nothing that improves their understanding. especially operator understanding and ability to respond to abnormal situations. Estimating economic benefits of advanced control in the success of process management. Control system can disentangle interactions and move multiple MVs in a Operators are often not able to make sense of what the controllers are doing. The human machine interface (HMI). ARC matches nonlinear-APC. • Increased yield of more valuable products (50%). Enhance operator ability Diminish operator ability A good dynamic model can provide a strong training simulator that enhances Advanced control can be confusing and reduces process understanding. that ‘‘throughput Increase’’ was an economic benefit associated with an APC project. Their evaluation criteria included: 5.2. which displays the process and controller activity to the operators. • Reprocessing cost reduction (10%). FLC and Expert systems. In any one How does advanced control impact the operators’ ability to particular application. They reported the primary reasons for economic benefits. and related to control implementation and operation. understandability. the authors’ experience. from improved control. predictive maintenance.6 Editorial / ISA Transactions 50 (2011) 2–10 Table 1 Advanced control strategy impact on humans. by revealing the right way to coordinate Control displays may not keep the operators engaged with the process. stays tuned. averaging about 50%. • ARC was best in all categories except in miscellaneous benefits • Quality giveaway reduction (40%). asking about how to determine the economic benefits of APC. and others. control. the first 5 benefits dominate. constraint control. same features. For instance. Automation can correct human error. IMC. benefit. 28 suppliers). 70% of the respondents stated monitoring. hard and soft). the payback period. MPC. APC includes rors. In their study. If operators are not engaged and the control action is not understandable. industrial APC experts (66 responded. fault/situation diagnosis. Each group evaluated multiple performance criteria especially useful for knowledge validation and dissemination. self-tuning and adaptive • Robustness (unexpected upsets. • Process stability improvement (55%). MV movement. not the magnitude of the benefit): • When any controller contains solutions to all process problems. • Energy consumption reduction (55%). personnel training. Economic benefits of control and advanced control • Cost (initial. ease of Bauer and Craig [9] reported the results of a survey of adjustment. process testing to obtain models). execution er. knowledge dissemina. alone. linear programming. production they had in-house APC software and expertise to implement it. • Balance of CV and MV performance (ISE. More than 50% of the 38 users indicated that calibration errors. take manual corrective action in response to process events? This but overall. for relative new personnel. then their ability to The survey [9] also asked APC experts to state the extent manage abnormal events in the process progressively diminishes. Obviously there are many economic benefits. and DoF (constraint) handling. product changeover. guaranteed solutions within a defined time). Listed in decreasing priority (the percentages Their conclusions include: represent the number of respondents citing this as an economic • None of the control approaches are operator-convenient. Tests included automated process health monitoring. engagement. future nonlinear control. and the popularity of the benefits tion. indicates that owners chose to operate closer to specifications . statistical process constraints. APC or ARC decisions should not be based on CV performance • Other (10%). process and instrument faults. equipment. • Operating manpower reduction (7%). diminish or enhance that ability (see Table 1). control. and process management. split range control. from the above benefits. • Miscellaneous benefits (process knowledge validation and are approximations of the proportion of respondents listing that dissemination. • DoF handling (future impacts) requires model-predictive con. and complicated controllers can either throughput. • DoF handling (excess and insufficient number of MVs. • Responsiveness increase (12%). convenience. The numbers in parenthesis propagation of noise). trol. one benefit may have been the critical issue. 5. Primary benefits • Operator issues (education. and how can How do the modeling outcomes of advanced control benefit one estimate the potential impact from an APC application? process understanding by the engineer and subsequent process en. maintenance.2. control. This provides the process owner an opportunity to choose tive maintenance. which matches and abnormal management? The better the process knowledge. With the • Throughput increase (70%).1. it seems is grounded in operators’ understanding of both the advanced that most applications used the improved control to increase controller and the process. algorithm robustness. And of them. A reduction in variability means tighter the better will be process management decisions related to predic.

understanding how to use it. inverse. wanted to end the project. an APC application is the expected time over which the economic a paint can opener is the best tool. But. linear.1. $450k. may not be functioning properly.3. several basic control loops understood for good manual control—Use FLC. override. ‘‘When you have a hammer. Hugo [13]. However. re-tuning or CPI are about 95% with the remaining 5% of loops using something restructuring of lower level controllers changes relative gains that else (MPC. For instance. yielding improved plant controllability • If SISO. Some of the Bauer and • Production loss due to installation downtime (4%). implementing APC are: feedforward. . Being familiar with the shape distribution or demand. such problems may significantly improve process performance. Editorial / ISA Transactions 50 (2011) 2–10 7 and process constraints. or to the can lid. Honeywell [11] reports 6 months on Again. Application demographics favor the upper-level MPC experiences. GMC. Canney [10] estimates that • Traveling expenditure (5%). changing valve sizes changes implementing PID. etc. cascade. Refreshingly. and non-stationary—Consider adaptive to APC is the mechanism to get the resources needed to take care controllers. significant issues. Craig [9] respondents cited 5%–10%. half of the APC experts choose a control added? Benefits are commonly claimed for MPC pushing the plant to technology that they are already familiar with. which means increased throughput and • Internal manpower cost (35%). cascade. feed- forward. nonlinear. ‘‘Were any measurements and Craig [9] report that ‘‘. but I continue to use the screwdriver. the right tool. This action substantially improved • If SISO.000. This makes APC Canney [12] estimates an average MPC implementation cost of very high on the investment priority list. a change in product to open a can of touch-up paint. so much so. if the tool works once. and the strategy may no longer has many enhancements for gain scheduling. and reduced energy consumption. disappointingly rapidly. and Hugo [13] each mention process to another. primary control devices and strategy. Ford [14]. use PFC. be matched to how the plant is now being operated. only qualitatively understood. well understood. scheduled PI(D). This requires and stability. It translates an operator’s experience in one Kern [15]. However. that ‘‘. [16]. or ARC. Alternately. Unless the economic incentive choices technically more favorable. Wade recalls an estimate of APC project half-life of about one year. reasonably functional. . of the basics. Estimates of dysfunctional for the new process. the knife may become the chosen tool This would suggest that a criterion to consider when justifying to open the paint can. tain meaningful study results. Technical aspects authors’ experience. or is it due experience with that technology on similar processes’’. A garage-person may use a screw driver. Estimation of costs of implementing APC • If MISO. a model of the ill-behaved dynamics (integrating.). it appears that many applications degrade into disuse. Here is a guide for controller selection: tem that is in need of renovation is often a major contributor to improvements. in a process study made for most situations where interactions are and constraints are not the purpose of designing ARC. adaptive. Its high reliability makes it appropriate for this. • Control software (upgrade) (45%). etc. or Estimates for the number of MPC product units implemented product mix) change and the models or relationships become world-wide since the 80s range from 10. Rhinehart says. and DCSs. PLCs. remains. MPC successes are often credited to fixing and upgrading the delay).13. but certain control problems make other constraints and nonlinearity. giving credit to the strong involvement dents who placed that category as one of the top three contributors and cooperation between the vendor staff and the user staff. nonlinear. a controller was tuned in order to ob. that the process owner re. Bauer and Craig [9] report • Control hardware (upgrade) (35%). • If MISO.000–15. or PMBC to ceived more benefit than expected from the ARC upgrade and clearly embody the process models. then there will Since the first stage in implementing MPC seems to be be a tendency to use it next time. • Consultant manpower cost (68%). overrides. and stationary—Consider inferential control. linear. because resources are not allocated to 6. The respondents also indicate that APC projects have a payback period of 3–9 months. Reasons for disuse include: process operating conditions (equipment reconfiguration. even an APC vendor [11] • If SISO. Perhaps management buy-in • If SISO. linear. . and could remove the primary APC justification—throughput increase. with well-behaved dynamics—Use gain the process performance. . and ill-behaved dynamics—Use IMC. to implementation cost. which was necessary for a successful APC realization’’. equipment revamp. constraints on present value of auxiliary variables—Use • Cost of technology (58%). but is this benefit due to MPC. throughput and quality. 5. were modified and optimized. • Other (5%). impact of good regulatory control and MPC on the economic even though it bends the lip on the can top’’. and fixing Products are widely available from most control vendor for PCs. but adequately credits economic improvement to ‘‘. rotation of operators and the proportion of control loops using PID controllers within the engineers is not accompanied by adequate training. ADRC. For instance devices and inferential measurements • PI(D)—It is simple. which is consistent with the 6. Use It is often stated that fixing basic problems with a control sys. or an improvement in process capacity of a table knife. . APC increases throughput by about 3%–5%. • If MISO. two frequently named profit factors’’. which are directly related. Some people to new knowledge of the constraint?’’ say. with constraints on future PV values—Use APC. Bauer benefits. companies do not justify the economic re-investment to Use the right tool for the job: A kitchen worker may need maintain the control system.14]. all the world looks like a nail’’. Bauer and Craig [9] report that the main costs associated with • If MISO. Interestingly. knowledgeable about where to find one. yield. nonlinear. well-behaved dynamics—Ratio. ‘‘I know I have a improving the basic control system. were • Maintenance cost (30%). When to use which? maintain the APC [9. Darby et al. Canney [12] estimates 9 months on his web site. . well-behaved dynamics—Ratio. based on favorable a new measured constraint. considering damage to the tool incentive will remain strong. numbers in parenthesis represent the number of respon- a particular project. it is difficult to separate the paint can opener somewhere. nonlinear. • If MISO. fuzzy. Wade recalls an instance where. . gain scheduling. for instance asks.

both technical and organizational issues . Darby significant change is impending. interactive. controller design and tuning. or time-constants change with state) or nonstationarity (gains and 6. providing a set of • If MIMO. operating rate). or 3 MVs—Consider ARC. implemented as ARC or MPC. classical control and to jump from one set of constraints to another as supervisory model-predictive control is needed to consider whether sensors optimizers economically adjust set points. But. staged decommission by solution. or lack • DeltaV Predict and PredictPRO (Emerson). primary or upper-level loops. if a Related to choosing an appropriate control strategy. DCS migration of ARC is typically time-consuming. decouplers). so these will have to be The reasons to use MPC are grounded in model-predictive ac- user-structured. • DMC Plus (Aspen Technology). ’’ and that ‘‘. (Notable. Non-technical aspects • FLSmidth Automation ECS/Process Expert for Cement and Mineral Applications. with significant dead time and one or more significant disturbance • INCA (linear. MPC performance degrades. . subject to constraints. For more There are some residual reasons that seem to be disappearing. For equivalent performance. interactive. Consider the following example: A single controlled variable • Connoisseur control and identification package (Invensys). GMC. An underappreciated issue with ARC is one of identifying • Applications with a small constraint set where economics the required models (e. .g. • Cascade and hierarchical applications. are • Situations where the controller undesirably causes the process critical . and would need dynamics for the CVs—Use a hierarchal control structure with dynamic models for each stage. loops that include: switch between MAN and AUTO. Issues that might lead to implementing • RMPCT (Honeywell). and in the future. or from a combination of simpler measurements? it needs the same sensors. complicated strategies. ARC is a viable option. Linear MPC is not the However. . perhaps nonlinear MPC is applicable or linear MPC default by the operators a year later is also often mentioned. and dissimilar • Processes that cycle through distinct stages. linear. subject to constraints (DoF < 0) or • Batch—nonstationary. which shape the identification and analysis. The commercially available choice. scheme. nonlinear. requiring model re-development. Nonlinear MPC the model parameters is not a viable option. Extracting information from [13. • If MIMO. a • 3dMPC & Expert Optimizer (both ABB). or a hierarchal • Situations where continual engagement of operators on the structure. measurements. HPC. DoF handling (economics and present and future constraints). • ControlMV. with MPC with only base regulatory controls implemented in the • eMPC (eposC) and Control Station’s LOOP-PRO. which variables should be included in the lower level ARC and which in the upper-level APC? From the Wikipedia Web site [17]: The answers to these questions depend what operating conditions ‘‘Commercial MPC packages typically contain tools for model will affect the plant today. or PMBC. PharmaMV and WaterMV from Perceptive Engineer- migration of MPC to a new DCS is typically not.2. with any control APC supervising a lower level ARC. subject to simple constraints. the Are valves adequately sized to provide a linear response. trial-and-error tuning of extent that the model needs to be changed. interactive. packages include: 6. issues? How should plant tests be designed to reveal plant dynamics without hitting future constraints? When is the model 7. are often limited to first-order plus dead time. • If MIMO. or non-stationary—Consider nonlinear MPC. Can reliable inferential • Applications where critical sensors or analyzers are subject to measurements be generated from time-compensation of the frequent faults. DMC. longer term plan to develop a more extensive MPC strategy. . . subject to constraints (DoF < 0) reasons to not use MPC. Batch) from IPCOS. not subject to constraints— 6. with gain scheduling or multiple models. disappearing reliable enough to be able to provide accurate. where engineering principles cannot be • Processes that are likely to change frequently or soon. the strategy in MPC include company expertise and knowledge. . . and either nonlinear • Processes with three or fewer interacting MVs. ARC will likely need retuning et al. a company with several different • APC Library (Siemens|PCS7). DCS systems. physical structure. which have intermediate which requires strong payout rates. [16] discuss the ‘‘. ADRC. without maintenance. and as the technology evolves. and problem with sensor faults is equivalent with both ARC and are they properly functioning to prevent sticktion and constraint MPC.8 Editorial / ISA Transactions 50 (2011) 2–10 • If MIMO. Regardless.) • If MIMO. This can be expensive. If ARC has the same functionality as an MPC. process is desired to keep them trained and able to react for abnormal event management.g. independent tuning may be needed for each stage. linear. If there are skills in the company and • ADMC & APCX1 (both Cutlertech). (The authors are not aware of commercial GMC or PMBC products. to the used to determine the model parameters. Why not use MPC? Use PFC. Such a strategy could be • Pavilion8 (Pavilion Technologies). Smith predictors. significant ‘art’ aspect to the application or restructuring also. nonlinear.to 9-month pay-back periods are often indicated. there are things that MPC is not best at doing. and multi-model MPC can cope with this. of a standard DCS platform (e. complete information in a timely manner.).14]. as well relative importance of the pros and cons associated with each as controller performance evaluation. . ing. large number of MVs. interactive.) tion. .3.. of time-constants change in time). variables are available as feed forwards. This is somewhat of are in the right locations. ’’ Experience with the process. • Control loops where the primary problem is nonlinearity (gains Payout is usually the consequence of throughput enhancement. the necessary capability in the distributed control system (DCS). . these or opportunities (DoF > 0). . Most modeling packages on the market are tailored to MPC not ARC. Commercial MPC products adequate? In a hierarchical structure.. and reduction in CV variability. and a complaint about earlier MPC implementations. fault-free. DCS systems) and a decision to standardize advanced applications • MACS (Capstone Technology). . well-behaved. which dictate a consistent constraint policy. and 2 coordinated MIMO control. etc. linear. and similar dynamics for all CVs—Use MPC (APC. opportunities (DoF > 0). Unless the process never variables and loops with relatively fast dynamics relative to the changes (product mix. large number of MVs. measuring enough variables.

RTO. ing systems to determine the value of intermediate and final products. • Creation of a modeling framework for control which combines • Expertune (Plant Triage). acoustic. and compensation. • Integration of planning. but RTO uses diagnosis. RTO changes can move from • Sustainment—monitoring and improving both ARC and APC. preventing degradation. But real time optimization (RTO) and APC • Automation of the hierarchal structure. nonlinear. impending flooding. disturbance) recognition. needs for control: • Shell Global (SMOC-II). transient engineers can understand both the process behaviors and the state. which would lead to autonomous changes in the system management rules. Plantwide control (coordinate control chart data).16. foaming. instantaneous prices and costs. and MPC products. impending log jams. accurate. that it is justified. in-process cost account- • Prime (RandControls)’’. • Honeywell (Profit MAX. With help from [18]. What variables should actions are strongly grounded in economic values. synthesis of control structure). the evaluation criteria. one for each section. total and for each loop. with no assessment of the impact perhaps within 6-sigma plans. only update a model when utility reveals • ControSoft (Mantra). on utilities cost or product variability. 8. etc. scheduling. clinker formation. Scheduling creates a • Autonomous abnormal event (fault. including adaptive. when inferential techniques use laboratory data for updates. by named products. and former Dot Products Nova and STAR).18] here are additional development • Perceptive Engineering (Perceptive). This is in contrast to controlling state vari- ables. room for ally the plant changes which makes the outdated model less improvement remains. the MPC challenges or concerns expressed earlier. Product improvement • Continuous monitoring of economic benefit. costs. in alphabetical order by vendor name. units. complexity for the application). Here are some development opportunities: than ideal).1. and EnTech). analysis of potential APC or RTO application. Although research and development in control and related • Evaluation of economic impact of ‘‘model degradation’’ (actu- automation support has provided wonderful tools. tomer satisfaction). • Better integration of laboratory analysis as feedback action • CyboSoft (Model Free Adaptive Control). • Improvement in HMI technology so that operators and process The event to be recognized might be steady-state. hierarchy and strategies Evaluating the value of intermediate products is even more for what and how to control in a plant? difficult than determining a value of the product (considering • Diagnostic techniques for poorly performing controllers.) monitoring waste. Additional perspectives to guide improvement per unit per situation basis to facilitate estimation of economic benefit of new applications. Profit Suite. im- constraints. transition start & stop. for priority shedding of diagnosis of phenomena (cavitation. How can be part of the conventional ARC and what should be inputs we prevent RTO and APC from bouncing operating conditions and outputs of the supervisory MPC? Should there be one between constraints when there is an appearance of a penny to supervisory MPC or does the plant isolation of effects indicate be saved? that several smaller MPCs. wave of change that progresses through the plant. • Normal event (stage completion) recognition and action trigger. with data and comparison to historical models. reflect operational priorities. SS models. Listed below are only companies with and undesired outcomes of control action. or distributed MPC. observation of data. draining complete. process-model • Control technology to achieve improvement. • ABB (Optimize IT). for instance. safety. • Emerson (Delta-V NN and FLC. According to [9. froth. controller solution. to supervisory applications. etc. etc. • Control of perceived situations from visual. overall plant optimization instead of local • The use of economic uncertainty to temper control action. (Notably there are many companies that status quo. • Look-up table of benefits from post application audits on a 8. economics or based on old values. DMC-Plus). sniffer • The use of multiple objective functions. set point adjustments based on good enough to stop plant tests. Editorial / ISA Transactions 50 (2011) 2–10 9 • MATLAB Model Predictive Control Toolbox. or other such ‘‘cyber employees’’ that observe. evaluation of • Development of guidelines to indicate when the model is last night’s loop performance. Aspen Target. initiating calibration. RTO is a • Autonomous health monitoring of the process and the control SS model sandwiched between dynamic operations. sequential units. This would be a supervisory recognition of desired provide APC or ARC services.) • Knowledge Process Solutions (IPC). evaluate. • The development of robust. (This is • Hyperion (DMCplusTM). and APC.. pending undesired confluence of events. already getting research attention and could alleviate some of • Ipcos (INCA). feeds. not to hold at the based. do not engineers (data analysis. • Shifting of the marketing of FLC and ES from feedback control • LineStream (ADRC). • Automated model development and adjustment by computer- • Adersa (HIECON. • Cutler Technology (ADMC1 Adaptive Multivariable Controller). taste. • Development of tuning procedures that are easy to implement • Automation of every routine function of the operators and (prevent ill-conditioning. RTO uses system. cus- • Creation of adaptive MPC that auto corrects the model. • Standard methodology or tool to estimate the cost benefit • Yokogawa (APCSuite). Automate validation of new observations • Aspen Technologies (AspenOne. emulsion stabilization. is better? . • Matrikon (ProcessACT). one constraint set to another. balancing need extensive simulation testing). but APC action is devoid of and advise operators and engineers. Using utility as • Capstone Technology Corporation 2(MACS). adjusting cycle times. empirical models and fundamental models (of appropriate • Gensym (G2 products). and PFC). • Universal Dynamics (Brainwave). scheduling. inventory. here are additions. Sales activities). • Development of MPC Cascades.14. agglomeration.

Journal of Process Control 2008.1. an IFAC symposium. of model-based and conventional control: a summary of experimental results. control educators related to the preparation of students within [11] Honeywell process solutions. Badgwell TA. Undergraduate process control experience control. ‘‘Model predictive control’’. WMCanney@ModelPredictiveControl. www.04. • Add process control laboratory experience to the undergrad- Houston. loop structure and tuning. Control theory is not the essential issue [17] Wikipedia. p. etc.10 Editorial / ISA Transactions 50 (2011) 2–10 • Development of technology for making MPC action more Acknowledgements aggressive when CV is changed by unmeasured disturbances? • Automatic and robust updating of inferred property rules The authors appreciate the review and feedback from Dave (inferential sensors. information such as process order. Experimental comparison of control strategies. and know how to correct and prevent it. 13106 Dogwood Blossom Trail. intelligent controllers who can baby-sit their process.org/wiki/Model_ for undergraduates seeking to go to industry.: +1 405 744 5280. USA uate program. will need to understand the rasion d’etre of their process.11:733–64. Grouped neural network modeling for model predictive 8. Diminish Laplace transforms to the role of a historical language of communicating R. ISA Transactions 2002. a white paper. unit operation models. Jacques Smuts. Russell Rhinehart ∗ process and controller dynamics. http://hpsweb. permissible industrial tuning Tel. [7] Joshi NV. [1] Richalet J. • Inclusion of consistency relationships in the model. vol. Active disturbance rejection control: a paradigm shift in feedback MPC is applicable.12(6):573–80. but it Available online 24 November 2010 does not usually cover instrument system calibration. Lecture 3. Houston. Economic assessment of advanced process control—a survey and framework. The one control course in the ChE program is adequate to reveal the 20 October 2010 ‘‘tip of the iceberg’’ of PID feedback control to students. The key is tuning for disturbances! Principal CMiD Solutions. For the plant engineer. [9] Bauer M. Paruchurri VP.com [accessed 12.net. APC. ISA Division Newsletter: There is too much academic emphasis on tuning for set point changes—in part due to IMC/lambda tuning Mark L.06. 2006. Use pilot-scale equipment. Hydrocarbon • Let go of the technology legacy of PID control to a set point.2. Laplace transforms need to be understood only as a carrier of 423 Engineering North. [3] Gao Z. New York (NY. The topics discussed in this paper also present a challenge to [10] Canney WM.06. [2] Han J. communication network. http://en. 57. E-mail address: wade@wadeco.83–8. Koc University. 3.18:2–18.10].wikipedia. Do not let the joy of the mathematics mask technology. Gupta A. Drop frequency analysis and z-transforms from undergraduate courses. 17522 Bushy Ruver Ct. Process control through nonlinear modeling. Safety Instrumented 6338.2010]. TX 77095. TX 77065. and Alan Hugo. USA • Add Automation Engineering degree programs to universities. APC: a status report (The Patient Is Still Breathing!). valves. Tel. Paper 239. The future or advanced process control promises more benefits and sustained value. 2399–405. control system design. • Automatic and robust updating of steady-state models used in References RTO (real time optimization) to keep them true to the plant. Riggs JB. Sullivan GR. identification. 14–6. 2009. [18] Qin SJ. Students Maverick technologies. signal Inc. fax: +1 405 744 optimization. O’Donovan D. [4] Lee PL. Keynote understand health. Simpler control methods often provide better results. Comparison steady-state gains. Control Engineering Practice 1997. Do not use bench-top engineering-science experiments or computer simulators for Harold L. Predictive functional control: principles and • Improved robustness to field instruments (sensor transmitters. Tel.41(2):195–202. health monitoring. Computers and Chemical Engineering 1998.06. predictions are corrected by future estimated residuals. Oil & Gas Journal 2003.honeywell. Here are some perspectives on how education should be com/Cultures/en-US/NewsEvents/SuccessStories/Success_BaASF [accessed 28.mavtechglobal. A survey of industrial model predictive control supporting theme. BASF ammonia plant increases production and undergraduate engineering programs for automation and control achieves ROI in six months with profit controller. the primary course objective. In: Proceedings of the 2006 American control • Models of the unmeasured disturbance so that future model conference. recognize symptoms.35(10):3547–59. Control material and energy balances. [6] Subawalla H. Include [5] Rhinehart RR. OK 74078-5021.5(7):885–96. and gain Stillwater. Rhinehart RR. [15] Kern A. VLE. Harmse M.III(7):86–90. 2010. They will need to understand and recognize misbehavior [16] Darby ML. soft sensors) which estimate the CV value Schnelle. changed to support the automation engineering workforce needs: [12] Canney WM. Oklahoma State University. Present the math predictive_control#Overview page was last modified on 15 May 2010 at 11:18 [accessed 02. Pandit HG.com [accessed 05.101(16):48–54.13(1):19–23 [in Chinese]. data logging. and prepare engineers to become the parents and coaches of [14] Ford JR. [13] Hugo A.08]. 1990. Automate Unit Operations Laboratory process E-mail address: darbymark@sbcglobal.: +1 281 304 9502. ∗ Corresponding author. or adequate courses to obtain a minor in automation. No. Nikolaou M. Rhinehart RR. 1 Systems.com. Students need to experience Principal Wade Associates. industrial applications. and implement a cure. Darby 1 approaches. . final elements). Generic model control. 2009. equipment.. [8] Ou J. School of Chemical Engineering. InTECH.edu. transmission. Craig IK. USA): Springer. p. Control Engineering Practice 2003. Processing 2000. isolation). An inferential update.10]. wiring protocol. Rhinehart RR. ARC. DCS structure or operation. for the controllers. May/June 2010.: +1 713 927 8709. as constraints in model optimization. USA value. Control and • Static transformations to linearize the process I/O so that linear Decision 1998. Auto-disturbance rejection control and its applications. careers. etc. They will need to ADCHEM 2009. 2 practices. sensors. In: Proceedings of the international symposium on advanced control and chemical processes. etc. Wade 2 the chemical engineering lab. electronic aspects (grounding. • Continued improvements in methods for closed loop plant Industrial and Engineering Chemistry Research 1996. Murugan P. MPC: current practice and challenges. dead time value. and analysis that is fundamental to control as a secondary. diagnose the disease. Paraphrasing a comment from Greg Shinskey in a 2003 E-mail address: rrr@okstate.