Process control technology plays a significant role in improving and maintaining efficient process operations. It influences the strategic and operational goals of enterprises, economic results, development and quality of products, continuity of production and competitiveness within the marketplace.

Over the past years, industries have invested heavily in model predictive control (MPC) controllers. Many within the industry believe MPC is the only way for advanced control that facilitates process improvement with the high project return on investment. In some organizations, MPC tools are the company standard; other methods are taken as nonstandard low-level technologies. Some organizations refuse to accept other technologies no matter how well they perform. Nonprofessional control engineers always believe that dynamic matrix control (DMC) software is versatile, and they try to use DMC packages for basic stability/control problems.

Myth vs. facts.

Too often, many techniques can be applied for advanced process control (APC) and process improvement. For example, intelligent control (neural net, fuzzy logic and expert systems) are very commonly used in APC, and some other techniques such as supervisory control, system identification/modeling skills are most indispensable in APC. The objective of this article is to analyze current misapplication, challenges and engineering management of advanced control technology for process improvement and to present the prospect of advanced control technology while educating how to use advanced control correctly and efficiently.

Advanced control background.

Since the advent of the distributed control system (DCS) in the 1970s, it has been widely used. DCS provided a tool for easy implementation of existing control strategies such as cascade, feed forward, nonlinear control, Smith predictors, constraint control and even decoupling control. These are control schemes based on the proportional-integral-derivative (PID) single-loop feedback controller, and they provided the platform of distributed and supervisory control, which is called (advanced) regulatory control.

What it is. Advanced control is a systematic studied approach to choosing relevant techniques and their integration into a co-operative management and control system that will significantly enhance plant operation and profitability. APC includes more sophisticated strategies, such as intelligent control, adaptive algorithms and MPC tied to empirical modeling. As an improvement over typical process control, APC represents an enhancement in the performance of control strategies that results in more consistent production, process optimization, better product qualities and less waste. While regulatory controls maintain mass and heat balances, advanced controls manipulate the mass and heat balances to achieve the best performance or quality.

With the development of information/computer technologies, having a powerful server connected to the plant collecting real-time data opened up numerous possibilities for complementary technologies. For example, complementary technology is an inferential estimation technology (soft sensor) that would infer the required composition of the stream to be used in the control system, in the absence of online analyzers or long delays in the measurement signals.

Expert systems. Another example of advanced control is the expert-systems technology that captured engineers’ imagination back in the early 1980s. It is based on capturing knowledge of the best operators and transforming real-time data into useful information through reasoning and analysis. The technology has had success with applications such as special detection and signal processing, startup/shutdown, emergency and abnormal condition monitoring and diagnostics.

Fig. 1 shows the plant operations pyramid and the commonly used techniques of advanced process control. The pyramid as it stands, represents the data transfer rates for the applications. The inverse of the pyramid is also true for the computations per solution.

Fig. 1. Plant operation pyramid.

Disadvantages and misapplication of MPC

MPC technology no longer relies on traditional servo-control techniques, e.g., feedback control was first designed to handle effects from unmeasured disturbances and have done a fairly good job for about 100 years. MPC assumes that the knowledge regarding the process is perfect and that all disturbances have been identified. MPC is an open-loop system. There is no way for an MPC to handle unmeasured disturbances, other than to readjust at each controller execution, and the bias is similar to an integral-only control action.

This partially explains MPC’s poor behavior when challenged by disturbances unaccounted for by the controller. MPC has been sold as the ultimate solution to every plant’s control and optimization needs. But, as engineers point out, for a multitude of reasons, the results from MPC implementations often produce short-lived, sub-optimal and/or poor outcomes. The products now in the marketplace have been over-used and are forced to perform functions for which they were neither designed, nor should have been allowed, to do.

But MPC products were so heavily marketed and over-sold that the few voices of reason that may have existed were overwhelmed and could not prevail. Over the last few years, MPC is applied too much and is continuing to receive records of lackluster performance. It is estimated that more than 50% of applications are in “off mode” or do not work at all (worse than regulatory control); only about 10% are fully working well, according to some industry experts.^1,2

The limitations of MPC have been thoroughly exposed, although probably widely ignored.³ Many take MPC as a versatile control package. Result: There are many misapplications. Concerns for MPC include:

• MPC is based on tested local model. Working in its tested operating point, re-vamp work is inevitable once the operating point changes. Correcting model mismatch requires retesting and remodeling—very expensive maintenance action item. MPC controllers, particularly complex ones implemented on complex units with many interactions, require constant attention from highly trained (and highly paid) control engineers. This is a luxury that few operating companies can afford.

• Hiring contractors to do an APC project will take 5–12 months since they need more time to become familiar with the processes and to test the processes. All of this effort costs more money; normally, a small project with 6–12 MVs, will costs around $150,000–$200,000. Contractors are normally professional in software package that they are using, and may not be professional in process and control system design. So, they try to do projects in their ways that may work for a while, but it is NOT the best design mode.

• If we interview operators or observe their operation, we find that they are always changing the MPC limits. This is because MPC design cannot satisfy the operation, and they try to control the system by moving limits. Clearly, this corresponds to manual control. When feeds change or the outflow changes, the MPC is turned off. The system is unstable due to big swings because the material balance is broken.

• MPC packages tend to hide technical stuff, and good ones will assist through commissioning. MPC engineers must be able to understand the process objectives and the features that make control difficult—only then can the MPC be set up with correct DVs, MVs, CVs, etc., and simple compensation for nonlinearities. We can imagine that MPC could be applied to minimize the need for good engineering. But, that is a false economy. Licensors of MPC software will tell you that their algorithms will “optimize” operation by operating at constraints using the least costly combination of manipulated variable assets. That is certainly correct. Mathematically, that is the way the linear programming (LP) or qualitative programming (QP) works. In practice, however, any actual “optimizing” is marginal at best. This is due to several reasons.⁴

• MPC takes ALL of the credit for improved performance. The benefit calculation is traditionally based on vague industry experience and some assumptions that are not always correct. For example, the performance will be consistent, and all achieved benefits are due to the MPC.⁵ In fact, all MPC performance is not consistent because it is based on a local tested model. The achieved benefit relies on the whole control system design strategy, including regulatory control design, PID parameter tuning, advanced regulatory control scheme and advanced control design (inferential control, fuzzy logic, system identification, supervisory control, MPC, etc.). More important, once we modified the control strategy, tuned parameters and updated the design, we will find that the performance had already significantly improved. Is MPC really needed?

• Why are some examples successful applications? First, the engineers or designers have very good working knowledge about the process. They fully know how the variables behave, including control-action direction, operation limits, dead time, dynamic response, steady state, etc. Secondly, the process control is a “slow” system—any disturbance can be slowly eliminated by feedback control as long as the control-action direction is correct. Consider the MPC modeling process: the model is built based on a local test data, and then the designer will fabricate it according to their process knowledge. Is this the right direction? Is the dead time okay? Is the steady state and model gain fine? The model will be modified until the designers are satisfied based on their process knowledge. There is no model validation in MPC packages; the “prediction” function that they have is using the modeling data for data checking—this is nothing. You must have more than 99% data match in this “prediction.” In addition, this knowledge-based model with many parameters does change online based on operation: limit ranges, bias, model gain, etc. So, we can understand why MPC is successful in certain applications within a short operation range. As long as the action direction is correct, the model works. Few industry people ask if this is the best way to control and what is it worth?

MPC for simple control system.

If one uses MPC where a PID controller with feedforward could work just as well, it is just a misapplication. For example, what about using a multivariable controller (MVC) to control a fired-heater outlet temperature by adjusting the fuel? Even in the case where you have multiple control variables and a single manipulated variable, a set of PID controllers with some intelligent design would likely work better than the MVC in a single-tower control, or even in a couple of towers together. Other widely misapplied MPC applications are, for inventory control, basic stability control, long steady-state time processes, directly controlled valves using MPC, etc. PID with feedforward or many other control techniques are a simpler and better performing solution.

Combustion MPC design problem.

A good combustion-control system will save energy and improve the unit efficiency. This will depend on the fuel and air flow both in steady-state and dynamic modes of the process. In some petrochemical furnace systems, the fuel and air are independently controlled by temperature and oxygen. This absolutely devalues the combustion efficiency, in particular, for dynamic processes, e.g., when changing the temperature setpoints or disturbance happens.

Few combustion units are successfully controlled by MPC since the system needs more intelligent factors. Most MPC combustion-control designs use the fuel and air as MVs, and try to rely on the local tested model to achieve the firing rate demand. Obviously, this design does not consider efficient combustion. The correct design is to use the ratio of fuel and air as the MV. The idea for efficient combustion is to have air lead the fuel on increases in demand for and fuel to lead air on decreases in demand. On load increases, the air is increased ahead of the fuel. On load decreases, the fuel is decreased ahead of the air. This motivates us to develop cross-limiting combustion control strategies.

Material balance issues in MPC design.

A process system consists of inputs, outputs and internal dynamics; they work in a fixed balance relationship at all times. Sacrificing this relationship to get a local control/optimization is a short-sighted method, and it eventually will compromise the control/optimization, leading to an unstable state. It is necessary to maintain the materials balance at all times.

Consider these scenarios: 1) the feeds change; 2) outflows have a large change; and 3) tower reflux has major change. All of these conditions move the system into an unstable state or the systems experience wide swings due to a break in the balance relationship if there is no material balance control. MPC is widely used in process systems—the feed is normally feedforward to the MPC control system. When we test the system, we have to test it in a wide range, although, the MPC is still in a local model. The model gain is not able to adapt to the material-balance relationship when system inflows or outflows change.

Let’s study one example: Feed, F_in = 100; overhead outflow, F_OH = 20; bottom outflow, F_BT = 30; sidedraw, F_SD = 50. Our design is to control the sidedraw to maintain the whole system material balance and stability, and to optimize the production using reflux (R), F_OH and F_BT. MPC can be used for the optimization, but it cannot be used to control the material balance, because it totally depends on the local model by step test. The model gain cannot be adaptive to all situations.

For example, the current gain is F_SD = F_in = 50 =100, by optimization, the flow distribution change to F_OH = 10, F_BT = 20; so the sidedraw has to be F_SD = 70. The gain becomes F_SD = F_in = 70 = 100. It is clear that MPC does not work if it is still using the old gain. A material-balance supervisory control must be used to adapt all feed changes and outflow changes due to production optimization in maintaining system stability. This motivates us to develop a principle-material-balance-supervisory control system.

Advanced control challenges

Control technology developed very rapidly, and process control is now ubiquitous within industries. With increasing reliance on information technology, systematic decision-making strategies are essential for effective and efficient performance including petrochemical/refinery industry systems. To maintain its successes, the industry must be flexible and adaptable to new technologies, external pressures and changing markets. All of these challenges require systematic improvement methodology and advanced technologies for optimization, control and planning. The most commonly used advanced control techniques include:
• Adaptive control: Controller changes over time (adapts)
• Intelligent control: Fuzzy logic, neural network, expert systems
• Supervisory control: Optimization by certain objectives
• Efficiency inferential without analyzer: Optimization and modeling
• Model predictive control: Multiple inputs and/or outputs decouple solutions
• Nonlinear control techniques: Nonlinear gain scheduling, neuro-fuzzy control and challenging to derive analytic results
• Process modeling and system identification: Principle and empirical models
• Optimal control and stochastic control: Controller minimizes a cost function of error and control energy or the controller minimizes variance.

Developments.

Most large companies have their own R&D or engineering group to manage these challenges. They normally develop the control methods for one unit and then transfer them to others. These developments are very necessary and helpful for company’s long-term business goals. It is not good or difficult to hire contractors for some specific developments. However, some R&D or engineering groups do not have the abilities to develop control technology for their facilities, although they may have some process engineers with PhDs within the groups.

Some development work can be very ridiculous. For example, consider tower-flood prediction by using a simple material-balance calculation, and then applying it to every unit. Likewise, consider using a simple MPC controller to manage onstream-time to evaluate performance, and requiring all units to keep the controller on whether it works well or not. The superintendents always show off their controller onstream time is above 90%. If the time is less than 90%, this will affect operators’ bonuses. Here are some of practical process engineering APC topics and development issues that we always face:
• DMC gain scheduling: An adaptive solution is needed.
• Combustion control strategy for furnaces and heaters: Cross-limiting or fuzzy logic reasoning are needed.
• pH control problems: Nonlinear control gain scheduling is needed.
• For some tower level cascade flow control systems, the level control is not important, the flow is important and is required to be as stable as possible: Limit constrained selective control strategy, and supervisory control strategy are needed.
• A tower feed or production flow change will cause instability or big swing: Material balance supervisory control is needed.
• Controlling a tower bottom flow based on level and pressure drop in regulatory control: Selective control and logic selection mechanism are needed.
• Inferential model development: System identification and fuzzy logic or neural network techniques may be needed.
• Quality measurement is based on lab samples; there are no online analyzer. Accurate inferential model can control these processes: Rule-based fuzzy logic control or expert-system control strategy is needed.

APC engineering management

An important APC benefit is from management. A knowledgeable, judicious management is able to pick up a reasonable technology, maximize benefits and minimize costs. The limitations of MPC have been thoroughly exposed, although widely ignored.

Management of many companies has been seduced by the popular myth that MPC is easy to implement and to maintain. This is a misconception fostered at the highest management levels by those most likely to benefit from the proliferation of various MPC software packages. But this is not an overly pessimistic picture of APC as it exists. Advanced control is very promising provided that management clearly and correctly understands the APC concept. A judicious management should encourage people to develop the best approach for specific applications using innovative technologies; to build reasonable process performance tracking tools; to pay more attention to basic regulatory control and tuning work; and to aggressively seek change and/or improvements.

Falsely seduced by technology.

There is a wrong concept: MPC/DMC is better than traditional control. Many managers and process engineers believe this notion that MPC/DMC controllers can be deployed into every plant system. Nobody is asking: Is it necessary? Does it provide benefits? How does it work compared with other techniques? Does it involve higher costs (design, test, commissioning, maintenance, revamp, etc.)?

A simple question should be considered: MPC/DMC are not suitable for most situations in refineries and petrochemical facilities. Table 1 lists a typical comparison for a small process control system, say, 8–12 MVs using improvement-based intelligent control design and software package based on standard MPC design.

Some control-engineering groups are willing offer software packages (e.g., DMC) installer; MPC packages exactly provide the tool for those who are eager for instant success and quick, short-term profit. They can even collude with the vendors to push selling their technology and services. There are incidents when a client wanted to have the MPC implemented on a particular unit even after they were told that it wouldn’t add any value. Who cares whether it would work or not, these clients only wanted a MPC nameplate on their unit.

For political reasons, many ridiculous things can be seen in plants. For example, a simple controller is designed as reflux subtract on an operator setting constant to control material balance of a tower. Although it does not work, especially when feedrate changes or other flow changes occur, it is reported this design can save energy at more than $1 million/yr, which corresponds to 5–6 months of total energy usage by this tower. This “advanced control” work was recognized by the superintendent and also published in the internal newspaper; nobody dared to say anything—surely all were blinded by technology.

Capital projects.

Most companies define MPC projects as capital projects, or almost 100% high-return projects are DMC based so that they can apply for funding and use outside resources.

The argument is: time is money; using a contractor can get the project implementation done as soon as possible and reap a return. Conversely, people forget that:

• Contractors need more time to study the system; their expertise is in the software package. Contractors may not be familiar with the specific application and may not have enough control technical skills. Result: The company will have to spend more time on the entire project, normally, even a small project may need 6–12 months.

• Extra cost will be needed to hire contractors.

• The designer always ignores the nature of real control problems and can pay more attention to the garish function of the software package.

• Instant benefit is often too tiny to define a true capital project if considering cost, maintenance, revamp, etc., and this tiny benefit is often easily obtained from basic regulatory control change and tuning. Typically, the short-term benefit of a DMC/MPC is 1%–2% improvement, while a true optimization technology-based APC long-term benefit is about 10%–20%, conservatively speaking.

Professional control engineer.

Technically, the professional control engineer is the right person to determine what control technology should be used; what package is suitable; what resource should be used, especially if there is no political or management interference. Control engineers must be proficient in control and improvement skills and technologies. Process knowledge is necessary but not critical in control engineering. Process engineers or DCS engineers have too much superstition about MPC because they may not have the knowledge about state equation, z transform, system identification, impulse dead time analysis, frequency response and so forth. Thus, process engineers try to use a software package to replace the real control system analysis and design. They try to use their process knowledge to analyze the system instead of practical data statistical analysis. A professional control engineer should also have strong knowledge about process improvement methodology (six sigma), control technologies (algorithms, control design, modeling) and process performance evaluation techniques (statistical analysis).

Control technique standardization.

Standardizing a control technique using one software package for company-wide plants is typically the wrong decision. The benefits from standardization are often touted. This is consistent with a recent trend in some large corporations, led by process engineering group or IT departments, to “standardize” technology solutions and then to “translate” the standard solutions from one problem to the next. The argument is: ease of implementation, low-cost maintenance, reduction in training, flexible movement across the whole organization, etc. Equally important, management is also concerned should the designer moves/leaves, nobody can take over or maintain the system. Thus, the automation group is told to standardize the company’s control strategy, DCS, MPC package; thus the operating company can use this standard exclusively for all control problems. Obviously, such an approach ignores the nature of real control problems in the plants and relegates control engineers to software transfer; furthermore, this standard technique will waste huge dollars, and lose significant opportunities for real improvement.

Advanced control prospect

MPC technology has its limits. It is the local optimizer and it does work well for one class of simple process. In particular, MPC technology works very well in labs or simulations because there is no unexpected issues such as model mismatch, extra disturbances, unknown condition changes and strong nonlinearity.

Fortunately, there are many other technologies for APC. It is worth noting that APC benefits will rely on the capability of technology, expertise of the designer and the reliability of APC methodology. It is vital for management to audit APC benefits scientifically, professionally and reasonably. It is strongly recommended that an audit team reviews and approves the APC projects and audits the benefits after the project. The benefit can be estimated as:

APC Benefit = (Optimum – Current operation) 3 C% 3
E% 3 R%

where C denotes capability of technology to capture benefits for the application

E represents the expertise of implementation team for the application

R represents the reliability of APC project methodology for the application.

These points should be kept in mind before APC engineering:

• Execute a detailed APC definition and justification study first. The study will provide a basis for investment and control technology determinations.

• Design the APC to solve operating problems and optimize the process/operations. Use the principle of control design techniques—simpler solutions to solve lower level problems, with ascending, higher level solutions to achieve more complex control, optimization objectives. Encourage control engineers to think openly and use more control strategies/technologies to improve systems. MPC is not versatile; designing it correctly and efficiently is the urgent work to be accomplished.

• When implementing APC, the purpose is to improve the process using the controller. Designers should pay more attention to process performance instead of controller evaluations. Professionally, there are many methods/calculations to evaluate controller performance depending on the objectives. For example, consider options such as constraint and capability analysis, along with performance evaluation in terms of user-specified benchmarks, but we have to know the objective first. Having standardized companywide key performance indicators (KPIs) and comparing them between the sites is too difficult, and sometimes not necessary. It is incredible to use MPC/DMC onstream time as a performance measure to evaluate the control system. There are many examples that demonstrate a controller showing high onstream time, but the process operating performance is very bad. The company will still use the controller because the superintendents want to have a higher number. Performance evaluation and benefit analysis are very necessary, and professional team should do this task.

• It is vital to track process critical variables (cpk or process capability.) The critical variables can be defined as CV in MPC, and/or something like energy efficiency or production recovery. This kind of performance is absolutely standard and can be easily compared between different sites for different purposes.

MPC is very attractive for some process engineers, but it is losing attention from many aspects (theory, modeling, cost, practical performance, maintenance, software itself). It is time for industry to have a revolution in APC; we need a control-oriented package. MPC can be embedded in the DCS, if needed. A judicious, intelligent-supervisory control system package will emerge soon; this will be the new era of APC.

Supervisory control.

It is a general term for controlling many individual controllers or control loops; a supervisory control scheme offers the prospect of process optimization and human operation intelligence. It is expected that the supervisory control package will provide more powerful functions and feasible supervisory algorithms. More important, it will provide the platform for users to develop user algorithm or functions, such as a material/energy balance model, fuzzy logic, neural network, genetic algorithm, optimization algorithm, decouple model, feedforward compensation, etc. MPC may be embedded into the supervisory system, but it will provide more modeling tools and model math expression. It will be a function block like a PID controller, just needing a few parameters for the user; No special software training is necessary.

In Summary

The automation technology is widely applied in a multitude of different industries. An extension of advanced control technology is an emerging advanced technology. The bottom-line driver for applying the advanced technology is its widely demonstrated capability to improve process performance; reduce costs for equipment and maintenance; and increase system stability, reliability and the capability of system fault tolerance.

APC design and implementation is state-of-the-art in modern control engineering, and is full of challenges both in theory and practice. After several decades of APC applications, it is time for control engineers, both in industries and academies to think about how to use it correctly and efficiently; to pave the path for its next direction; and, more important, move forward to the development and conception. But, the development and application of next-generation APC is definitely attractive and promising for industry. HP

LITERATURE CITED

¹ Ford, J., “Jim Ford’s views on advanced process control,” Hydrocarbon Processing, November 2006, pp.19–20.
² Friedman, Y. Z., “What happened to simple useful APC techniques?,” Hydrocarbon Processing, March 2008, p. 126.
³ Hugo, A., “Limitations of model predictive controllers,” Hydrocarbon Processing, January 2000, pp. 83–88.
⁴ Ford, J., “APC: A status report—the patient is still breathing!” White Paper from APC website.
⁵ Kern, A. G., “Outlook for multivariable predictive control,” Hydrocarbon Processing, October 2008, p. 33.

The author
	Dr. Jin Wang is working in the development of advanced control, system optimization and process improvement at Nalco Co. He holds BS and MS degrees, and a PhD in process control and industrial instrumentation with first class honors from Northeastern University and the Chinese University of Hong Kong. He held assistant professor position in the chemical engineering department of West Virginia University Institute of Technology, and research position at Carnegie Mellon University. He joined industry as a senior control engineer and engineering scientist for Lyondell and PPG Inc. He is a professional in advanced control engineering. Dr. Wang’s major interests include advanced process control, modeling, optimization, measurement system, robust and adaptive nonlinear control, intelligent control systems, nonlinear output regulation, with special emphasis on process improvement applications to refinery, petrochemical/chemical, metallurgical, power and other fields. He has authored over 20 reference journal papers and 40 conference papers. Dr. Wang is the senior member of IEEE, Control System Society and a senior member of ISA.