Model-Based Predictive Control: No Longer a Black Art

Aug. 12, 2013
No longer a black art or an expensive add-on, advanced process controls have become much easier—and cheaper—to install and maintain thanks to technical developments that are also opening the door to a wider range of applications.

Any talk of model-based predictive control (MPC) still conjures up images of dollar or euro signs in the minds of many engineers. For in the past, MPC and other forms of advanced process control (APC) have been something of a black art demanding lots of development work by highly skilled specialists.

The situation seems to be changing, though. Sure, production chiefs like Attila Bodocs still see euro signs in their mind’s eye. But now the signs represent something different at the Algyo Gas Plant, which the Budapest-based MOL Hungarian Oil and Gas Co. operates near Szeged, Hungary. For Bodocs and his colleagues, the signs now signify a savings of more than €1.2 million ($1.6 million) a year in energy costs alone on six distillation columns producing propane, butane and pentane.

Results like this have helped to change the perceptions about MPC that many in the process industries have held for a while. As users take another look at MPC and other APC technologies, they are finding that developments are often making the technology easier to deploy, thereby opening the door to a larger set of applications.

At MOL, perhaps the most important development was the embedded nature of the MPC technology in DeltaV distributed control systems (DCSs) from Emerson Process Management ( Because the technology comes already preprogrammed and embedded in the DCS, Bodocs found embedded MPC simpler to install than the more conventional versions that run on supervisory systems. He also reports that the plant’s operators find it easy to use and understand.

Like other forms of MPC, this embedded version is an APC technology that consists of a dynamic model of the process and an optimizer. The model predicts what will happen in the process in the future, and the optimizer determines the best set of moves in the manipulated variables that will bring the controlled variables to their targets and keep all constraints within bounds.

The technology relies on one algorithm to optimize the process, instead of a set of independent control loops. “In most processes, variables interact with each other,” says Pete Sharpe, Emerson’s director of industry solutions application development in Glen Allen, Va. “A change in one manipulated variable will affect more than one controlled variable and constraint.” Using the dynamic model, the controller accounts for these interactions and stabilizes the process at an optimum operating point for a given set of constraints, such as cost and yield.

Browse a library of models
The dynamic model used by MOL’s Algyo Plant is Distillation Optimizer, which is based on the SmartProcess libraries of predictive control models developed by Emerson’s engineers. “We’ve built a library of models for common processes, such as distillation columns, fractionators, boilers, blenders,” explains Sharpe.

Each module already contains the necessary calculations for describing the process. For a distillation column, for example, the calculations include such parameters as internal vapor-liquid traffic, pressure-compensated temperatures, and energy per ton. DeltaV’s PredictPro optimizer then uses these parameters to improve the stability of the process and to optimize the columns’ reflux ratios and other parameters.

For this reason, operators no longer need to run the process at a higher-than-necessary purity to guard against anomalies like storms and variations in feed composition from throwing the process out of spec. Because the multivariable models predict outcomes accurately and permit optimizing operating parameters quickly in response, operators can run the process closer to specifications.

At MOL’s Algyo Gas Plant, the result was a 35 percent drop in energy consumption, saving €1.2 million ($1.6 million) a year once the first five columns went online with predictive modeling. The savings exceeded the original goal of €734,000 ($973,210) set at the outset of the project and generated a return on investment in less than two months. The savings only increased once the sixth column received a new online analyzer and could go online with MPC, too.

Model-based predictive modeling is just one of the embedded APC technologies loaded in Emerson’s DeltaV controllers. The controllers also contain neural networks and fuzzy logic for solving nonlinear control problems. “Most processes are nonlinear with process gains that change over the range of operation,” says Sharpe. Although nonlinearity is often too small to matter, fuzzy logic and the other nonlinear APC techniques can improve accuracy in applications when it is large enough to affect results.

Once these APC applications are activated, users can deploy them, much like they would use control blocks in conventional programming. “You just drag your distillation block from the library and wire it in, configuring your flows, reflux, temperatures and so forth in the windows,” says Sharpe.

Because these advanced techniques are embedded in the controller, a supervisory control architecture is no longer required. “It’s just another block you have for solving control problems,” says Sharpe. “There are no database transfers, OPC drivers, and DCS programming.” Consequently, buying extra hardware and hiring experts to make it work is unnecessary.

>> More on Advanced Controls: Click here for additional information on advanced controls.

Driving these abilities has been a combination of increasingly powerful process controllers and PC hardware. “They have created a platform for bringing the most sophisticated technologies to standard components of a process control system,” says Ronald Nijssen, industrial automation consultant at Siemens Industry Inc. He reports that Siemens has added advanced engineering tools and libraries to the standard tool set in its Simatic PCS 7.

Consequently, APC tools are accessible to more users. “The functionality of tools and run-time functions is increasing from version to version, resulting in a process of ‘democratizing’ techniques that were only used for specific, high-end applications before,” says Nijssen. “In addition, the integration of the existing high-end solutions has become much easier due to standardization of protocols and alignment of control and engineering strategies.”

Automation simplifies use
The goal of embedding MPC and other APC techniques directly into controllers is to make them available to smaller applications. “Some industries, like refining, use APC quite extensively and have lots of experience with the big APC projects,” notes Sharpe. “But there are opportunities for smaller projects like the small distillation column here or there. With a week’s worth of work, you can get APC working on it.”

The sweet spot for DCSs with embedded MPC is projects that have a relatively small number of variables, according to Ric Snyder, senior product manager in the Information Software and Process business at Milwaukee-based Rockwell Automation Inc. ( For complex, large-scale projects, he recommends a more conventional, multivariable MPC scheme on a supervisory system. “In fact, MPC was developed in the mid-80s for large oil refineries,” he says. “Shell and Exxon did a lot of that early research.”

Since then, this more conventional style of MPC has also undergone a number of improvements, such as the emergence of what Snyder calls hybrid modeling. Before hybrids, model developers would choose one of two methods for building a process model: an empirical method based on crunching huge amounts of historical data, or a theoretical method based on mass balances and thermodynamics.

“Hybrid modeling takes the best of both worlds,” notes Snyder. “It uses whatever information is available. If it’s mostly empirical, then you use the theoretical equations to make the empirical model better. If it’s mostly equations, then you use whatever empirical observations that you have to make those models better.”
He attributes this ability to the continuing evolution of computing technology. “It’s rare to find a process today that doesn’t have a historian connected to it,” he says. “There is a huge amount of data out there that wasn’t available 15 years ago. You can mine it quickly and efficiently with the techniques available today.”

Likewise, hybrid controllers can also handle a mixture of linear and nonlinear control problems simultaneously. Rockwell Automation is finding these hybrids useful in nonlinear applications like polymer production, where a 10 percent change in a parameter can cause a 50 percent change in the melt index. “Because a linear controller has a fixed gain, it doesn’t know how to deal with these things,” says Snyder.

So, since buying Pavilion Technologies in 2007, Rockwell Automation has been tapping into that company’s expertise in neural networks and inferential sensors to develop nonlinear models for predictive control. “We’ve been focusing on making this hybrid approach a lot easier to use and useful for a wider range of applications,” notes Snyder.

Automated step tests
A simultaneous trend in MPC today has been a steady effort among providers to enhance the robustness, reliability and usability of their model-based systems. Anywhere from 20 to 50 percent of older models in the field either has been turned off or is only used on certain products, reports Snyder. “These models are very accurate for a certain range of operation,” he says. “But as companies started stretching production and making more with the existing equipment, the models did not extrapolate well and reflect what was going on in these new operating ranges.”

To solve these problems and widen the range of applications, MPC providers have been developing internal automation that simplifies the technology’s installation, use and maintenance. This automation is, in part, a response to a common complaint among users that previous generations of MPC software required a highly skilled expert, not only to install it but also to maintain it as processes change with periodic upgrades and normal wear and tear.

Hankinson Renewable Energy of Hankinson, N.D., is already reaping the benefits of such automation developed by Honeywell Process Solutions ( By installing Honeywell’s Profit Controller on a grain drying operation, Hankinson, a producer of fuel-grade ethanol, was able to control moisture content for maximum yield, while minimizing natural gas consumption. The result was an annual savings of nearly $800,000.

One tool responsible for these results is Honeywell’s Profit Stepper, a tool for automating step tests. Before such automation existed, engineers would perform these tests manually. “We did all of our step testing one variable at a time so we didn’t mess up the model,” says Perry Nordh, Honeywell’s product manager for advanced control and optimization layers.

Now, besides testing all variables at the same time, the automation also looks at multiple moves. It moreover accounts for cross-correlations, checks for frequency problems, and manages control limits. “Not only does that give you safer plant test and more accurate models,” says Nordh, “but the length of the implementation is also much shorter.” He estimates that letting the computer do the work makes the average job go about 40 percent faster.

Besides exploiting the computing power available today to automate as much of MPC as possible, Honeywell has also adopted an approach called layered optimization. The approach differs from traditional optimization schemes in that its dynamic models are very similar to the control models. “Essentially, we’re taking advantage of all of the modeling work that you do for your controller,” says Nordh. “We reuse all of that work in the optimization layer.”

In the conventional approach, the steady-state models for the optimizer are usually built from first principles. Because the control model usually is not developed that way, the conventional approach requires an effort to coordinate these two disparate models. A potential problem with having these disparate models is that the steady-state model can sometimes generate targets that the dynamic limitations of the plant make impossible to hit.

Not only does the layered approach permit managing both layers in the same environment, but its dynamic optimization also does not present an infeasible target. “Because we don’t have to wait for a steady-state condition for the model to converge, we can immediately move in the right direction—towards either an energy or quality optimum, depending on your setup,” says Nordh. For these reasons, he believes that the layered optimization has advantages for both timing and long-term maintenance.

Automated model maintenance
Another user benefitting from the automation available on today’s multivariable MPC is Ecopetrol’s Barrancabermeja Refinery, a facility supplying 85 percent of the refined fuels in Colombia. Because of past success with DMCplus from Aspen Technology Inc. of Burlington, Mass. (, Ecopetrol’s engineers decided to give the latest version of Aspen Inferential Qualities a try on a new hydrodesulfurization unit and new diesel and gasoline hydrotreaters.

As expected, the automation built into the software streamlined the implementation. Engineers from the vendor and Ecopetrol were able to develop the appropriate equations and have a model working on the units in about nine weeks, according to Robert Golightly, a product marketing manager at AspenTech.

Another key technology for the project was AspenTech’s Adaptive Process Control, which automates the maintenance of the model to ensure that it remains a true representation of the plant as its behavior evolves over time. Golightly identifies the automation of model maintenance as the most important technical advancement in multivariable MPC. “Traditionally, this maintenance required repeating most of the initial workflow for developing the controller,” he says.

Now, automation like AspenTech’s Adaptive Process Control can maintain the models by working in the background, performing the necessary tasks without disrupting the process significantly. The goal is to turn model maintenance into a built-in process that occurs while the controller is online and continuing to optimize the plant, rather than as an external project performed periodically.

>> ISA-88 for Better Batches Every Time: This standard is making industrial systems more flexible. Visit

By monitoring the process continuously, the system finds mismatches between the model and the plant. Not only does the software collect and clean the relevant data, but it also schedules and conducts regular assessments of the model using the data and key performance indicators. Small-amplitude step tests occur in the background during production without full-time supervision from an engineer. Using the results of these tests, the software automatically realigns the models to eliminate the mismatch.

Ecopetrol’s Andres Rodriguez reports that, since installing the MPC, the refinery has increased the life of the catalyst in the diesel and gasoline hydrotreaters and reduced the sulfur content in its products. By tightening control over the process, it also improved its stability, allowing it to run closer to quality specifications and reducing the amount of product that the company was giving away. All in all, MPC has saved Ecopetrol $3.7 million a year—$2.7 million on the two hydrotreaters and $958,000 on the desulfurization unit.