Real-World Applications of Advanced Process Controls

As a rule, engineers live by the KISS rule—Keep It Simple, Stupid. Sometimes, though, a simple solution isn't good enough.In many process industries such as petrochemicals and food processing, advanced techniques are required to solve feedback control problems that are too tough for simple PID loops.

By Vance VanDoren, Control Engineering September 1, 2001

Key Words

Process & advanced control

Adaptive control

Process controllers

Open systems

As a rule, engineers live by the KISS rule—Keep It Simple, Stupid. Sometimes, though, a simple solution isn’t good enough.

In many process industries such as petrochemicals and food processing, advanced techniques are required to solve feedback control problems that are too tough for simple PID loops.

Ocean Spray

For example, PID loops were having a tough time maintaining constant temperatures at the Ocean Spray Cranberries’ juice bottling plant (Henderson, Nev.). Desert temperatures in excess of 100 °F would wreak havoc on the cooling water used to adjust the temperature of the juice as it is being bottled. Worse still, some of the erratically cooled juice would end up recirculating through the pasteurizer, where temperature had to be maintained at a specified point for a specified period mandated by the U.S. Food and Drug Administration.

Furthermore, the pasteurization process is key to creating the flavor and color of the final product. Poor control of pasteurization temperature leads to bitterness, unpleasant colors, consumer dissatisfaction, and lost revenues.

During the summer months, the plant’s original PID controllers could take up to 30 minutes to stabilize the pasteurizer’s temperature at start-up. Ocean Spray’s solution was to implement a predictive, adaptive, model-based controller called BrainWave from Universal Dynamics Technologies (Vancouver, B.C., Canada). Brainwave helped get the pasteurizer temperature up, keep it there, and maintain it at setpoint.

Start-up time was also reduced by 70% thanks to a feedforward scheme that anticipated incoming fluctuations in juice temperature. This allowed the controller to take corrective action before the pasteurizing process was adversely affected. And because Brainwave is adaptive, it was able to track changes in the behavior of the process, and modify its model and control strategy accordingly.

J&L Steel

An adaptive controller has also been successfully applied to a temperature control problem at J&L Steel’s annealing facility (Midland, Pa.). Annealing is the process of passing steel through a furnace, then rapidly cooling it to soften, toughen, or otherwise modify its physical properties. The desired effect is achieved by regulating the temperature of the furnace and the rate of cooling.

The facility’s conventional PID controls had been successful at maintaining the furnace exit temperature during steady-state operations. However, PID proved too slow to handle upsets when the dynamic behavior of the heating process changed due to variations in line speed, welds passing through the furnace, changes in the cross section of the strip, etc. As a result, some coils of steel had to be reprocessed.

Enter QuickStudy from Adaptive Resources (Pittsburgh, Pa.). J&L installed the QuickStudy adaptive controller on the furnace to account for the variable behavior of the heating process. First, QuickStudy’s modeling facility was applied to the input/output history of the process to characterize its behavior with a set of mathematical equations, i.e., a model. QuickStudy was then put on-line and allowed to continue fine tuning its models with the latest input/output data.

With up-to-date models of the heating process available all the while the line is in operation, QuickStudy has a better understanding of how the process behaves and is better able to control the furnace temperature. Zone temperatures now remain within 5 °F of setpoint.

Air Liquide

Houston-based Air Liquide America Corp., a unit of Air Liquide Group (Paris, France) has used a model-based predictive controller in a similar manner, starting with generic models of the process, and fine tuning them on-line. In this case, however, the process is an air separation unit (ASU) in McMinnville, Ore. The ASU filters, dries, cleans, and cools atmospheric air, then extracts nitrogen, oxygen, and other products from it using a series of distillation columns.

To coordinate all associated flow rates, Air Liquide decided to implement a multivariable predictive controller capable of manipulating more than one process variable at a time. The controller selected—GMAXC from Intelligent Optimization (Houston, Tex.)—was also required to pursue multiple objectives, instead of simply maintaining its process variables at predetermined setpoints.

That job was left to a series of single-variable PID controllers, each assigned the task of regulating its independent feedback loop. GMAXC was responsible for choosing their setpoints, so that they would collectively act to maintain the quality of the distilled products, maximize yield and throughput, and counteract transient disturbances. This scheme is often called hierarchical control , with a supervisory controller at the top of the hierarchy providing setpoints for the regulatory controllers at the bottom.

The GMAXC controller used a model of the overall process to predict future effects of its current setpoint selections, allowing it to guide the process along what should turn out to be the most profitable path. And, since the model could be automatically updated while the ASU was in operation, GMAXC was theoretically capable of optimizing the overall performance of the process even as its behavior changed.

It worked. Air Liquide’s hierarchical control system has been able to meet its objectives, while maintaining the plant’s stability without operator intervention. By increasing the efficiency of the distillation process, the control system has also reduced the electricity required to generate each ton of product. GMAXC paid for itself in less than two months.

Help for lone controllers

Still, Air Liquide’s plant engineers thought they could do better. Some of the stand-alone PID controllers elsewhere in the plant were performing poorly, especially the loop regulating liquid level in the high-pressure distillation column. Aggressive PID tuning for the controlled flow entering the column had caused excessive oscillations, whereas sluggish tuning had reduced those oscillations, but also lowered the product yield. Even when optimal tuning was achieved, subsequent changes in process behavior (such as changes in the flow rate of the uncontrolled streams entering and leaving the column) would render the optimal tuning obsolete.

Air Liquide turned to the CyboCon model-free adaptive controller from CyboSoft, part of the General Cybernation Group (Rancho Cordova, Calif.). Like Brainwave and QuickStudy, CyboCon is an adaptive controller capable of tracking changes in the behavior of the controlled process, but it does not use a process model for the purpose. [For more on CyboCon’s alternative approach to adaptive control, see “Model Free Adaptive Control,” Control Engineering Europe, Feb./March ’01, p. 25, and online at www.controleng.com.]

CyboCon was able to provide tighter control than PID without causing the flow rate to oscillate. The overall process now runs more smoothly and product yields have improved. Distillation column purity has also improved and the plant is now setting new production records. Air Liquide hopes to do even better by combining GMAXC at the supervisory level and CyboCon at the regulatory level in future applications.

AngloGold

Another hierarchical control system was implemented by AngloGold North America, a division of AngloGold (Johannesburg, South Africa), and Meridian Gold (Reno, Nev.) at their Jerritt Canyon mill (Elko, Nev.). Here ore is crushed, dried, and ground to prepare it for extracting the gold it contains.

The problem at Jerritt Canyon was that new deposits of harder ores required more grinding to achieve the required particle size. Additional horsepower and a hierarchical control system were installed to alleviate bottlenecks and increase plant throughput. As at Air Liquide, increasing throughput was intended to lower the fixed cost per ounce of gold recovered.

However, excessive throughput can overload an ore mill and bring it to a grinding halt. Maximizing particle size can increase throughput without risking an overload, but can also hinder gold recovery in downstream extraction operations.

A hierarchical control system incorporating AutoPilot supervisory control software from PSE/Optima (Salt Lake City, Ut.) was installed to assess throughput/ particle size tradeoffs and operate the mill at the highest possible efficiency. Unlike GMAXC, though, AutoPilot’s decisions were made according to a set of rules designed to mimic the actions of expert human operators, hence the term— expert system .

AutoPilot’s expert system monitored several process variables, including the mill’s power consumption, which drops precipitously when overloading begins. It then determined optimal setpoints for the incoming ore feed rate, air flow rate through the grinder, and the speed of the particle size classifier. These values were downloaded to the regulatory level of the control hierarchy for implementation by three independent PID loops.

By effectively coordinating all three control loops, AutoPilot has been able to increase throughput by 8% and eliminate the bottleneck caused by the grinder’s limited capacity. The new system paid for itself through increased production in 10 weeks.

ASARCO

A virtually identical control problem was addressed at the Mission Mill copper processing facility operated by ASARCO (Sahuarita, Ariz.). A hierarchical control system was implemented to maximize the ore grinder’s throughput, but in this case the supervisory control software—Connoisseur, from The Foxboro Co. (Foxboro, Mass.)—based its decisions on the future behavior of the process rather than the past behavior of expert operators.

The Connoisseur controller was configured to optimize overall plant performance by analyzing the process model for optimal operating conditions. As at the Jerrit Canyon gold mill, the optimizer tried to maximize feed rate, either at the mill’s physical limits or until some other controlled variable, such as particle size, reached its maximum allowable value or constraint . Whenever the feed rate was already at its maximum, the optimizer tried to minimize particle size to increase copper recovery downstream.

Connoisseur can account for two classes of constraints—”hard” and “soft.” Hard constraints on the process variables are never violated. Soft constraints on the setpoints that Connoisseur feeds to the regulatory PID controllers may be violated if necessary to satisfy process variable constraints. Connoisseur includes a prioritization scheme for defining which soft constraints are most important.

By constantly pushing the process to operate as close as possible to its constraints, Connoisseur was able to maintain a high throughput without causing significant overloads. Particle size and variations in particle size have been reduced and the feed rate has been improved during periods of maximum ore availability. The mill now remains fully loaded at the point of highest grinding efficiency.

Los Pelambres

Yet another application of hierarchical control for ore grinding has brought real-time expert systems, on-line adaptive modeling, and optimization together at the Los Pelambres copper minerals concentrator in central Chile. KnowledgeScape from KnowledgeScape Systems, a division of Baker Process (Houston, Tex.) served as the supervisory controller.

Again, the object was to operate the grinding mill at the highest possible feed rate no matter how hard or soft the ore. To do so, KnowledgeScape employed an expert control and optimization system based on genetic algorithms (a form of artificial intelligence) to emulate the functions of the expert plant operators. KnowledgeScape also used a neural network to continually model the grinding process.

These features allowed KnowledgeScape to evaluate alternative operating conditions for future effectiveness, then implement the best set, provided it was going to outperform the current control algorithm. KnowledgeScape’s decisions have proven to be so accurate that over 50% of the mill’s regulatory setpoints are now being generated by Knowledgescape and production has increased by 8%

Perhaps these applications of advanced controls pass the KISS test after all. They seem to have made life simpler for every one of their users.