What if your drying operation could run close to the upper moisture limit boundary of your product specification, yet never exceed it? Say your product is powdered milk. If you overdry it, you give away product and spend more money on energy. Plus, you’re not running to the capacity that is possible to achieve because it takes longer to dry the product (hence, less product throughput). With conventional control (e.g., PID), you’re chasing ever-changing environmental/process conditions and incoming product variability. So, by the time you try to compensate, you’re off-spec again, and your product is probably too dry. What are your options?

Article Index:

While lower energy costs might not entice you to consider some advanced control techniques such as model predictive control (MPC), would higher capacity and additional throughput spark your interest? Though you might not have a dryer application, MPC might be able to help with other problems, even production line bottlenecks.

“With conventional PID control loops, the adjustment to the final control element is based purely on error and changes in error [process variable - loop setpoint - process variable],” says Brett Kachenmeister, Matrix Technologies industrial systems division department manager. “Because of this, PID algorithms have difficulty dealing with processes that have a long dead time and/or a long lag time. In many instances, they also have trouble correcting for any major disturbances to the process.” Therefore, the product quality may suffer, depending on how critical it is for the process target to be maintained. Sometimes, the addition of a feed-forward algorithm to the PID loop can help, adds Kachenmeister.

However, many typical MPC algorithms do a much better job of optimizing performance by dynamically adjusting the model and predicting changes to the final control element on the future time horizon based on previous performance. When the MPC model is built correctly, the algorithm can properly adjust for disturbances before the process is affected, rather than afterwards, says Kachenmeister.

Advanced control—What is MPC?

“MPC is classified as an advanced control technology,” says Chetan Chothani, president of Adaptive Resources Inc., a member of the Control System Integrators Association. It’s important to note the key word in MPC is model since most other advanced control technologies are not model based. “They start with extremely simple systems such as a ‘rule-based classification system,’ which uses a simple “If … Then … Else” structure,” states Chothani. “The more complex the process, the more convoluted this structure becomes until it is completely unwieldy.” It is, by design, suboptimal. Fuzzy logic is often included in the advanced control category, although it is more of a data preprocessing technique as opposed to a control methodology. Neither of these techniques actually utilizes a “model” in the more traditional sense of the term, explains Chothani.

“Neural network models, based on pattern recognition models designed after the workings of a human brain, are often used for advanced process controls,” adds Chothani. These models can be effective, particularly in soft sensor-type applications, and are naturally advantageous for nonlinear relationships. But they include certain disadvantages such as long-term drift, boundary conditions and extrapolation, says Chothani. “Consequently, these models are not included in the traditional definition of MPC.”

According to Mike Tay, product manager for Pavilion Products, Rockwell Automation, “Model predictive control is an advanced-process, model-based control technique that utilizes explicit dynamic models of the relationship between objective variables [targets and constraints] and existing process-driving variables [manipulated, existing PID loops and external, measured disturbance variables].”

The models are principally developed via empirical methods, are fast to develop and require only data. But they can oversimplify a problem and have difficulty extrapolating to new operating regions. On the other hand, fundamental models, which are relatively straightforward, can resolve many of the empirical models’ limitations. However, they can be more difficult to develop and more mathematically challenging to solve. In addition, they can take more time to process, adds Tay.

“MPC is an algorithm that utilizes a ‘model’ of the process to predict future behavior,” says Chothani. “Based on the deviation of the predicted process variable from setpoint, MPC takes corrective action in the present to minimize the integrated deviation from the setpoint over the entire prediction horizon. Included in the MPC algorithm is the ability to use feedback to correct its prediction and control actions.” (Figure 1 on this page shows a typical execution of an MPC algorithm.)

The first step in implementing an MPC application is analyzing the process and developing an understanding of how the variables may interact, says Chothani. At this point, the goal is to find as many potential variables for the model as possible and allow the modeling algorithm to determine causal relationships between them. The model parameters will then determine the level of correlation between the variables, and decisions can be made about which variables to include in the final model, says Chothani.

The earliest formulation of an MPC model was a simple step-response model, adds Chothani. “This model is developed by introducing a step change to the manipulated variable, the process input and the recording of the process variable over time. The model is then expressed as a series of gains over discrete time slices.” (See Figure 2 above.) A model that includes multiple variables will have a series of step responses and evaluate the relative contribution of all the changes over the relevant time history in all the relevant variables and predict how the process variable will respond in the future. Next, the controller will generate a series of potential future moves, predict future behavior based on these moves, evaluate the best series of moves to minimize the sum of the error over the prediction horizon and implement the first move, says Chothani.

“MPC provides a robust control, compensating for modeling inaccuracies or nonlinearities and, at the same time, includes the ability to anticipate future events and take control actions accordingly,” explains Anders Sehested, GEA Process Engineering product manager. “We apply system identification techniques to fit our ‘grey box’ models. In principle, we employ small perturbations of the system to understand the dynamics and interactions to be modeled.” MPC allows the current control input to be optimized while keeping future control inputs in consideration.

GEA Process Engineering uses MPC in its DRYCONTROL evaporation and spray drying process to provide it with disturbance rejection/feed-forward functionality to handle processes with time delays. MPC provides operation with multiple inputs/outputs (MIMO), allowing setpoint changes without affecting other controlled variables. MPC also provides optimization of a defined cost function, allowing for optimal operation, while considering constraints (stickiness of powder) and profitability (energy consumption, water content of powder and capacity).

“MPC can also function when measurements are seldom made and rely on the model between samples,” adds Sehested. (See Figure 3 on page 110.) “MPC can control product qualities such as powder residual moisture, whereas  PID cannot.”

Who should use MPC?

Obviously, not every food and beverage application is a candidate for MPC, but it definitely pays to look very closely to see if an application might benefit from it. “Most food and beverage processes can be properly controlled using conventional PID controller loops,” says Matrix’s Kachenmeister. “And, in many cases, their effectiveness can be increased by coupling them with proper feed-forward algorithms to deal with the majority of expected disturbances. These feed-forward algorithms are typically derived from empirical data sets from the actual process disturbances and are rarely linear. However, a polynomial regression is usually needed to make sure the feed-forward algorithm makes the proper adjustment,” says Kachenmeister.

In contrast, some food and beverage processes that have more indirect effects on the end product, such as the cascade temperature control of a product sterilizer, would certainly benefit from using an MPC algorithm, continues Kachenmeister. “For example, on many production lines, the same equipment is used to process many different products that frequently have or require vastly different process dynamics to maintain end product quality. With MPC, these types of thermal processes can be improved and will ultimately yield higher product quality and potentially use less energy.”

GEA’s Sehested lists processes that can benefit from MPC:

  • Processes with time delays, high-order dynamics and process interactions/cross couplings
  • Process control with constraints/boundaries
  • Nonlinear processes where the robustness of linear MPC can compensate for the prediction error
  • Processes where optimal operation is sought as a combination of several process states and actuator use
  • Processes like evaporation, drying, mixing, distillation and membrane filtration could be good candidates as they all have some of the above characteristics, and MPC could facilitate production closer to specifications, reduce giveaway, increase capacity and reduce energy cost.

An application with a measurable specification also is an ideal candidate for MPC. “In addition, process applications that use material balance and energy balance to retain quality are good candidates—for example, processes that use heating and cooling to develop a quality product,” says Rockwell’s Tay. Applications that deliver high value in evaporation, drying, standardization, utility centers and line management (such as balancing an integrated processing line), among others, are also good MPC candidates, adds Tay.

By the numbers, MPC can show a compelling reason to add it to, for example, a dryer application since it allows the user to run the process closer to the desired target. “Consider a  processor that has a product with an acceptable moisture range of 2 percent from 8 to 10 percent [moisture],” says Chothani. “If the process variability is +/-1 percent, the moisture target must be set in the middle of the range at 9 percent. However, if the process variability can be reduced to +/-0.25 percent by using MPC, the target can be set at 9.75 percent and still stay within the acceptable range. This allows the processor to increase the amount of moisture in its product by an average of 0.75 percent, reducing raw material costs and improving profitability.”

With GEA Process Engineering’s DRYCONTROL process, the MPC structure is built on top of an application’s current PID control, and when MPC is shut off, PID control is reinstated, says Sehested. As seen in Figure 4 on page 112, with MPC turned on, less product variation is visible, and it is possible to operate closer to the maximum allowed moisture level.

Are more sensors needed?

“In many cases, MPC is installed with existing sensors and control systems. However, a modern control system that supports integration process data required for the MPC [generally some upstream and possibly downstream process measurements] is required,” says Rockwell’s Tay. Occasionally, there is a clear advantage of having additional sensors, although this is dependent on the value and justification of the added costs, such as ambient humidity to support MPC on a dryer or feed quality measurement on an evaporator.

Generally speaking, additional sensors are not required, because advanced controls add value through consistent maximization against process limits, adds Tay. (See Figure 5 on page 114.) However, any delivery team should review the available measurements and process conditions to determine their adequacy and identify any missing requirements that may add significant improvement to the control capability of the final MPC solution.

“For many purposes, soft sensors/inferred values can be calculated,” says GEA’s Sehested. However, the soft sensor will need periodic manual feed reference values to eliminate any offsets in the model. “For critical-value, high-precision process parameters, GEA Niro utilizes inline instrumentation to get exact real-time measurements and enable the process to go as close as possible to any process constraints or specifications.” GEA Niro installs sensors on all significant sources of process disturbances, e.g., on the drying chamber itself to monitor the drying conditions, and a sensor for quality check of the final powder, according to Sehested.

“The most important sensor in the process is the process variable,” says Adaptive Resources’ Chothani. “So, for example, you must have a good moisture measurement in place. More often than not, this is the Achilles heel of an MPC strategy.” Many final product attributes are analyzed offline in a lab environment, sometimes using archaic measurement techniques and units. To effectively use MPC, online real-time measurements for these same variables must be determined. If the methodology or units cannot be directly calculated, a correlation must be developed that can provide a reasonable level of confidence in the control algorithm. So, if a total moisture measurement is not available, a reasonable correlation may be developed using a surface moisture (measurement) and other variables, says Chothani.

The second-most important factor, according to Chothani, is ensuring the manipulated variable can be modulated and provides a reasonably linear response. For example, if the manipulated variable is an exhaust fan, it is not enough that one can turn it on or off; a VFD must be added to the fan controls to allow modulation.

The third-most important factor is adding sensors for key disturbance variables, as it is always desirable to convert unmeasured disturbances to measured disturbances. For instance, if the humidity of the exhaust gas is a critical variable for control, a humidity sensor must be installed. Or, if ambient conditions such as temperature and humidity affect the primary variable, these types of sensors must be installed, adds Chothani.

“The best approach is to instrument fully the first application and collect as much data as possible, perform an analysis, run proof-of-concept testing, verify the results and finalize the system setup to be rolled out across the board,” says Chothani. Of course, this approach  can be used when the same application is to be rolled out to multiple sites. However, it is not very cost effective in one-off applications. Instead, engineering knowledge and judgment must be applied to select the variables and sensors.

What does MPC cost?

The costs of implementing MPC are highly dependent upon the application, the technology provider and the measurement and control updates that may be required. “The key thing to note is that you do not need MPC for all the loops in your process, just for the critical loops that can benefit from it and provide economic payback,” says Chothani. “MPC software itself can be obtained for as little as $5,000 for one loop to as much as $100,000+ for large implementations.” MPC control engineering could take as little as two to four man-weeks for simple applications and man-months for more complex applications.

“We generally see a one-year or faster payback on MPC projects,” notes Rockwell’s Tay. While these are additional, nonrequired investments above and beyond investments in standard control systems, they pay for themselves wholly in direct process improvements, such as  reduced off-spec or downgraded product, increased product yields, reduced energy consumption per pound of good product and increasing the throughput of product.

“Payback scenarios depend heavily on how well operators can mitigate effects from disturbances, e.g., changes in raw material and weather/humidity,” says GEA’s Sehested. “For example, large milk powder plants typically see a payback in six months to one year, based on between 0.2 and 1.0 percent higher residual moisture in the final powder. That is, the more moisture is sold as powder, [the] less moisture has to be evaporated, while staying within the specifications. The basis for this is a reduced variation of the process, allowing the plant to operate closer to the specification.”

Sehested reports this type of application can realize an extra capacity of 3 to 10 percent and 5 to 10 percent reduced energy consumption. Two other advantages are the plant produces a more uniform product, and MPC reduces the risk of fouling by retaining the point of operation defined in a safe window.

Rockwell Automation has had recent successes with line optimization, balancing a series of processing stages and equipment that turn raw food into packaged products, reports Tay. “The ability to incorporate many of the individual process unit operations, such as freezing, drying, frying, blending, etc., and coordinating an entire line to maximum throughput while maintaining quality limits have delivered high returns.” Bringing line rates up and down in a dynamic, balanced fashion has been very successful, and it allows the management of rate changes, outages and overall product quality in a new way. Operators have been very receptive to advanced system control, especially as they had challenges in the past keeping different operating units in sync, concludes Tay.

For more information:
Brett Kachenmeister, Matrix Technologies, 419-897-7200, blkachenmeister@matrixti.com, www.matrixti.com
Chetan Chothani, Adaptive Resources, Inc., 412-431-4662, cchothani@adaptiveresources.com, www.adaptiveresources.com
Mike Tay, Rockwell Automation, 512-438-1400, metay@ra.rockwell.com, www.rockwell.com
Anders Sehested, GEA Process Engineering, anders.sehested@gea.com, www.gea.com