To view the recording of Deminar #7, click on the above picture. If you want to just view the slides click on Deminar #7 - PID Control of True Integrating Processes

Time is money. If you can get to optimum setpoints faster during fed-batch operations and for startup and product transitions of continuous operations, the increase in production revenue can be significant. For continuous operations there may also be an appreciable decrease in the processing, recycle, and waste treatment costs of off-spec material.

For cascade control, the speed of the secondary PID setpoint response largely determines the ability of the primary PID to get to its setpoint quickly and reject disturbances in the primary loop. A slow secondary PID setpoint response may require detuning of the primary PID to prevent interactions between the secondary and primary loops.

In Deminar #7 we explored how we could use PID structure options, setpoint feedforward, and bang-bang control to improve the setpoint response for integrating (e.g. batch) processes. The concepts are also applicable to the continuous process startup and transitions. The demos showed a big reduction in rise time (time to reach setpoint) by the use of "PID on Error" instead of "I on Error, PD on PV." The benefit of the additional bump from derivative action on error is rather marginal for the small rate setting used. In other words most of the speedup in the setpoint response could be achieved by "PI on Error, D on PV" unless there is a large secondary lag and hence a large rate time setting. The use of setpoint feedforward helped reduce overshoot, rise time, and settling time by about 25%. For deadtime dominant self-regulating processes, the improvement would have been more impressive. The most dramatic improvement occurred for full throttle bang-bang control. With some adjustment of logic and resting value as noted on slide 6, the bang-bang logic can also be effectively used for self-regulating processes. You can try out setpoint feedforward and bang-bang control on the virtual plant website starting August 20.

If you are intrested in reducing batch cycle time and startup time, check out Deminar #7 at 10:00 am CDT on Aug 11.

To attend the event, go to http://bit.ly/JC-LiveMeeting
Use the information below to connect (if you're not using the available computer audio):
• Toll-free: +1 (877) 771-7176
• Toll: +1 (225) 383-1099
• Participant code: 264679

Whenever I see real control valves with digital positioners and diaphragm actuators, I get a bit giddy with excitement. If on the other hand I see on-off valves installed to perform the role of process control, I just shake my head in dismay. If flows are turned on or off, there is very little process control opportunity. Flows, whether process or utility, are the levers for the process. If we can only jerk the levers around, we will have a jerky process. The Feb-Mar 2010 InTech article "Key Design Components for Final Control Elements" details this perspective as well as the essential design features needed. If you have throttled flows not only do you have a means of affecting but also a way of optimizing the process. It would be a rare coincidence if the flows were exactly at their best value at the right time. There is almost assuredly an opportunity to increase capacity or yield or decrease energy use by changing the flow to reduce variability and/or moving a measurement closer to it optimum operating point. Sure there are options to sequence the turning of flows on and off but such pre-programmed actions lack the feedback correction needed to deal with disturbances, non-idealities, and unknowns in industrial processes. Unfortunately, graduates from chemical or biochemical engineering programs may mistakenly be thinking they can set the flows per the process flow diagram and process design simulation program. Sure they probably had a course on control theory, but maybe all they got was a mathematical view of process control isolated rather than integrated with process research, development, and design.

If the fixed flow mindset results in the use of on-off valves and missing feedback measurements, the opportunities are difficult to identify and may require years and a bunch of money not only for the field instruments and valves but also for the piping and equipment modifications. Just think if you want to install a thermowell and there is no nozzle on the vessel or column in the right location? Also, on-off flows create the step disturbances you would hope would be relegated to control theory textbooks.

Dynamic simulations can show the way but a large expensive automation project can be a hard sell without an installed example. If on the other hand there are sensitive throttling valves and process measurements, opportunities can be trialed and implemented by taking advantage of the ever increasing incredible capability being built into the modern DCS. The key characteristic is sensitivity, which is the smallest change in the controller output or process variable that the valve and sensor, respectively will consistently respond to. Once the sensitivity threshold is reached the output will change by the full amount whereas the output will only change by a quantized amount that is a resolution limit, the other major component of precision. Often the term "resolution" is mistakenly used instead of sensitivity. Resolution, which has a stair-case response, was mostly an issue with rack and pinion actuators and older A/D converters with wide signal ranges (e.g. 1980s generation DCS thermocouple input cards). The resolution today of digital I/O far exceeds the sensitivity capability. The consistent precise response to change is more important than an exact match between input and output for valves. For example, valve span or bias errors (offsets) are clearly not much of an issue because the feedback loop will correct for them provided there is a full range of control possible. Measurement span and bias errors can also be corrected by upper loops or operating procedures, but accurate besides precise measurements are important for closing material balances for process analysis, diagnostics, and optimization as discussed in the Jan-Feb 2010 InTech article "Advances in Flow and Level Measurements Enhance Process Knowledge, Control"

Wireless measurements offer the opportunity to move the transmitters to find opportunities and the optimum location if the process and equipment design engineers had the understanding to provide the connection options. Wireless pH offers the ability to develop inferential measurements and prove the best electrode technology as revealed in the Jan-Feb 2010 InTech WEB Exclusive article "Opportunities for Smart Wireless pH, Conductivity Measurements"

I hesitated at first to include sample time as one of the exceptional opportunities in process control because in most loops it is not issue. Then I realized I should give my perspective on the effect of sample time for the following reasons:

(1) Since we live in a digital world, sampled data is the norm. Just from the volume of applications, the opportunity is large

(2) There are no clear guidelines for various types of process control applications

(3) In some applications conventional sample times can cause severe safety and performance issues

(4) In most cases the tuning of the controller dictates that sample times could be significantly slower. If DCS module execution times and wireless communication time intervals could be increased, controller loading is reduced and wireless battery life is prolonged, respectively

(5) If we want more at-line analyzers to provide measurements of stream compositions that tell us what is really going on in the process and offer the opportunity for a more advanced level of control, we need to understand and address sample processing and analyzer cycle times

(6) If we want to move to more wireless measurement that give us the flexibility and intelligence for process control improvement, we need to understand and address wireless communication intervals

I am considering sample time as the time between updates in sampled data in the broadest sense. The following discussion should be useful for determining whether DCS scan or module execution times, wireless communication time intervals, model predictive control execution time, and at-line analyzer cycle time will affect control system performance.

If you are pressed for time you can skip the discussion below and just check out ProcessControlSampleTimes.pdf

There is considerable confusion as to when sample times affect the ability of a control system to compensate for unmeasured disturbances. The following is my quick attempt to provide some concepts to sort out fact from fiction and provide some guidance.

The performance of a control loop depends upon the tuning. Specifically, the peak and integrated errors are inversely proportional to the controller gain. The peak error is not affected much by the integral time setting. However the integrated error is proportional to the integral time. Thus, a loop with good dynamics can be made to perform as poorly as a process with bad dynamics by sluggish tuning. The effect of slow sample times is hidden by large integral times or small controller gains. Thus, it is critical for any comparison, that tuning criteria be specified. In fact there is an implied deadtime as a result of the tuning of the loop as derived and discussed in Advanced Application Note 5. The tuning of the controller puts a practical limit on how fast the sample time must be for the effect to be negligible.

If a controller is tuned for maximum performance, the peak error is proportional to the loop deadtime to process time constant ratio. The integrated error is proportional to the deadtime squared. These statements are strictly true only when the process time constant is large compared to the loop deadtime. The loop dead is the sum of final element deadtime (e.g. valve pre-stroke time delay, deadband, and sticktion), process deadtime (e.g. mixing, thermal, and transportation), automation deadtime (e.g. sensor lag, transmitter damping, and sample times), and small process time constants. All of the time constants smaller than the largest time constant become effectively deadtime in the first order plus deadtime approximation used in industry. Process and automation system dynamics places an ultimate limit on loop performance. There is a corresponding ultimate limit on the sample time.

The relationships between process dynamics (e.g. total loop deadtime), controller tuning, and loop performance is detailed in the Theory section in Chapter 2 of Advanced Control Unleashed, and Appendix C in New Directions in Bioprocess Modeling and Control. All of my books and many of my articles take advantage of the fundamental understanding gained from these relationships.

The effect of sample times can be accessed in terms of practical and ultimate limits on performance. Critical loops where peak errors can cause destruction or environmental releases such as compressor surge control, furnace pressure control, exothermic reactor temperature control, and RCRA pH control, the tuning is necessarily aggressive. As a result the practical limit is much closer to the ultimate limit. For a discussion of cases where exceptionally fast sample times are needed, checkout the April 2, 2007 entry "Analog Control Holdouts."

For excellent final elements, clean sensors, and transmitter damping settings of 0.2 sec, we can suggest practical and ultimate sample times for different types of processes with typical dynamics. The ultimate limit (fastest conceivable sample time requirement) is set to be less than 1/10th of the sum of the minimum loop deadtime and minimum process time constant with some consideration as to maximum practical controller gains to reduce valve cycling and noise amplification. For any loop with a a large control valve, the minimum loop deadtime is about 1 second for an unmeasured disturbance (unless volume boosters have been added to the output of the positioner) so the ultimate limit on sample time is about 0.1 second. The practical limit reflects current tuning practices (much slower tuning to insure a smooth gradual response despite unknowns and nonlinearities). For integrating processes, the process time constant shown is the inverse of the integrating process gain (denoted by single exclamation point). The double exclamation point denotes a runaway (positive feedback) process time constant. Consultants says it is impossible to generalize but I think some guidance is helpful to the user with the realization there are always exceptions and the actual process dynamic and tuning should be identified by automated online tuners and adaptive controllers (e.g. DeltaV Insight). I didn't consider ultimate sample times slower than 60 sec. Note that slower sample times will affect the deadtime identified. A Rough Guide to DCS and Measurement (e.g. Wireless) Sample Times is offered in ProcessControlSampleTimes.pdf

For many digital devices the update is available near the beginning of the sample time (latency is negligible), which means the average deadtime from the sample time is about half the sample period. For at-line analyzers (field analyzers with automated sample systems), the result is not available until the end of the sample processing and analyzer cycle time, which translates to an average effective deadtime that is about 1.5 times the time interval between updates in the analyzer output signal. Theses deadtimes determine the minimum peak error for an unmeasured step disturbance at the input to the process.

The detrimental effect of sample time is greater than deadtime in that for continuous sources of dead time such as process transportation and mixing time delays and small process time constants, there is a continuous train of updates. For sampled data there are no intervening values. Consequently, the effects can be worse. For example, there is aliasing of oscillations where the indicated amplitude is smaller and the period is larger than actual. There can be jitter due to variations in latency and lack of synchronization of digital data that introduce variable time delays and noise for rapidly changing signals.

The PIDPLUS modification of the traditional PID developed for wireless applications helps the PID deal with the sample time from digital devices and communication, and at-line analyzers. The improvement is most dramatic for self-regulating processes but is also significant for integrating processes as seen in the tests documented in ControlStudiesPIDPLUS1.pdf. The PID-Plus algorithm also breaks the limit cycle from the resolution limit from the deadband setting for exception reporting of wireless devices because integral action is only done when there is a measurement update.

As I reflected on my career, I reaffirmed that what drives me is gaining a deeper understanding and sharing what I have learned, hopefully with a few laughs along the way. Throughout my career I sought with an open mind the knowledge and insights of the leaders in process modeling and control. I then used simulations to rapidly explore process relationships and to prototype control improvements that incorporate process understanding. The knowledge prepared me to solve tough plant control problems.

During my career at Monsanto I wrote a bunch of articles in the 1980s for InTech on my time in the plants with some humor introduced to help make the material more accessible and memorable. These articles were compiled and published in the book A Funny Thing Happened on the Way to the Control Room available for viewing as an E-book in the April 3, 2009 list of my books on this website. This is my favorite book, I didn't write much in the way of articles or books in the 1990s. I was on the road most of the time.

When I retired from Monsanto-Solutia in 2001 (sans package), I taught at Washington University. The students were great but after the course and lab was developed, it became routine. Also, I felt isolated.

I tell people I flunked retirement. I moved to Austin in September 2004 and started a second career as a part time consultant at Emerson Process Management. This gave me a chance to keep up to date with the latest new tools besides continue my exploration of process control opportunities. Plus it felt like home since Monsanto and Fisher Controls were one for most of my career.

I have been blessed with access to the best minds. In Monsanto's Engineering Technology I got to work with the leaders in process modeling and control. Some went on to distinguished chairs at prestigious universities, several were inducted into the Process Control Hall of Fame, some served as presidents of ISA and AIChE, and others left to become the principal technical resources for leading simulation companies. Here in Austin in Applied Research I get to work with the brains behind DeltaV. Plus my second career is more balanced. Except for the spike in work this year, I take a total of 4 months off each year to travel to see relatives, friends, and neat places and to write books.

Key points of my articles written in my post retirement years provide a quick overview of what I have been doing. The entries on this website in July will focus on the dozen articles I have written since retiring from my full time job. Here are the articles from 2003-2004.

"Has Your Valve Responded Lately", Control, May, 2003
"What is Your Flow Control Valve Telling You", Control Design, May 2004

Putman publications decided to do an encore publication in a second magazine. Some nomenclature typos were corrected in the reissue of the article in Control Design.

1. Deadband originates from backlash in the linkage and connections between the actuator and the plug, disc, or ball. Stick-slip comes from friction in stem packing and seals around the sealing of the plug, disc, or ball for process isolation

2. Deadband from linkage and connection backlash and stick-slip from trim and packing friction create deadtime for slowly changing controller outputs

3. Deadband will create a limit cycle in any control system where there are two integrators in series, such as a PI controller on an integrating process (e.g. level)

4. For deadband, the limit cycle amplitude is the ratio of deadband to controller gain

5. For stick-slip, the limit cycle amplitude is the product of the open loop gain and the stick-slip

6. For both deadband and stick-slip, the limit cycle period is proportional to the controller integral time and inversely related to the controller gain

7. Large actuators can have a large stroking time for a large change in signal

8. The size of the changes signal typically used to checkout control valves will not reveal the deadband or stick-slip and make all but the largest valves look good

9. A volume booster can reduce the stroking time of big actuators but has a large deadband. The booster should be put on the positioner output to quickly drive through this deadband. The booster bypass must be opened enough to prevent fast cycling from the positioner output looking into the booster's small inlet volume

10. Unstable oscillations can break out for large disturbances when the integral action in process loop becomes faster than the valve response. The integral time must be greater than the product of the valve slewing rate, disturbance size, and controller gain. (Not mentioned in the article but frequently discussed on the this website is that position read back from digital positioners and the PID dynamic reset limit option can automatically prevent the controller output from outrunning the valve)

11. Limit cycles are attenuated (filtered or washed out) by vessels or columns. The ratio of the attenuated to original amplitude is proportional to the period of the oscillation and inversely proportional to the residence time (volume/flow)

12. The control valve with the best response is a sliding stem valve with a digital positioner. If one must use a rotary valve, avoid tight shutoff and high friction packing and use a diaphragm actuator with a short shaft and splined connections between the actuator shaft and the stem of ball, disc, or plug. Make sure the stem is cast with the ball, disc, or plug to avoid another connection with backlash

Postscript: Rotary valves designed by piping manufacturers have a lot of deadband and stick-slip as discussed in the July 2009 Control Talk column "Downturn Turndown" in Control magazine.

"The Next Generation - Adaptive Control Takes a leap Forward", Chemical Processing, September, 2004

1. Nearly all controllers are detuned (backed off from maximum performance) to some degree to provide a smooth response and to deal with the inevitable changes in the process dynamics

2. Older technology adaptive controllers had these undesirable features
a. The process had to be disturbed or oscillated (e.g. patter recognition)
b. The dynamics were embedded in tuning settings
c. No real insight as to where the process has been or where it is going
d. Tuning method was fixed
e. Always playing catch up even if same situation was seen a thousand times

3. The next generation adaptive controller can
a. Normal changes in a controller's set point or manual output are used
b. The process dynamics are displayed and historized
c. From changes in the process dynamics, plant problems can be diagnosed
d. Several tuning methods are available
e. Tuning settings identified can be scheduled for preemptive action

4. "The information on changes in the process model may be directly used to monitor loop performance and to provide more intelligent diagnostics. The models can provide the dynamics for simulations and identify candidates for feedforward control and advanced control techniques. For example, loops dominated by a dead time or exhibiting disturbance models for multiple variables, are prime candidates for model predictive control. The dynamic process models in general can be used to create or adapt real time simulations for prototyping new control strategies, exploring "what if" scenarios, and training operators. Process gains that decrease or time constants that increase with feed totals are ripe for real time optimization of the run time between defrosting or cleaning and catalyst reactivation or replacement. The beauty of this route is the models and tuning settings are available from the adaptive controller for a higher level of control by a better knowledge of the topology"

"Advanced Control Smorgasbord - A Lot of Tasty Choices", Control, May, 2004

The online version is missing the following introductory sentences at the beginning of the first paragraph.

"By the time I was assigned to my first electronic control room project, some very smart engineers had already developed most of the techniques to exploit PID controllers.
Relative gain arrays and simple decoupling of the controller output were used to analyze and deal with interaction on a steady state gain basis. The outputs from PID controllers, whose process variable was a constraint variable, were sent to a signal selector to form an override control scheme to maximize or minimize a manipulated variable."

1. Previously, advanced process control (APC) required software packages at $100K a clip, separate computers, special interfaces, and consultants to do the studies and implementation. The total bill could easily approach or exceed a million dollars for a medium project, the biggest chunk being the consultant's time charges. Even a greater consideration was that the process knowledge to exploit or to just maintain the system disappeared when the consultants left the site

2. At the turn of the century, APC technologies were integrated into the basic process control system. License fees were minimal and whole cost of implementation decreased by a factor of twenty or more by the automation of the configuration, displays, testing, simulation, and tuning

3. In the time it takes to read this article, a model predictive controller or neural network could have been configured

4. Perhaps the biggest opportunity for driving the application of APC is the development of online process performance indicators

5. The key variable for process performance monitoring is the ratio of the manipulated flow to the feed flow

6. The controlled variable is best expressed and plotted as a function of the flow ratio (e.g. pH versus reagent to feed ratio, column temperature versus reflux to feed ratio, exchanger temperature versus coolant to feed ratio, and stack oxygen is versus air to fuel ratio)

7. The process efficiency is seen in difference between the actual and optimum ratio rather than in the gap between the actual and optimum controlled variable

8. A novel method has been developed to use model predictive control (MPC) to simultaneously adapt multiple first principle process model parameters

9. For closed loop process control, consider
a. PID for tight control of integrating or runaway processes
b. MPC for multivariable control, interactions, and optimization

10. For online property estimators for continuous processes, consider
a. ANN for highly nonlinear predictions with uncorrelated inputs
b. LDE for lag dominated linear predictions with uncorrelated inputs
c. PLS for steady state predictions from large number of correlated inputs

ANN is an artificial neural network, LDE is a linear dynamic estimator, and PLS is a projection to latent structures or partial least squares prediction discussed in Chapter 8 of Advanced Control Unleashed

In this final part of this series, we look at what model predictive control (MPC) can do for the following applications:

(1) Extend rangeability
(2) Improve resolution
(3) Enable preferential use of flows based on cost
(4) Send flows to multiple destinations possibly based on priorities
(5) Provide counteracting effects

MPC is the more powerful solution for an optimization problem (applications 3 and 4). MPC also offers the simultaneous manipulation of multiple flows, objective oriented tuning knobs, and manipulated variable costs that make the optimization more a science than an art. An experienced regulatory control person can make a PID do almost anything but many plants don't have that experience base. The MPC offers a solution that a person with some basic knowledge of the process and dynamics (process gain, time constant, and dead time) can understand. In my experience in teaching process control to chemical engineers at Washington University in Saint Louis, it was easier for students to understand and use an MPC because it was process oriented. On the other hand, the PID had dozens of parameters with a hundred different opinions on how to set them. If you don't believe me, check out the 484 page documentary of setting 3 of the 20+ PID parameters in the Handbook of PI and PID Controller Tuning Rules, which doesn't get into structure, options, and windup. Just think about trying to teach the PID nuances and heuristic rules well enough to turn a new employee loose on an optimization problem. I think you have a much better chance of success if the neophyte is armed with an MPC. With the disappearance of mentors and in-house technical courses all but the basics of PID control may well be lost. The manager of a process control group at a large refinery told me that he starts up with a PID but quickly moves every loop to an MPC because he doesn't have a Shinskey in his group.

We will see that MPC can also be used for applications 1 and 2 and thus cover the range of opportunities we discussed last week for valve position control (VPC). The principle drawback of the VPC solution is the lack of tuning guidance, no embedded economics, and no move suppression in conventional PID controllers to address multiple objectives (tight control of the critical process variable and the minimization of costs and variability from unnecessary movement of the expensive and large flow).

I first explored an MPC solution for application 3 for the classic case of the manipulation of fast but high cost flow and a slow but low cost flow. The solution as outlined in AdvancedApplicationNote1.pdf involves setting up one of the flows as an optimization variable. Normally one would pick minimization of the high cost flow but this made the fast flow less available for tight control in my tests because the optimization routine had a tighter than expected grip on this flow even when the penalty on error (PE) for the optimization variable was greatly reduced. The high cost and fast flow tended to ride its low limit. I achieved better load rejection performance by setting up the MPC for maximization of the low cost but slow flow. For this setup, the maximization of the low cost flow took a back seat to the tight control of the critical process variable when I reduced the PE for this optimization variable.

If the MPC allows the user to write to the relative costs manipulated variables based on the priorities of each manipulated flow, MPC offers a solution for application 4 without the addition of an optimization variable.

I next explored an MPC solution for the manipulation of a big (coarse) valve and a small (fine) valve. The solution as outlined in AdvancedApplicationNote2.pdf involves setting up the small flow as a second controlled variable. I was able to get good load rejection and set point response while minimizing the use of the big valve. I reduced the PE of controlling the small valve at its optimum position. The stick-slip limit cycle from the big valve can be broken by writing a zero to the move size limit for the big valve when the small valve is within an acceptable throttle range. This MPC solution can be extended to applications 3 and 4 by writing a set point (target) for the small valve based on costs or priorities. If it is the simple of case of trying to minimize the small flow because it is expensive, the optimum set point corresponds to a minimum throttle position that doesn't have excessive seating or sealing friction and hence stick-slip. Is the 1st or 2nd MPC solution better? The more I think about it, I think the solution outlined in the second application note offers more flexibility and is easier to set up but maybe that is because I am an old VPC guy and this MPC is a smarter way of doing VPC.

An MPC could be set up for application 5 but I am not sure whether the advantage of the built-in knowledge of the dynamics of the valves outweighs the disadvantage of the MPC inherent approach for simultaneous movement of the manipulated variables. Application 5 really demands sequential manipulation of the flows so that you are not wasting energy or raw materials. To force sequential manipulation, it appears to me you would have to have extremely high penalties for both valves being open and be able to deal with the discontinuity of the split range point with an MPC that is expecting a linear model.

With all of the advanced algorithms and smart instrumentation available today, we can sometime lose sight of what are the real limits to loop performance. While it doesn't take an Einstein to figure this out, as a former physicist, I found an interesting analogy.

Einstein's reasoning that nothing can travel faster than the speed of light lead to incredible insights and revolutionary equations. For example if you substitute the speed of light for velocity into the equation for kinetic energy, you now have the famous equation that relates mass and energy (energy is equal to mass multiplied by speed of light squared). You also end up with a unification of space and time and warping by gravitational fields.

The absolute limit to feedback control system performance is the total dead time in the loop, which is the summation of all the final element, process, measurement, I/O, and controller execution time delays. A feedback control system cannot correct for something it hasn't seen yet and hasn't been able to change yet in the process (see "Funny you should Ask a Process Control Engineer" in the Funny Thing E-book). http://www.modelingandcontrol.com/FunnyThing/page-123.asp

The fastest closed loop time constant (Lambda) possible is the deadtime. If you substitute deadtime for Lambda into the controller gain equations for Lambda tuning, you end up with the Simplified Internal Model Control and factored Ziegler Nichols equations for the highest controller gain with a relatively smooth response. This unification of equations for controller gain was documented in Appendix C of New Directions in Bioprocess Modeling and Control. This Appendix also provides the derivation that the performance achieved in terms of integrated absolute error (IAE) for an unmeasured load upset is proportional to reset time and inversely proportional to controller gain. BioprocessModelingControlBookAppendixC

The implications of this for sample delays in terms of there being an additional implied dead time for detuned controllers is explored in Advanced Application Note 5.
http://www.modelingandcontrol.com/repository/AdvancedApplicationNote005.pdf

The hype of some advanced process control (APC) algorithms may lead one to believe this limit can be violated. Many of the early APC algorithms significantly increased the loop deadtime (See "Advanced Control Algorithms- Beware of False Prophecies in the Funny Thing E-book). While model predictive control (MPC) can potentially help dead time dominant systems, the original execution time (e.g. 1 minute) of separate MPC software packages was so large their applicability was restricted to slow processes. With the advent of the MPC embedded in the DCS, the execution time can be as fast as 1 second which means MPC can be applicable to all but the fastest processes (e.g. liquid pressure control and furnace pressure loops).
http://www.modelingandcontrol.com/FunnyThing/
http://www.modelingandcontrol.com/2008/08/tipsntechniques_tnt_tuning_fur_1.html

Deadtime compensators such as the Smith Predictor can make the PID algorithm think there is no deadtime in the loop. You can get fooled as well if the PID faceplate shows the compensated PV that has the deadtime removed from the consequences of its own actions instead of the original PV. Deadtime compensators allow the user to increase the controller gain. If the deadtime compensation is perfect, the increase in controller gain can be huge. However, many sources of deadtime are variable and unknown.
http://www.modelingandcontrol.com/2007/06/deadtimes_secret_identity_part_1.html
http://www.modelingandcontrol.com/2007/06/deadtimes_secret_identity_part_2.html

For PID controllers an underestimate of deadtime can lead to instability if one goes for the gusto of ultimate performance and pushes the limit beyond the original unfactored Ziegler Nichols equation for controller gain. For deadtime compensators and model predictive control, you can also get into some oscillations for overestimates of deadtime.
http://www.modelingandcontrol.com/repository/AdvancedApplicationNote003.pdf

Finally, there is another limit to control loop performance, signal resolution. You can't control to a degree finer than the resolution of the final element, measurement, or I/O. The resolution limit of digital devices today is nearly negligible (e.g. 16 bit A/D) but some older DCS (e.g. 12 bit A/D) could cause noticeable stair-steps in the temperature response from wide range thermocouple input cards and the standard input card. It is strange to me that the standard input card for many variable speed drives still uses an 8 bit A/D that significantly restricts the resolution of the final element. Today most of the resolution limit seen in control loops is from control valves. The principal cause is stick-slip and is usually lost in the smoothing afforded by process volumes and the compression of data (except for pH and other high gain processes) unless you have on-off or isolation valves posing as control valves or choose to save money by not buying digital positioners.
www.ControlDesign.com/articles/2003/164.html
http://www.ChemicalProcessing.com/articles/2007/200.html
http://www.controlglobal.com/articles/2008/063.html

Not well recognized is that for PID control of integrating processes, valve deadband causes an implied resolution limit in the PV for a reversal of direction that is the deadband multiplied by the integrating process gain and size of the correction needed in the controller output to balance out the disturbance. Real control valves with digital positioners have a deadband that is less than twice the resolution limit.