The teachings and writings of Greg Shinskey have had a profound effect on many process automation professionals, including these authors. Greg McMillan’s comments follow; we start with thoughts from Sigifredo Nino.
The first time I met Greg Shinskey was in the winter of 1993, when I attended his “Process Control Systems Seminar” in Atlanta. Since that time, I have had the opportunity and pleasure of implementing many of his ideas throughout my successful career as process control engineer-turned-consultant. Thanks to my professional relationship with Shinskey, I have become closely familiar with his thinking, methods and creative process. His tutelage and experience continue to guide my work today, and this unique experience has given me huge insight into his approach, more so than any of his books and papers ever could have, no matter how many times I read them. Nonetheless, there is a wealth of priceless process control information in his prolific written production, which I recommend you tap into.
If Shinskey’s name and expertise is unknown to you and doesn’t ring any bells, please allow me to give a brief introduction. He wrote his first process control article under the title: “For Gas-Phase Reactors… Design for Control of Temperature,” published in Chemical Engineering on October 5, 1959. His most recent, “Killing Model-Based Control Dead Time” was published by Control in May 2013. If you tally that up, that is over half a century of invaluable contribution to process control and a confirmation of his profound interest, commitment and dedication to developing techniques and practical methods for real-life applications.
The slow disappearance of process control expertise led me to tell him something that I found saddening and disappointing, to say the least: “Future generations of process control professionals will at best have to discover Shinskey from a dusty library book.” Today, there is less and less understanding and appreciation for what process control is and what it can do for chemical facilities in nearly any industry. Furthermore, true experts in process control are an endangered species. The industry, as many others, has fallen in love with tools, software and gadgets that cannot replace a robustly designed control strategy. You cannot write the best software in the world if you do not know how to program. You can buy the latest video card with cutting edge software, have the best processor and have the best 4K display on the market, but poor coding will still result in a lackluster program. Without a deep understanding of the fundamentals, sub-optimal performance in any domain is guaranteed. There will always be the need for in-house expertise in process control, and Shinskey’s books and papers will always be a source of valuable information whenever there is a need for an ingenious and practical solution.
October 1971 marked the beginning of Shinskey’s quest of closing the gap between industry and academia. In his article, “To teach creativity” (Instrumentation Technology, p. 34), he criticized the professors at higher education institutions for their lack of contact with the real world and the consequent continuous proposition of solutions to problems that don’t exist in real life. The theory is great. Theory allows for exploration of alternatives not yet seen in the field. But theory cannot be the be-all-end-all to the success of a process engineer. It must be realistic, and applicable to real life. We don’t need to just think about it. We need to do it.
Shinskey also advocated for integrating process and control design in chemical engineering curricula. I am not sure to what extent this has happened, as it is rather common to find (very) deficient control strategies in the plant floor. There is still a way to go, and is up to us to continue in the direction Shinskey pointed.
So, what makes a good process control engineer? In his writing, “Reflections on CPC-III” (1986), Shinskey saw himself as a practitioner with the right balance between theory and practice. Someone who was able to gather knowledge from academics like E. Bristol and practitioners like C. Ryskamp, and apply the theories of the former in real life while finding the theory to support the practical application of the latter.
Because exceptional understanding of the process is prerequisite for excellent control application, Shinskey was repeatedly requested to make decisions other engineering fields were responsible for, e.g. fossil fuel combustion and compressors. He prudently declined, insisting he is an expert in process control and not an expert in all kinds of processes.
Shinskey’s deep and functional understanding of the engineering of chemical processes is simply brilliant. I have witnessed his proficiency while diagnosing and creating control designs that are well supported by solid applied engineering concepts and fundamentals and that work, despite the occasional skepticism. Shinskey’s firm belief is that the thing that makes an engineer great is his/her understanding of the basic sciences. Great engineering does not originate from the ability to use advanced tools.
Now, a tale from the field. According to Edgar Bristol, Shinskey gave the “more dramatic name” Relative Gain Array (RGA) to his “New Measure of Interaction for Multivariable Process Control.” Shinskey returned from a refinery where he was discussing a control strategy that wouldn’t work. He proceeded to calculate the RGA and came up to the result of 26, effectively saying that it wouldn’t be possible for the strategy to work. From that point on, RGA became a key element in the design of his strategies, notably for distillation control.
Derived from his process control teaching around the world, some of Shinskey’s books were translated into several languages: Process Control Systems into Japanese (1967), Romanian (1969), Italian (1971), Russian (1974), Spanish (1996) and Chinese (2014); Distillation Control into Japanese (1977); pH and pIon Control into Japanese (1977); and Energy Conservation Through Control into Russian (1981) and Japanese (1981).
In addition to Gregory McMillan’s comprehensive summary below of Shinskey’s most important contributions to process control, I must add:
- A successful process control application starts with the deep understanding of the process, followed by identification of the control objective. The lack of those two elements cannot be resolved, neither by tuning the PIDs nor by the use of a given ingenious control law, however clever we may think it is.
- The importance of acknowledging that certain processes are composed of several interacting lags, notably distillation column compositions and heat exchangers, that modeling those processes as first-order-plus-deadtime is not satisfactory, and that the controller tuning should be specific for those cases. I have personally been involved in cases where this approach for tuning distributed processes has given outstanding results: A power boiler main steam temperature control and a high-purity distillation column composition.
- Controller robustness, defined as the minimum change in the internal process parameters that can bring the loop to the stability limit (sustained oscillation), is inversely related to control loop performance, which is proportional to the product of the proportional band, integral time and the change required in the controller output to bring the process variable back to setpoint following a load disturbance. However, an improperly tuned Smith Predictor can give the loop both low robustness and low performance.
- A good load disturbance elimination capability in a PID controller is characterized by the overshot trajectory of the manipulated variable during the rejection of a load disturbance.
- Feedforward is considered a “high performance controller” by Shinskey. Minimization of the error in a control loop can be accomplished by reducing the proportional band and the integral time (subject to the stability limit), and minimization of the amount of effort, namely the change in the controller output. Feedforward reduces the amount of effort the feedback controller needs to perform by as much as 100 times compared to what the controller output would change on feedback alone.
- Valve position controller is probably one of the most valuable concepts he invented, due to its ability as a practical optimization technique.
Shinskey proposed control strategies across many industries, remarkably oil & gas; fossil power generation; pulp & paper; water treatment; heating, ventilation and air conditioning; mineral processing and food. And in doing so, he proposed strategies for several unit operations, including distillation, exothermic reactors, heat exchanging, steam turbines, steam plant management systems, combustion systems, multiple-effect evaporators, reciprocating and centrifugal compressors, refrigeration, evaporation, solids drying and, of course, pH control.
Process control is a well-established engineering field—it’s still needed today, it will be needed in the future and lest we forget, Francis Greg Shinskey has arguably made the most important contribution to our profession.
A perspective by Gregory McMillan
The one person who has done the most by far to advance the practical understanding and performance of PID control at both the basic and advanced levels is Greg Shinskey. What particularly distinguishes Shinskey is that the solutions originate from a deep and pervasive understanding of process principles and dynamics, and the largely overlooked extensive capability of PID control. Each of his books is the greatest source of knowledge on the respective subject. His articles and papers are eye openers that awaken people to what really works best. In this tribute to what I hope you will realize is the greatest mind in process control, I seek to provide recognition and synopsis of the most important knowledge conveyed in his publications, on which I have built my career.
It is impossible to summarize everything I have learned from Shinskey. I will focus on the key points, concentrating on PID with the ISA Standard Form, a parallel form where the proportional mode gain setting affects all three modes. The proportional band that is the proportional mode tuning setting in Shinskey’s works is simply 100% divided by the PID controller gain in the ISA Standard Form.
Insight into loop performance: Shinskey’s simple equation for the integrated error (IE) for a load disturbance shows the importance of the best tuning for rejecting load disturbances and reveals the effect of deadtime. The equation was developed for a load disturbance on the process input, the most common disturbance. The equation shows the IE is proportional to the integral time and is inversely proportional to the controller gain. Shinskey subsequently shows that for lag-dominant processes, the minimum integral time is proportional to the deadtime and the gain is proportional to the time constant-to-deadtime ratio, revealing that the minimum IE is proportional to the deadtime squared. The time constant mentioned here is the largest time constant in the open-loop response, and the deadtime is the total loop deadtime, which includes the deadtime in the valve, measurement and controller response as well as the process response.
Correspondingly, the maximum magnitude of error (peak error) for lag-dominant processes is inversely proportional to controller gain and hence proportional to the deadtime to time constant ratio. For deadtime-dominant processes, there is a negligible reduction in the peak error from PID action and the IE is consequently proportional to deadtime. In actual process applications, it is better to use integrated absolute error (IAE). The IE for a PID tuned for minimum IAE may be about 30% smaller than the IAE, due to some oscillations cancelling out positive and negative errors, but the metric is still very insightful and useful to provide simple but pervasive guidance.
There is significant benefit from a large process time constant that makes the response lag-dominant. Methods that have the disturbance on the process output bypassing the process time constant fail to recognize this potentially incredible effect. For example, deadtime-to-process constant ratios for well mixed vessels can be as low as 0.01, enabling a minimum peak error that is a factor of 100 less than the error predicted if the disturbance was on the process output. Also, derivative action can be used that can reduce the integral time by 50%, reducing the IE by 50%.
Understanding the effect of deadtime on tuning provides the insight that the goals for improving loop performance are to minimize deadtime in the design of the system, and then to tune the controller to minimize the IE and peak errors for the maximum deadtime including any unknowns. Of course, nonlinearities and uncertainties as to the open-loop gain (e.g., product of valve gain, process gain and measurement scale gain) and open-loop time constant (largest time constant) must be considered and the tuning set for the worst case (largest deadtime and open-loop gain, and smallest open-loop time constant). If the PID is not tuned per the actual deadtime, the control loop will do as poorly as a loop with greater deadtime where money has not been invested to reduce loop deadtime.
In fact, you can easily estimate the increase in deadtime from detuning. This fundamental understanding is not apparent in the many publications and presentations that do tests and provide conclusions on the effectiveness of the PID and the consequences of PID tuning and process variable update rate and filtering. By ignoring this equation, you can prove almost any point you want by how you tune the PID and by ignoring load response. There are more than 40 years of examples, including many papers by academics who try to show the value of their special feedback control algorithm, not realizing that Bohl and McAvoy in a 1976 landmark paper (“Linear Feedback vs. Time Optimal Control, II. The Regulator Problem,” Industrial Engineering Chemistry, Process Design and Development, vol. 15, no. 1, 1976, p. 30-33) proved that the PID provided essentially optimal single-loop control for load disturbances. Also, a simple feedforward signal incorporated into the PID output that has timing and measurement errors totalizing less than 1% can improve PID performance for measured disturbances by a factor of 100.
Comparison of tuning methods: Tests to promote the merits of a feedback control algorithm or a tuning method often fail to detail the inherent tradeoff between robustness and performance predicted by Shinskey. Not revealed is that the most aggressive but non-oscillatory settings, when the dynamics are well known and fixed, are similar to the Ziegler Nichols Reaction Curve method settings with the PID gain simply cut in half. Most studies to showcase a proposed innovation compare test results for an unmodified Ziegler Nichols Ultimate Oscillation method that is inherently oscillatory, typically showing quarter-amplitude damping. The minimum IAE has one-seventh amplitude damping.
It is important to realize almost every aggressive tuning method will have an oscillatory response and insufficient robustness. An oscillatory response is not appreciated in most processes because it increases variability indices and can cause resonance. Also, nearly all loops have gains, deadtimes and time constants that are not constant and often not well known. What you need to do is get beyond ego and be practical as to objectives of IE and peak error. Also, the load response should first be tested in tuning the PID. This can simply be done by momentarily putting the PID in manual, making a step change in the output and immediately returning the controller to automatic before the process starts to respond.
Most tuning methods are tested by making a setpoint change, and proposed improvements in tuning are recommended to reduce overshoot. Not realized is that the best load response tuning settings could be used and a lead-lag applied to the setpoint to minimize overshoot. The lag time is simply set equal to the integral time and the lead time equal to the half of the lag time to reduce the time to reach setpoint. The lead time can be decreased to minimize overshoot. A lead time of zero and a lag time equal to the reset time would correspond to PID structure of integral on error and proportional and derivative on process variable (I on E, PD on PV). A two degrees of freedom (2DOF) structure can be used instead of a setpoint lead-lag to accomplish the same objectives. Note that for secondary loops, a lag time or setpoint filter should generally not be used because it seriously slows down the ability of the primary loop to make corrections. For production rate changes where reactant flows are ratioed, setpoint filters are used to prevent the short-term upsets in mass balances from different flow loop response times. This is preferable to the practice sometimes cited of detuning the reactant flow controllers to match the slowest flow loop response, reducing the ability of these flow controllers to deal with flow disturbances such as pressure changes. To summarize, the PID should generally be tuned for best load disturbance rejection, which incidentally was the goal of Zeigler Nichols methods. An exception is a mammalian bioreactor temperature and pH response, where the cells are extremely sensitive to any overshoot, the load disturbances originating from the cells are extremely slow, and the time to reach setpoint is of no concern since the batch cycle time is a week or longer.
Important points were made by Shinskey that the Lambda tuning method as documented in the literature did not give a good IE for lag-dominant processes. Publications need to emphasize that lambda is set relative to deadtime with a minimum setting of one deadtime for a non-oscillatory minimum IE if system dynamics are well known and constant, and set as three deadtimes for much more robustness. If you want to showcase the very minimum IE despite being oscillatory and vulnerable, lambda would be set equal to about 60% of the deadtime. Also important is the use of a derivative setting that is at least 50% of the deadtime even if the secondary time constant is not identified. Of greater importance is to realize that when the time constant is more than four times the deadtime, the process is classified as near-integrating, and integrating process rules are used where lambda is now an arrest time for load disturbance rejection.
For deadtime-dominant processes, putting a low limit on the integral time to be about 50% of the deadtime, which is close to the minimum integral time in Shinskey’s settings for a high deadtime-to-time constant ratio, will prevent approaching integral-only control. These practices, used by myself and experienced practitioners knowledgeable in what Bill Bialkowski taught, need to be better documented in literature to resolve Shinskey’s concerns. The June, 2017 Control Talk column with Mark Darby, “Opening minds about controllers, part 1,” provides an independent confirmation of the capability of these recommended lambda tuning practices that are largely missing in the literature.
Effect of digital and analyzer sample time: There is an incredible amount of misinformation about the effect of sample times, whether we are talking about transmitter digital or wireless update rate (time between updates in transmitter output signal), analyzer cycle time, or PID execution rate (time between PID executions). Because tests do not take into account when the disturbance arrives in the sample period, latency and the effect of tuning on the results, as seen in Shinskey’s IE equation, erroneous conclusions have been stated, often in terms of underestimating the effects. If the disturbance arrives immediately before the sample, what the PID sees is only delayed by latency.
Shinskey clearly details the effective deadtime as one-half the sample time for zero latency. This is seen in the phase lag and by the consideration that the average time that a disturbance arrives is in the middle of the sample time. Additional deadtime from latency is the time required for calculations or analysis before a result can be sent. For analyzers where the result is available at the end of the cycle time, the latency is the cycle time, giving a total deadtime that is 1.5 times the cycle time (0.5 x cycle time from sampling and 1 x cycle time from analysis time). Additional deadtime comes from the sample transportation delay, equal to sample system volume between sample point and analyzer divided by sample flow. Hence, remotely located analyzers in particular need a high sample flow rate achieved by a high sample recycle flow.
The increase in deadtime will decrease the maximum controller gain and increase the minimum integral time, and thus result in a two-fold increase in IE. There are additional effects in terms of PID execution rate. The effectiveness of derivative action starts to significantly decrease if the PID execution rate is greater than one-tenth the rate time, which subsequently requires a larger integral time. Also, the effective integral time must be incremented by the PID execution rate. A small signal filter has a similar effect. Consequently, the numerator of the IE equation is the integral time plus the PID execution rate and signal filter time.
External-reset feedback: The positive feedback implementation of integral action offers the ability to use external reset feedback of a signal that is feedback of what is being manipulated. The most common use is in cascade control where the external-reset feedback to the primary PID is the process variable of the secondary PID. This feedback passes through the filter whose time constant is the integral time. This structure inherently prevents the integral action from changing the secondary PID setpoint faster than the loop can respond. This is critical for secondary loops that are not more than five times faster than the primary loop. Slow secondary loops causes a devious problem in that for small setpoint changes or small disturbances there does not appear to be a problem. Thus, during testing and tuning that typically involves small changes the user may be unaware of the problem. The burst of oscillations that occurs for large disturbances and setpoint changes is confusing. For more on this easy solution for cascade control see “The power of external-reset feedback” (Control, May 2006, p. 53.
External-reset feedback (ERF) enables an incredible spectrum of opportunities. ERF can prevent the PID output from changing faster than a control valve can respond. ERF can stop a limit cycle from valve stiction resolution limit in loops without process integrator and a limit cycle from backlash deadband in all loops. It can also enable the use of a much higher gain resulting in tighter control in loops with poor positioner sensitivity or large stroking times. However, there is an offset when a limit cycle stops and additional deadtime associated with the PID output working through the resolution or sensitivity limit and deadband. Of course, the better solution is a control valve, positioner and actuator assembly with best resolution and sensitivity and least deadband and stroking time. All this and much more is detailed in the May 2016 article, “How to specify valves and positioners that do not compromise control.” ERF can provide fast open and slow closing of a surge valve without retuning by putting rate limits on the valve signal. ERF can provide a gradual optimization and fast getaway by a valve position controller (VPC) using external reset feedback of the process variable and intelligent up and down setpoint rate limits of the PID being optimized. For the only known block diagram showing the positive feedback implementation of integral action, see slide 42 of the online presentation, “ISA Mentor Program WebEx PID Options and Solutions.“ This incredible capability of the PID and much more is discussed in the Control Talk blogs “PID Options and Solutions - Part 1,” posted 9/28/2016, and “PID Options and Solutions - Parts 2 and 3,” posted 11/1/2016. The presentations review the equations for IE and peak error in terms of tuning settings (practical limits to performance) and in terms of deadtime and time constant (ultimate limits to performance). The blogs provide summaries and links to the three-part ISA Mentor Program WebEx presentations.
ERF is also the basis of a PID option that will only update the integral contribution when the process variable changes by an appreciable amount. This prevents oscillations from the periodic updates of wireless transmitters and analyzers. If the wireless update rate or analyzer cycle time and analysis time cause a total loop deadtime greater than the 63% response time of the process, the PID for self-regulating processes does not need to be retuned for the additional deadtime, and in fact, the PID integral time can be decreased and PID gain increased to be the inverse of the open-loop gain. This opportunity is introduced in the InTech July-August 2010 article, “Wireless: Overcoming challenges in PID control & analyzer applications.” Since then, it has been realized that this technique can even deal with incredibly large and variable time intervals between lab analyzer results. However, the user must recognize that the additional deadtime from periodic updates will decrease the ability to handle load disturbances. The additional deadtime should be added to the IE numerator.
Deadtime compensation: Simple insertion of a deadtime block in the external reset feedback to give what Shinskey labels as a PID+td provides deadtime compensation better than a Smith Predictor that requires accurate knowledge of the open-loop gain and open-loop time constant. The PID+td gain can be increased and integral time greatly decreased. The deadtime (td) is added to the integral time in the IE numerator to show the benefit still deteriorates as the deadtime increases for load disturbances. Contrary to common knowledge, a deadtime compensator does not eliminate deadtime in the loop and the improvement in performance is greatest for lag-dominant rather than deadtime-dominant processes.
Also, the robustness is decreased as a tradeoff to the increase in performance. Sensitivity to a decrease in deadtime is introduced that is not seen in a PID. The PID+td sensitivity to a decrease in deadtime may be five times greater than the sensitivity to an increase in deadtime in terms of the start of oscillations. Fortunately, the deadtime block can easily be updated for transportation delays. The setting of deadtime block might benefit from a slight underestimate of the total loop deadtime, which is the opposite of what we do in the tuning of a PID without deadtime compensation or in the dynamic compensation of a feedforward signal. The PID+td also performs much better than the model predictive control (MPC) often cited as necessary for deadtime-dominant loops. For more on misconceptions, see “Common automation myths debunked” (Control, April 2017, p. 42).
pH control: Shinskey showed how to model the titration curve and hence the nonlinearity of pH by a simple charge balance that I found can be solved by a simple interval-halving search technique to find the pH that provides a zero charge balance. Based on the guidance offered by Shinskey in his book on pH, I was able to extend the charge balance to acids or bases with three dissociation constants and conjugate salts. The equations presented in numerous books on ionic equilibrium do not have this insight, which leads to specific complex solutions that fill the book. The extension and importance of the charge balance is detailed in “Improve pH control” (Chemical Processing, March 2016, p. 33).
Shinskey also pointed out how the open-loop gain identified from a step test was the line between the start point and end point on a valve or process nonlinear curve. Consequently, the gain is only the slope of the valve or process curve for small changes. For this and many other reasons, signal characterization is beneficial to eliminate the change in tuning and control loop performance for changes in the size of the step for tests and disturbances, as discussed in the 10/25/2015 Control Talk blog, “The unexpected benefits of signal characterizers.”
Another key understanding given by Shinskey is that contrary to what is often shown in books and courses on control theory, the temperature, composition and pH response of a process is not the result of the sum or difference of process inputs, but is the ratio of flows, typically a manipulated flow divided by a process feed flow. The result is an additional nonlinearity when the process variable is plotted versus the manipulated flow. The nonlinearity is inversely proportional to the feed flow. Thus, for temperature loops operating at a low heating or cooling demand, there is an increase in process gain, and if the jacket does not have a constant recirculation flow, there is an increase in the jacket deadtime from the increase in jacket transportation delay. Oscillations can break out at low production rates or early in the batch for continuous or batch operations, respectively. This insight is also the key understanding of the value of flow ratio control, where the temperature, composition or pH controller corrects the ratio often by a simple bias often not considered in MPC. To see the pervasiveness and importance of these ratios, check out the 5/30/2015 Control Talk blog, “Essential feedforward control and operator interface tips” and the 1/31/2017 Control Talk blog, “Uncommon knowledge for achieving the best feedforward and ratio control.”
I cannot possibly convey the total and incredible value of what Shinskey did, as seen in his articles, books and papers. No other person has come close to providing the core, first-principle knowledge, often the opposite of what is taught or read about in control theory. Unfortunately, we have the bizarre situation where publishers do not understand anything about the true and unique value of the written knowledge. Consequently, all of Shinskey’s books are out of print by the original publishers. You can get a copy of Process Control Systems Fourth Edition from the Schneider Electric education department by contacting the training manager, Bob Schwarz, at firstname.lastname@example.org.