Stan: A process control specialist may be expected to walk into a control room and within a few hours develop wondrous solution to a problem whose real cause not apparent. I had to do this in my career repeatedly on startups when time was precious and patience was slim.
Greg: I also had this eye opening experience that got more technically complicated when I moved into Engineering Technology and was tackling problems that plagued production units for years. Process control improvement comes down to putting process knowledge in the control system. How you get this information from the plant documentation, and operations, process engineering, and research is a problem it itself. Documentation is more focused on details to build and operate the plant than how to control the plant. Research reports may stay in the research department. Chemists, operators, and chemical engineers don't generally understand dynamics or the basics of process control. What they can offer is very important in turns of what does and doesn't work and process relationships. I have found Tip #51 Seek Conversations with Knowledgeable People (7/13/2012) post on the ISA Interchange Site to be essential for success. To get another perspective, we asked James Beall, our guide in our recent December and January columns, how he approaches the challenge of intertwined problems that have evaded solution.
James: I have had to tackle a double effect evaporator system with downstream centrifuges and a lab analysis just once a day, over 20 pieces of equipment, multiple sources and destinations, and a recycle stream of unknown composition. I drew a process sketch that combined the automation system components from the Piping and Instrument Diagram (P&ID) and key streams from the Process Flow Diagram (PFD). Every week I asked more questions and used the answers to improve the sketch.
Greg: There is a non-self-regulating effect of recycle streams where impurities and inerts can accumulate. I know a plant whose capacity did not achieve the promises of debottlenecking projects for decades because a long-forgotten impurity was building up. A research report from over 50 years ago buried in the archives provided the revelation that could have saved millions of dollars and decades of frustration and speculation. There must be an analysis of all components in the recycle stream and an intelligent purge of impurities and inerts. Recycle streams can cause your thinking and the control system to go in circles. Feedback control adds self-regulation. A 1993 series of papers by Bjorn Tyrus and Bill Luyben "Dynamics and Control of Recycle Streams" concluded there must be a flow control of the recycle stream somewhere in the recycle system to prevent a "snowballing" effect. Also, the makeup of a reactant recovered from the recycle stream in a distillation column must fed back to the reactor by tight level control in the distillate receiver. A solution often needs to address the short term effects by feedback control and rely on optimization levels to find the best recycle flow. A paper Model Predictive Control for Process Improvement by ISA Mentor Program protégé Flavio Briguente in the OSU Automation Society Newsletter Spring 2013 showed how an innovative application of advanced process control greatly reduced the pH variability in a reactor with a recycle stream.
Stan: James, what do you think will be the key to a solution for your complex recycle system?
James: I will determine basic models of the process responses to better understand process relationships and operation. I did some initial step tests and found that steady state was reached in the composition response within 2 to 3 hours, instead of the 24 hours expected by operations. The temperatures were also lining out in the same time frame. Part of the misunderstanding may be caused by the once per day sample analysis. During the tests we took extra samples. Process engineers are not particularly in tune with dynamics. Operations may have a more realistic view.
Stan: Since operators spend 12 hours whereas process engineers an occasional few hours with the process, I would expect the operators would have a better sense of time. However, people in general have problems with anticipating the effect of dead time. I found this in retirement.
Greg: Operator graphics put too much focus on digital numbers and an obsession with values after the decimal place of no meaningful value. Operator displays showing the future trajectories of the controlled and manipulated variables is one of the advantages of model predictive control. I think all process variables should have an intelligent trend chart time span to show the trajectory of the past and into the future as noted in the Control Talk Blogs "Checklist for Loop Analysis by Trend Charts" (7/26/3012) and "Future PV Values are the Future" (6/29/2012).
Stan: James, how do you intend to use the knowledge of dynamics gained?
James: I plan to use the step response models in a simulator to help develop and test my new MPC control scheme. This will allow process engineers and operators to view and confirm the desired response from the MPC control scheme.
Greg: I think there is a great opportunity for step response models gained from auto tuner and adaptive tuner software to provide models. You can save a lot of test time by using a near-integrator approximation for slow continuous processes. You get the dead time and an integrating process gain, which is the maximum ramp rate divided by the change in the manipulated variable. The integrating process gain can in turn be converted to a time constant. The steady state gain can be approximated as the difference in controlled variable divided by the difference in manipulated variable for two different operating points. The time constant is then this steady state gain divided by the integrating process gain. You may even want to use a near-integrator approximation for a true integrator to prevent a model from ramping to a limit before all of the controls are nailed down. All calculations are done in percent of scale of the controlled and manipulated variables.
Stan: I can't overemphasize the importance of including the total loop dead time besides the process time constant. Tieback models that focus on a steady state gain do not give any sense of dynamics. The controller gain depends on all three terms.
Greg: Without dead time, I would be out of job. The controller could immediately see and correct for any change whether a load or a setpoint within the limits of noise. The importance of dead time is emphasized in Tip #70 Minimize Dead Time (7/27/2012) post on the ISA Interchange Site. Of course just the act of putting the controller and a model in a virtual plant creates dead time from the controller, model, and interface execution times.
Stan: Often the models need to be sped up for analysis and operator training. In this case, both the dead time and time constant should be shortened by the same factor so the controller gain is the same. The reset time should be shortened by the same factor as well, since the reset time is proportional to the time constant for the Lambda self-regulating process and is proportional to the dead time for the Lambda integrating process tuning method.
Greg: If the process time constant is greater than 4 times the dead time, the process should be treated as near-integrating and lambda integrating process tuning rules used as noted in my Control Talk blog series "Processes with no Steady-State in PID Time Frame" (3/02/2013). A question remains. Before you can do the tests what is the best guess for step response model dynamics? For new plants, the steady state gain can be found from process flow diagrams, instrument scale ranges, and valve and pump sizing. Some rules of thumb can get you started for the dead time and time constant. The process dead time for the composition response from reactant or reagent addition is on the order of 1 minute for vessels and 10 minutes when evaporation or distillation is involved. For small reagent flows, the dead time estimate can be out the window due to injection delay from dip tubes. The dead time for temperature response is larger due to thermal lags. If there is an analyzer, the model dead time is increased by the sample transportation delay and 150% of the analyzer cycle time plus 100% of the multiplex time. A ballpark number for the composition process time constant for reagent and reactant addition is 20 times the process dead time. If evaporation or distillation is involved the process time constant is about 5 times the process dead time. Once the plant is running, the initial dynamics of the model can be improved. An adaptive controller can be readily setup to adjust the dynamics of the step response model online to better match the operating conditions of the plant. After a setpoint or manual output change, the adaptive controller can adjust the integrating process gain to make the ramp rate of the model PV match the ramp rate of the process PV. For self-regulating processes, the adaptive controller can adjust the model steady state gain to get the model PV to match the actual PV. For near-integrating processes, the process time constant can be estimated from the steady state and integrating process gains.
A rap song "P.I. Diddy" composed by Ken Lane while attending my short course "Effective Use of PID Controllers" for the ISA New Orleans Section:
by Ken Lane
I've got the wrong kind of action
… The integral action
The insidious effect is gonna leave me in traction
Pulled from all angles,
Stick-slip, now I dangle
Overshoot … now I'm tangled
Tangled in a structure
Steps ahead to my destruction
Overshot the landing
Deviated from my function
Full throttle, bang-bang
There is dead time in my block
Loop it back around
And put your Shinskey on my clock