The following technical discussion is part of an occasional series showcasing the ISA Mentor Program, authored by Greg McMillan, industry consultant, author of numerous process control books, 2010 ISA Life Achievement Award recipient and retired Senior Fellow from Solutia Inc. (now Eastman Chemical). Greg will be posting questions and responses from the ISA Mentor Program, with contributions from program participants.
Luis Navas is an ISA Certified Automation Professional and electronic engineer with more than 11 years of experience in process control systems, industrial instrumentation and safety instrumented systems. Luis’ questions on evaporator control are important to improve evaporator concentration control and minimize steam consumption
Luis Navas’ Questions
Which criteria should I follow to define the final control strategy with model predictive control (MPC) in an existing PID strategy? Only one MPC for all existing PIDs? Or may be 1MPC + 1PID or 1MPC + 2 PIDs? What are the criteria to make the correct decision? What is the step by step procedure to deploy the advanced control in the real process in the safest way? Which are your hints, tips, advice and experiences regarding MPC implementations?
PID control of a double-effect evaporator
Greg McMillan’s Initial Answer
In general you try to include all of the controlled variables (CV), manipulated variables (MV), disturbance variables (DC), and constraint variables (QC) in the same MPC unless the equipment are not related, there is a great difference in time horizons or there is a cascade control opportunity like we see with Kiln MPC control where a slower MPC with more important controlled variables send setpoints to a secondary MPC for faster controlled variables. For your evaporator control, this does not appear to be the case.
We first discuss advanced PID control and its common limitations before moving into a MPC.
For optimization, a PID valve position controller could maximize production rate by pushing the steam valve to its furthest effective throttle position. So far as increasing efficiency in terms of minimizing steam use, this would be generally be achieved by tight concentration control that allows you to operate closer to minimum concentration spec. The level and concentration response would be true and near integrating. In both cases, PID integrating process tuning rules should be used. Do not decrease the PID gain computed by these rules without proportionally increasing the PID reset time. The product of the PID gain and reset time must be greater than the inverse of the integrating process gain to prevent slow rolling oscillations, a very common problem. Often the reset time is two or more orders of magnitude too small because user decreased the PID gain due to noise or thinking oscillations are caused by too high a PID gain.
I don’t see constraint control for a simple evaporator but if there were constraints, an override controller would be setup for each. However, only one constraint would be effectively governing operation at a given time via signal selection. Also, the proper tuning of override controllers and valve position controllers is not well known. Furthermore, the identification of dynamics for feedback and particularly feedforward control typically requires the expertise by a specialist. Often comparisons are done showing how much better Model Predictive Control is than PID control without good identification and tuning of feedback and feedforward control parameters.
While optimization limitations and typical errors in identification and tuning push your case toward the use of MPC, here are the best practices for PID control of evaporators.
- Measure product concentration by a Coriolis meter on evaporator system discharge.
- Control product concentration by manipulation of the heat input to product flow ratio.
- Use evaporator level measurements with an excellent sensitivity and signal noise ratio.
- When possible, use radar instead of capillary systems to reduce level noise, drift, and lag
- Control product concentration by changing heat input to feed rate ratio. If production rate is set by discharge flow, use PID to manipulate heat input. If production rate is set by heat input, use PID to manipulate product flow rate.
- Use near integrator rules maximizing rate action in PID concentration controller tuning.
- Use a flow feedforward of product flow rate to set feed rate to minimize the response time for production rate or product concentration control.
- For feed concentration disturbances, use feedforward to correct the heat input based on feed solids concentration computed from density measured by a feed Coriolis meter.
- The actual heat to feed ratio must be displayed and manual adjustment of desired ratio be provided to operations for startup and abnormal operation.
- To provide faster concentration control for small disturbances, use a PD controller to manipulate a small bypass flow whose bias is about 50% of maximum bypass flow.
The use of model predictive control software often does a good job of identifying the dynamics and automatically incorporating them into the controller. Also, it can simultaneously handle multiple constraints with predictive capability as to violation of constraints. Furthermore, a linear program or other optimizer built into MPC can find and achieve the optimum intersection of the minimum and maximum values of controlled, constraint, and manipulated variables plotted on a common axis of the manipulated variables.
I have asked for more detailed advice on MPC by Mark Darby, a great new resource, who wrote the MPC Sections for the McGraw-Hill Handbook Hunter and I just finished.
Mark Darby’s Initial Answer
It is normally best to keep PID controls in place for basic regulatory control if they perform well, which may require re-tuning or reconfiguration of the strategy. Your case is getting into advanced control and optimization where the advantage shifts to MPC. Multiple interactions and measured disturbances are best done by MPC compared to PID decoupling and feedforward control. First principle models should be used to compute smarter disturbance variables such as solids feed flow rather than separate feed flow and feed concentration disturbance variables. Override control and valve position control schemes are better handled by MPC. More general optimization is also better done with an MPC. Remember to include PID outputs to valves as constraint variables if they can saturate in normal operation. If a valve is operated close to a limit (e.g., 5% or 95%), it may be better to have the MPC manipulate the valve signal directly using signal characterization as needed using installed flow characteristic to linearize response.
Join the ISA Mentor Program
The ISA Mentor Program enables young professionals to access the wisdom and expertise of seasoned ISA members, and offers veteran ISA professionals the chance to share their wisdom and make a difference in someone’s career. Click this link to learn more about how you can join the ISA Mentor Program.
Here are some MPC best practices from Process/Industrial Instruments and Controls Handbook Sixth Edition, by Gregory K. McMillan and Hunter Vegas (co-editors), and scheduled to be published in early 2019. This sixth edition is revolutionary in having nearly 50 industry experts provide a focus on the steps needed for all aspects to achieve a successful automation project to maximize the return on investment.
MPC Project Best Practices
- Project team members should include not only control engineers, but also process engineers and operations personnel.
- First level support of MPC requires staff with knowledge of both the MPC and the process. Site staff needs to have sufficient understanding to troubleshoot and answer the questions of operations. Larger companies often have central teams for second level support and to participate in projects.
- Even in companies with experienced teams, it is not unusual to use outside MPC consultants. The right level of usage of outside consultants is rarely 0% or 100%.
- It may be tempting to avoid the benefit estimation and/or post audit, especially when a company has previous successful history with MPC. But doing so carries a risk. New management may not have experience or understand the value of MPC, leading to the inevitable question: “What is MPC doing for me today?”
- The other temptation is to forgo needed instrumentation or hardware repairs and proceed directly with an MPC project, arguing that MPC can compensate for such deficiencies. This carries the risk of not meeting expectations and MPC getting a bad reputation, which will be difficult to erase.
- Regular reporting of relevant KPIs and benefits is seen as the best way of keeping the organization in the know and motivating additional MPC applications.
MPC Design Best Practices
- Develop a functional design with input from operations, process engineering, economics staff, and instrument techs. Update the design as the project progresses, and after the project is completed to reflect the as built MPC.
- Not all MPC variables must be determined up front in the project. Most important is identifying the MVs. The final selection of CVs and DVs can be made after plant testing, assuming data for these variables was collected.
- The use of a dynamic simulation can be useful for testing a new regulatory control strategy. It can also be used to test and demonstrate an MPC, which can be quite illustrative and educational, particularly if MPC is being applied for the first time in a facility.
- If filtering of a CV or DV for MPC is required, it needs to be done at the DCS or PLC level. The faster scan times allow effective filtering (usually on the order of seconds) without significantly affecting the underlying dynamics of the signal. In addition, filters associated with the PVs of PID loops should be reviewed to ensure excessive filtering is not being used to mask other problems.
- The use of a steady-state or dynamic simulation can be useful for determining thermo-physical equation parameters for PID calculated control variables (e.g., duty or PCT) and MPC CVs, estimating process gains, and evaluating possible inferential predictors.
- With most MPC products, adding MVs, CVs, and DVs is a straightforward task once models are identified. This allows starting with a smaller MPC on one part of the unit, and later increasing the scope as experience and confidence is gained.
- Inferential models can be developed ahead of the plant test, which allows the model to be evaluated and adjustments made. For data driven regression based inferentials, one needs to have at least confirmed that measurements exist that correlate with the analyzed valve. Final determination of model inputs can be made during the modeling phase.
- A challenge with lab-based inferentials is accurately knowing when a lab sample is collected. A technique for automating this is to install a thermocouple directly in the line of the sample point. A spike in the temperature measurement is used to detect a sample collection.
- When implementing a steady-state inferential model online, it is often useful to filter inputs to the calculation to remove phantom effects such as overshoot or inverse response.
MPC Model Development Best Practices
- A test plan should be developed with operations and process engineering. It will need to be flexible to accommodate the needs of the modeling as well as operational issues that may arise.
- Data collected for model identification should not use data compression. A separate data collection is recommended to minimize the likelihood of latency effects such as PVs exhibiting changes before SPs.
- The data collection should include all pertinent tags for the units being tested. This can allows integrity checks to be made, and models to be identified for new CVs and DVs that may be added later to the MPC.
- Model identification runs should be done frequently, typically at least once per day. This allows the testing to be modified to emphasize MV-CV models that are insufficiently identified.
- The plant test is an opportunity to answer operational or process questions of which there are differing opinions, such as the effect of a recycle on recovery or on a constraint. This can help to develop consensus on a new strategy.
- MPC products that include automatic closed-loop should provide the necessary logic to change step sizes, bring MVs in and out of test mode, and displays to follow the testing.
- Lab sample collection for inferential model: include multiple samples collected at same time to assess sample container differences (reproducibility) and lab repeatability. When multiple sample containers are used, record the container used for each sample. Coordinate lab sample with collection personnel and record time samples are collected from the process.
- The MPC identification package should automatically handle the required scaling of MVs and CVs and differencing and/or de-trending.
- The ability to slice or section out bad data is a necessary feature. Note that each section of data that is excluded requires a re-initialization of the identification algorithm for the next section of good data.
- A useful technique for deciding on which MVs and DVs are significant in a model, and should therefore be included, is to compare their contribution to the CV response based on the average move sizes made during the plant test.
- Model assessment tools to guide model quality assessments are desirable. These may be error bounds on the step responses or other techniques to grade the model. A common technique for assessing model errors are bode error plots which express errors as a function of frequency. They can be useful for modifying the test to improve certain aspects of the model (e.g., reducing errors at low or high frequencies.
- Features to assist in the development of nonlinear transformation are desirable. Ideally, the necessary pre- and post-controller calculations to support transformations are a standard option in the MPC.
- Features that help document the various model runs and the construction of the final MPC model is a desirable feature.
- Even if an MPC includes online options for removing weak degrees of freedom, it is recommended that known consistency relationships be imposed as part of model identification.
About the Author
Gregory K. McMillan, CAP, is a retired Senior Fellow from Solutia/Monsanto where he worked in engineering technology on process control improvement. Greg was also an affiliate professor for Washington University in Saint Louis. Greg is an ISA Fellow and received the ISA Kermit Fischer Environmental Award for pH control in 1991, the Control magazine Engineer of the Year award for the process industry in 1994, was inducted into the Control magazine Process Automation Hall of Fame in 2001, was honored by InTech magazine in 2003 as one of the most influential innovators in automation, and received the ISA Life Achievement Award in 2010. Greg is the author of numerous books on process control, including Advances in Reactor Measurement and Control and Essentials of Modern Measurements and Final Elements in the Process Industry. Greg has been the monthly “Control Talk” columnist for Control magazine since 2002. Presently, Greg is a part time modeling and control consultant in Technology for Process Simulation for Emerson Automation Solutions specializing in the use of the virtual plant for exploring new opportunities. He spends most of his time writing, teaching and leading the ISA Mentor Program he founded in 2011.
Connect with Greg: