Gregory K. McMillan ( http://www.modelingandcontrol.com ) presents the process of assessing opportunities to apply advanced process control (APC), their potential benefits, and exposes some common
Gregory K. McMillan ( http://www.modelingandcontrol.com ) presents the process of assessing opportunities to apply advanced process control (APC), their potential benefits, and exposes some common myths.
Greg is a retired Senior Fellow from Solutia Inc. During his 33 year career with Monsanto Company and its spin off Solutia Inc, he specialized in modeling and control. Greg 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 “Process Automation Hall of Fame” in 2001, and honored by InTech Magazine in 2003 as one of the most influential innovators in automation. Greg has written a book a year for the last 20 years whether he needed to or not. About half are humorous (the ones with cartoons and top ten lists). Presently Greg contracts via CDI Process and Industrial as a principal consultant in DeltaV Applied R&D at Emerson Process Management in Austin Texas. For more info visit:
This presentation offers examples and a methodology for the identification of the benefits and solutions for advanced control
Pyramid of Technologies
Opportunity Assessment Methodology
Opportunity Assessment Questions
Model Predictive Control Primer
Example of Transition from Conventional to Advanced Control
MPC Valve Rangeability and Sensitivity Solution
MPC Maximization of Low Cost Feed Example
MPC Procedure and Rules of Thumb
What we Need
Columns and Articles in Control Magazine
06/06/09 TS is tactical scheduler, RTO is real time optimizer, LP is linear program, QP is quadratic program Pyramid of Technologies APC is in any technology that integrates process knowledge Foundation must be large and solid enough to support upper levels. Effort and performance of upper technologies is highly dependent on the integrity and scope of the foundation (type and sensitivity of measurements and valves and tuning of loops) The greatest success has been Achieved when the technology closed the loop (automatically corrected the process without operator intervention) Basic Process Control System Loop Performance Monitoring System Process Performance Monitoring System Abnormal Situation Management System Auto Tuning (On-Demand and On-line Adaptive Loop Tuning) Fuzzy Logic Property Estimators Model Predictive Control Ramper or Pusher LP/QP RTO TS
Loops Behaving Badly 06/06/09 1 E i = ------------ T i E o K o K c where: E i = integrated error (% seconds) E o = open loop error from a load disturbance (%) K c = controller gain K o = open loop gain (also known as process gain) (%/%) T i = controller reset time (seconds) (open loop means controller is in manual) A poorly tuned loop will behave as badly as a loop with lousy dynamics (e.g. excessive dead time)! Tune the loops before, during, and after any process control improvements You may not want to minimize the integrated error if the controller output upsets other loops. For surge tank and column distillate receiver level loops you want to minimize and maximize the transfer of variability from level to the manipulated flow, respectively.
Unification of Controller Tuning Settings 06/06/09 Where: K c = controller gain K o = open loop gain (also known as process gain) (%/%) 1 self-regulating process time constant (sec) max maximum total loop dead time (sec) All of the major tuning methods (e.g. Ziegler-Nichols ultimate oscillation and reaction curve, Simplified Internal Model Control, and Lambda) reduce to the following form for the maximum useable controller gain
Less blending, scrap, and rework or higher price for higher grade *
Lower utility costs (energy minimization)
Higher production rate (feed maximization)
Increased on stream time (fewer shutdowns)
Reduced maintenance (less stress on equipment)
Safer Operation (fewer shutdowns and less stress on equipment)
06/06/09 * The benefits for improved yield and less scrap or rework can be taken as an increase in capacity or a reduction in raw materials
Opportunity Sizing and Assessment (2% of COGs on the average in 50 processes)
Do a thorough opportunity sizing (OS) before the opportunity assessment using cost sheets, product prices, historian trends, business plans, research reports, technical studies, and simulations to establish actual, practical, and theoretical performance (e.g. yield, capacity) with operations and technology
Use plant process engineers to go through process, identify constraints, and offer ideas on opportunities to reduce gaps identified in OS to see and work way out of the current process box
Avoid temptation of canned solution or for consultants to come to conclusions before the plant people thoroughly discuss peculiarities and special problems. Get knowledgeable people to speak first and ask questions – hold off on solutions but offer concepts that people can use to generate solutions and be a good listener
Use historian to find loops in manual, limit cycles, slow or oscillatory set point and load responses, and controller outputs running near limits
Opportunity Sizing and Assessment (2% of COGs on the average in 50 processes)
Look for opportunities to infer compositions from fast lower maintenance measurements such as density, viscosity, mass spectrometers, microwave, and nuclear magnetic resonance
Seek applications of accurate mass flow ratios for material balance knowledge and control
Ask what would happen if a set point or operating mode is changed
Pick control technologies to address opportunities and give relative estimate of implementation cost and time (e.g. high, medium, low) and per cent of gap addressed
Ask plant process engineers to estimate percentage of gap addressed by each solution
Take advantage of momentum and group enthusiasm – start on “quick hits” immediately and set definitive schedule and assignments for others (avoid inertia of waiting for quote or study)
All the people you need to get started should be in the meeting, otherwise you have the wrong people
Are there limits to operating values that are important for product quality, efficiency, or for environmental, personnel, and property protection?
Can these limits be measured online, analyzed in the lab, or calculated?
Has there been down time attributed to violations of these limits? This can show up as an increase in the maintenance cost or number of failures of equipment, a decrease in the run time between catalyst replacement or regeneration, a decrease in the run time between clean outs or defrosts from a faster rate of fouling or coating, and trips from interlocks for personnel and property protection.
Has product been downgraded, recycled, returned, or pitched as the result of excursions beyond these limits?
Would operation closer to a limit significantly decrease utility or raw material use or increase production rate?
Have there been any environmental violations or near misses?
Does the operator pick set points to keep operating points away from limits?
Is there a batch operation with a feed rate that depends upon a process variable where the batch time could be reduced by increase in a feed rate by operating closer to process or equipment limits?
Auto tuners can compute controller tuning settings with an accuracy of more than one digit.
Act surprised when unmeasured disturbances, load changes, valve stick-slip, and noise cause each result to be different. Look forward to the opportunity to play bingo with the second digit.
You can just dump all your historical data into an artificial neural network and get wonderful results.
Forget about the same stuff that cause auto tuners to have problems. Use variables drawing straight lines because anything that smooth or well controlled must be important. Use the controlled variables (process variables) instead of the manipulated variables (controller outputs). Don’t try to avoid extraneous inputs or identification of the control algorithm instead of the process. If you want to purse a career in data processing, use every input you can find.
06/06/09 There were a lot of myths heard in opportunity assessment Here is the short list of the more humorous ones
Models can predict a process variable that is not measured in the field or lab.
Great way to spur creativity in training an ANN, developing a PLS model, and validating a first principal model plus it has the added bonus of the model never being wrong. Wait till your customers figure out something is wrong with the composition of your product. Discount as hearsay any suggestions that even the best models need periodic correction
Models don’t need to include process and measurement time delays
After all the following time honored traditions can’t all be misleading
Professors teach students to think steady state
Books on process control focus on continuous processes
Statisticians analyze snapshots of data
Operations want instantaneous results
Engineers think the temperatures, compositions and flows in the plant are constant and match what are defined on the Process Flow Diagram (PFD)
Process control does not apply to batch processes.
Use that time tested fixed sequence. After all, that batch cycle time is a tradition and the golden batch sure looks shiny.
Positioners should not be used on fast loops.
What was true for the good old days of pneumatic positioners and analog controllers must still be true today. Surely, digital positioners with tuning settings and digital control system scan times can’t make the original theoretical concerns less important than the practical issues of real valves. If you would rather believe the controller outputs are the actual valve positions, and just want valve problems to slip by, save some bucks on your project and only put positioners on slow loops. Just don’t stick around for start up.
The effect of a properly designed PRBS test averages out
Relay tuning methods may provide tighter control than loop
Software can automatically identify models from the normal set point changes made during startup and operation
To reduce variability in process outputs (temperatures and compositions), keep all the process inputs (flows) constant.
Keep believing that you can fix both the process inputs and outputs and don’t accept the notion that process control must transfer variability from process outputs to process inputs to compensate for disturbances. Just make the variability disappear.
Use process outputs for principal component analysis, neural network and partial least squares models regardless of control system design
Use the same process outputs (e.g. composition, temperature) after the loop is closed and variability has been transferred to process inputs (e.g. flows)
When Process Knowledge is Missing in Action 06/06/09 2-Sigma 2-Sigma RCAS Set Point LOCAL Set Point 2-Sigma 2-Sigma Upper Limit PV distribution for original control PV distribution for improved control Extra margin when “ war stories” or mythology rules value Benefits are not realized until the set point is moved! (may get benefits by better set point based on process knowledge even if variability has not been reduced) Good engineers can draw straight lines Great engineers can move straight lines
You need an advanced degree to do advanced control.
Not so anymore. New software packages used to form a virtual plant automate much of the expertise needed and eliminate the need for special interfaces. The user can now focus mostly on the application and the goal.
Models only apply to continuous processes.
Since most of the applications are in the continuous industry, this is a common misconception. While it is true that steady state simulations are not valid for batch operations since there is by definition no steady state, dynamic simulations can follow a batch as long as the software can handle zero flows and empty vessels. Model based control (MPC), which looks at trajectories is suitable for optimization of fed batch processes. The opportunities to improve a process’s efficiency by the use of models add up to be about 25% for batch compared to 5% for continuous operations
You need consultants to maintain experimental models.
No longer true. The ease of use of new software allows the user to get much more involved, which is critical to make sure the plant gets the most value out of the models. Previously, the benefits started to drop as soon as the consultant left the job site. Now the user should be able to adjust, troubleshoot, and update the models.
You don’t need good operator displays and training for well designed advanced control systems.
The operators are the biggest constraint in most plants. Even if the models used for real time optimization (RTO) and model based control (MPC) are perfect, operators will take these systems offline if they don’t understand them. The new guy in town is always suspect, so the first time there is an operational problem and there is no one around to answer questions, the RTO and MPC systems are turned off even if they are doing the right thing. Training sessions and displays should provide an understanding of the effect of future trajectories on actions taken by controller
Simple step (bump) tests are never enough. You must do a PRBS test .
A complete pseudo random binary sequence (PRBS) test may take too long. The plant may have moved to an entirely different state, tripped, or in the case of a batch operation finished, before a PRBS test is complete. As a minimum, there should be one step in each direction held to steady state. The old rule is true, if you can see the model from a trend, it is there. Sometimes, the brain can estimate the process gain, time delay, and lag better than a software package.
You need to know your process before you start a MPC application .
This would be nice, but often the benefits from a model stems from the knowledge discovery during the systematic building and identification procedures. Frequently, the understanding gained from developing models leads to immediate benefits in terms of better set points and instruments. The commissioning of the MPC is the icing on the cake and locks in benefits
Optimization by pushing constraints will decrease on-stream time .
Not true. MPC sees future violations of constraints to increase on-stream time
Batch Control 06/06/09 Reagent Optimum pH Optimum Product Feeds Concentrations pH Product Optimum Reactant Reactant Reactant Variability Transfer from Feeds to pH, and Reactant and Product Concentrations Most published cases of multivariate statistical process control (MSPC) use the process outputs and this case of variations in process variables induced by sequenced flows.
PID Control 06/06/09 Optimum pH Optimum Product Feeds Concentrations pH Product Reagent Reactant Optimum Reactant Reactant Variability Transfer from pH and Reactant Concentration to Feeds The story is now in the controller outputs (manipulated flows) yet MSPC still focuses on the process variables for analysis
Model Predictive Control 06/06/09 Optimum pH Optimum Product Feeds Concentrations pH Product Reagent Optimum Reactant Reactant Reactant Time Time Variability Transfer from Product Concentration to pH, reactant Concentration, and Feeds Model Predictive Control of product concentration batch profile uses slope for CV which makes the integrating response self-regulating and enables negative besides positive corrections in CV
Top Ten Signs of an Advanced Control Addiction
(10) You try to use Neural Networks to predict the behavior of your children.
(9) You attempt to use Fuzzy Logic to explain your last performance review.
(8) You use so much Feedforward, you eat before you are hungry.
(7) You propose Model Predictive Control for the “Miss USA” contest.
(6) You develop performance monitoring indices for your spouse.
(5) You implement adaptive control on your stock portfolio.
(4) You carry wallet photos of Auto Tuner trend results.
(3) You apply dead time compensation by drinking before you go to a party.
(2) You recommend a survivor show where consultants are placed in a stressed out old pneumatic plant with no staff or budget and are asked to add advanced control to increase plant efficiency.
(1) Your spouse has to lure you to bed by offering “expert options” for advanced control
Types of Process Responses 06/06/09 The temperature and composition of batch processes tend to have an integrating response since there is no self-regulation from a discharge flow Self-Regulating Process Gain K p = CV CO Integrating Process Gain K i = CV t CO d o 0 1 2 curve 0 = Self-Regulating curve 1 = Integrating curve 2 = Runaway Time (minutes) CV 0 CV Ramp Acceleration Open Loop Time Constant Total Loop Dead Time CO (% step in Controller Output)
What Does PID and MPC See of Future? (Long Term versus Short Term View) 06/06/09 time controlled variable (CV) set point manipulated variable (MV) PID loop only sees this present time MPC sees whole future trajectory loop dead time compensator sees one dead time ahead response PID is best if high gain or rate action is needed for immediate action to correct frequent fast unmeasured disturbances or a prevent runaway
Linear Superposition of MPC 06/06/09 time time time CV 1 = f( MV 1 ) CV 1 = f( MV 2 ) CV 1 = f( MV 1 MV 2 ) set point set point set point Nomenclature: CV is controlled variable (PV) and MV is manipulated variable (IVP)
Feedback Correction of Process Vector and Mirror Image Control Vector 06/06/09 time time time set point set point set point control vector process vector process vector process vector shift vector to correct model error actual CV predicted CV compute future moves for a mirror image vector to bring process to set point trajectory Most MPC packages use standard matrix math and methods (e.g. matrix summation and inversion)
Situations Where Model Predictive Control May be Beneficial
Process and Measurement Noise
Erratic or Stepped Measurement Response
Large Dead Times
Move Size Limits and Penalty on Move (Move Suppression)*
Multiple Manipulated Variables
No PID Control Expertise
06/06/09 * Enables regulation of the transfer of variability from CV to MV
Automated PRBS Test for Fed-Batch Reactor 06/06/09 Non-stationary Behavior (operating point is not constant) Test Data During Fed-Batch Operation
Linear Program (LP) Optimizer 06/06/09 For a minimization of maximization of a MV as a CV, a simple ramper or pusher is sufficient. If the constraint intersections move or the value of type of optimal CV changes, real time Optimization is needed to provide a more optimal solution. MV1 MV2 CV2max CV2min MV2max MV2min MV1max MV1min CV1max CV1min Region of feasible solutions Optimal solution is in one of the vertexes
How Well Can Coincident Constraints Be Handled?
Number of % Time % Time - % Time
Coincident Operator Override MPC
Constraints Can Hold Can Hold Can Hold
One 30% 90% 98%
Two 20% 45% 90%
Three 0% 30% 80%
MPC can hold constraints twice as tight as override and ten times as tight as operator if measurements and final elements precision is not an issue
Example of Basic PID Control 06/06/09 feed A feed B coolant makeup CAS ratio control reactor vent product condenser CTW PT PC-1 TT TT TC-2 TC-1 FC-1 FT FT FC-2 TC-3 RC-1 TT CAS cascade control Conventional Control
Example of Advanced Regulatory Control 06/06/09 feed A feed B coolant makeup CAS ratio CAS reactor vent product maximum production rate condenser CTW PT PC-1 TT TT TC-2 TC-1 FC-1 FT FT FC-2 < TC-3 RC-1 TT ZC-1 ZC-2 CAS CAS CAS ZC-3 ZC-4 < Override Control override control ZC-1, ZC-3, and ZC-4 work to keep their respective control valves at a max throttle position with good sensitivity and room for loop to maneuver. ZC-2 will raise TC-1 SP if FC-1 feed rate is maxed out
Example of Model Predictive Control 06/06/09 feed A feed B coolant makeup CAS ratio RCAS reactor vent product condenser CTW PT PC-1 TT TT TC-2 FC-1 FT FT FC-2 RC-1 TT RCAS MPC MPC MPC Maximize feed rate Model Predictive Control (MPC) set point set point
Example of MPC (Responses) 06/06/09 manipulated variables (MVs) TC-2 jacket exit temperature SP TV-1 condenser coolant valve IVP FC-1 reactor feed A SP TC-1 reactor temperature PV TC-3 condenser temperature PV FC-1 reactor feed A SP TV-2 reactor coolant valve IVP TV-3 condenser coolant valve IVP PV-1 vacuum system valve IVP FV-1 feed A valve IVP controlled variables (CVs) constraint variables (AVs) null null maximize MPC
Top Ten Signs You Have a Dysfunctional MPC Team
The recommended sizes of controllers range from 0x0 to 100x100
The models for the first controller fill up the hard drive
The model after 4 months of PRBS testing looks suspiciously like the model from the first bump test
The completion of the project is tied to the “Second Coming”
Food fights break out in the cafeteria over matrix design
Meetings kick off with kick boxing between consultants
More than one consultant onsite at a time is ruled a health hazard
A psychiatrist is chosen as the best possible project manager
The project over runs it’s Prozac budget
The creators of “South Park” request movie rights to the project
06/06/09 Model Predictive Controller (MPC) setup for rapid simultaneous throttling of a fine and coarse control valves that addresses both the rangeability and resolution issues. This MPC can possibly reduce the number of stages of neutralization needed MPC Valve Sensitivity and Rangeability Solution
06/06/09 MPC Valve Sensitivity and Rangeability Solution
06/06/09 MPC Valve Sensitivity and Rangeability Solution
06/06/09 MPC Valve Sensitivity and Rangeability Solution
06/06/09 MPC Maximization of Low Cost Feed Example
06/06/09 MPC Maximization of Low Cost Feed Example
06/06/09 MPC Maximization of Low Cost Feed Example Riding Max SP on Lo Cost MV Riding Min SP on Hi Cost MV Critical CV Lo Cost Slow MV Hi Cost Fast MV Load Upsets Set Point Changes Load Upsets Set Point Changes Low Cost MV Maximum SP Increased Low Cost MV Maximum SP Decreased Critical CV
Install flow loops or secondary loops to avoid direct manipulation of a valve
Reduce the data compression and increase the update rate of the data historian
Define and document baseline of operating conditions
Define and implement performance indices
For self-regulating responses, steady state = dead time plus 4 time constants
For integrating processes, time horizon is at least 5 dead times
Calculate the integrating process gain for level from vessel geometry and flows
Choose a step size that is at least 5x the noise level or resolution limit
Conduct a simple bump test for each manipulated and disturbance variable
Revise estimates of time to steady state or time horizon and step size
Conduct a Pseudo Random Binary Sequence (PRBS) test if needed
Choose simplest model (fluctuations of 10% in fit or parameters are insignificant)
Simulate the response for changes in targets, economics, and disturbance variables
Increase the penalty on move (move suppression) to reduce oscillation
Decrease the penalty on error and/or priority for less important controlled variables
Provide displays that show future predictions and process metrics
Train operations and engineering on use and benefits
Virtual Plant Setup 06/06/09 Advanced Control Modules Process Models (first principal and experimental) Virtual Plant Laptop or Desktop or Control System Station This is where I hang out
Virtual Plant Integration 06/06/09 Dynamic Process Model Online Data Analytics Model Predictive Control Loop Monitoring And Tuning DCS batch and loop configuration, displays, and historian Virtual Plant Laptop or Desktop Personal Computer Or DCS Application Station or Controller Embedded Advanced Control Tools Embedded Modeling Tools Process Knowledge
06/06/09 Actual Plant Optimization Reactant Ratio Correction Temperature Set Point Virtual Plant Online KPI: Yield and Capacity Inferential Measurements: Reaction Rates Adaptation Key Actual Process Variables Key Virtual Process Variables Model Parameters Error between virtual and actual process variables are minimized by correction of model parameters Model Predictive Control and LP For Optimization of Actual Plant Model Predictive Control and Neural Network For Adaptation of Virtual Plant Optimum and Reference Batch Profiles Actual Batch Profiles Multi-way Principal Component Analysis Super Model Based Principal Component Analysis Adaptation and Optimization