MPC Best Practices for ISA Certification of Automation Professionals (CAP) Tips

Here is a new section on Model Predictive Control Best Practices for the next edition of the Guide to the Automation Book of Knowledge that is the primary resource for the ISA Certification of Automation Professionals (CAP) Program. The following includes text that may not fit into the page allocation for Chapter 13 on advanced control. 

13.9 MPC Best Practices

•1.       Establish a user company infrastructure to make the benefits consistent.

•2.       Develop corporate standards to historize and report Key Performance Indicators (KPI).

•3.       Screen and eliminate outliers and bad inputs and review results before reporting KPI.

•4.       Train operators in the use and value of the MPC for normal and abnormal operation.

•5.       Installation must be maintainable without developer.

•6.       Improve field instrumentation and valves and tune regulatory controllers before MPC pre-tests.

•7.       Eliminate oscillations from overly aggressive PID controller tuning that excite nonlinearities.

•8.       Realize that changes even in PID loops not manipulated by the MPC can affect the MPC models.

•9.       Use secondary flow loops so that MPC manipulates a flow set point rather than a valve position to isolate valve nonlinearities from the MPC.

•10.   Use secondary jacket/coil temperature loops so MPC manipulates a temperature setpoint rather than a coolant or steam flow to isolate jacket/coil nonlinearities from MPC.

•11.   Use flow ratio control in the regulatory system so that the MPC corrects a flow ratio instead of using flow as a disturbance variable in the MPC.

•12.   Generally avoid replacing regulatory loops with MPC if the PID execution time must be less than 1 second or the PID gain is greater than 10 to deal with unmeasured disturbances.

•13.   Use inferential measurements (Chapter 12 linear dynamic estimators) to provide a faster, smoother, and more reliable composition measurement.

•14.   Bias the inferential measurement prediction by a fraction of the error between the inferential measurement and an analyzer after synchronization and eliminating noise and outliers.

•15.   Eliminate data historian compression and filters to get raw data.

•16.   Conduct tests near constraint limits besides at normal operating conditions.

•17.   Use pre-tests to get step sizes and time horizons.

•18.   Step size should be at least five times deadband, stick-slip, resolution limit, and noise.

•19.   Get at least 20 data points in the shortest time horizon.

•20.   Use a near-integrator approximation to shorten time horizon if optimization not affected.

•21.   Get meaning significant movement in the manipulated variables at varying step durations.

•22.   Get the MPC steady state process gains right for analysis, prediction, and optimization.

•23.   Use engineering knowledge and available models or simulators to confirm or modify gains.

•24.   Combine reaction and separation into the same MPC when the separation section limits reaction system performance.

•25.   Use Singular Value Decomposition (SVD) and Linear Program (LP) cost calculation tools to build and implement large MPC applications.

•26.   Reformulate the MPC to eliminate interrelationships between process input variables as seen by similar process gains in a column of the matrix.

•27.   Reformulate the MPC to eliminate interrelationships between process output variables as seen by process gains in one column of the matrix having the same ratio to gains in another column.

•28.   Make sure the MPC is in sync and consistent with targets by planning and scheduling people.

•29.   For small changes in dynamics, modify gains and dead times online.

•30.   For major changes in dynamics, retest using automated testing software.

For the discussion behind these best practices, see the Control Talk column series "Model Predictive Control - Past, Present, and Future, Part 1, Part 2, and Part 3"