Model Predictive Controller

Do you design and support Model Predictive Control (MPC) like DMC, RMPCT, PredictPro, Connoisseur or traditional Advanced Process Control (APC) inside a DCS or PLC?

When Model Predictive Controls (MPC) or Advanced Process Controls (APC) deteriorate due to bad models or changes in plant or process conditions, you may see ugly oscillations like below:

Model Predictive Controls (MPC) can have 10s or 100s of dynamic models. One or more could be wrong. Bad (wrong) Model Predictive Control (MPC) dynamic models produce a bias (model prediction error) between the predicted signal and measured signal coming from the sensor.

How to identify the bad (wrong) Model Predictive Control (MPC) models and how to generate correct models?

This has been a long challenging problem and a nightmare for the process control staff responsible for Model Predictive Control (MPC) quality and performance. Now COLUMBO offers a breakthrough, novel and state-of-the art solution like never before.

• DMC (Dynamic Matrix Control), RMPCT (Robust Model Predictive Control Technology), PredictPro, Connoisseur and other Model Predictive Control (MPC) systems are widely used in chemical, petrochemical, paper, power plant and oil refining industries.
• A Model Predictive Control (MPC) system can have anywhere from 3 to 50 MVs (manipulated variables) and 5 to 100 CVs (controlled variables).

• COLUMBO provides you a novel, modern and breakthrough method to tackle the most difficult Model Predictive Control (MPC) problems.
• With time, due to aging equipment, hardware changes, changes in process, operating and economic conditions and process nonlinearities dynamic models can change significantly causing the Model Predictive Control (MPC) quality to deteriorate.
• Model Predictive Controls (MPC) are often turned off with subsequent loss of benefits and profits. This can be a control engineer’s nightmare and headache for the plant management that spent a lot of money on the Model Predictive Control (MPC) project.
• The root cause of Model Predictive Control (MPC) problems and deterioration in quality are most definitely changes in the dynamic models inside the Model Predictive Control (MPC).
• The shapes of the Model Predictive Control (MPC) models is the heart and soul of the Model Predictive Control (MPC) system. Unfortunately, when Model Predictive Control (MPC) models change, or were wrong to begin with, it is a very difficult job to fix the models and identify the correct models.
• This is because several MVs (manipulated variables) are associated various CVs (controlled variables). Which model(s) is/are wrong? It is a hard problem to solve.
• PiControl’s COLUMBO provides an amazing new and novel method to identify the correct models. COLUMBO functionality is what people consider “magic” today. COLUMBO is a process control technology marvel with true breakthrough technology and algorithm.

• Currently practiced conventional known methods of model Identification typically involve making small step tests on the setpoints of slave PIDS in auto mode (often called closed-loop tests).
• Or they involve making step tests on control valves (final control element) in manual mode (often called open-loop tests).
• Note that closed-loop tests still involve steps in most cases and most methods. In these conventional methods, slave PIDs may be in auto mode but stepping on their setpoints is still required. The above is current and conventional thinking and methodology.

• Even in the new and latest DMC3 Calibrate method – step tests are superimposed on top of the MV move trajectories from a DMC controller that is active (on).
• So here, the DMC controller is on, but step tests are still needed for the DMC3 Calibrate to work. Step tests are “open-loop” by definition.
• Types of Dynamic Models – Transfer Function Models or Step Response Coefficients
• Most Model Predictive Controls (MPC) characterize process dynamics using either transfer function parameters or step response coefficients.
• COLUMBO can identify and improve models in both formats. COLUMBO can be used to improve the model accuracy and control performance of any Model Predictive Control (MPC) (COLUMBO is not Model Predictive Control (MPC) vendor specific) but works with any Model Predictive Control (MPC) system from any Model Predictive Control (MPC) vendor.

• ARMAX (auto regressive moving average with exogenous inputs) and Box Jenkins models have been used for several decades and are the main modeling methodology provided by Matlab from Mathworks and Labview from National Instruments.
• These methods do not work well with complete closed-loop data without any step tests at all and/or slave PIDs or MVs in cascade modes.
• They also do not work well with unmeasured disturbances and noise. These methods offer what they call “noise models” for separating noise from the true process dynamics, but when unmeasured disturbances comprise of a mixture of fast random noise, medium frequency drifts and slow unmeasured disturbances, the noise model concept is not useful.

• Most models are based on making small steps of about 1-3% of the current MV values. So, if a setpoint of an MV (a slave PID) is (say) 100 kg/h, then changes from 100 kg/h to 102 kg/h or 98 kg/h or so) are made.
• If the step changes are too large (say going from 100 kg/h to 120 kg/h and then to 90 kg/h), this may upset the process and the step cannot be help for too long and the process gain of the model cannot be unambiguously determined and there will be a lot of uncertainty in the data. COLUMBO overcomes the problem by analyzing closed-loop data with Model Predictive Control (MPC) running, a function not possible with other current Model Predictive Control (MPC) algorithms and software products.

• COLUMBO is the latest breakthrough technology for identifying open-loop dynamic models using completely closed-loop data. No step tests are required.
• Slave PID controllers can be in auto mode and even in cascade mode. How does this work? Please read below…

• Currently, Model Predictive Control (MPC) models are built using small steps (1-3% of current setpoints). When the Model Predictive Control (MPC) is running, it uses these models that were identified using the small steps to calculate the trajectory for various MVs in closed-loop mode.
• When a feedforward changes (measured disturbance changes), the Model Predictive Control (MPC) calculates compensating moves based on the models that were identified using the small steps.
• The difference now is that a measured disturbance signal (feedforward in Model Predictive Control (MPC) ) may have changed by 20% (not 1-2% like during step tests for model identification) and the larger change (20%) can lead to nonlinearities that cannot be seen during the 1-2% smaller moves.
• Also, during Model Predictive Control (MPC) closed-loop mode, several MVs are changed simultaneously (in a correlated manner) and this is also different compared to the model identification phase where typically only one MV is moved at a time to avoid complications due to correlations.
• So, there are two main differences between the Model Predictive Control (MPC) model identification stage (to identify Model Predictive Control (MPC) models) and the closed-loop stage with Model Predictive Control (MPC) running (to calculate MV trajectories in order to keep the CVs at their targets) and these are:

A) Moves are 1-2% of the3 current slave setpoints during model identification stage but could be 1-30% of the current slave setpoints with Model Predictive Control (MPC) running trying to provide control.

B) During model identification stage, only one MV is moved at a time followed by waiting time to try to see the new steady state. This is recommended by all Model Predictive Control (MPC) vendors as a good guideline to avoid correlation problems as if more than one MV (slave PID setpoint) is changed at the same time, then there is ambiguity and uncertainty regarding the impact of the multiple variables and subsequently dynamic models could be wrong due to correlations. However, when the Model Predictive Control (MPC) is running, the Model Predictive Control (MPC) moves various MVs (slave PID setpoints) simultaneously in order to keep the CVs at their targets.

• The differences between the open-loop model identification stage and the closed-loop stage with Model Predictive Control (MPC) on and running can be significant especially since there is nonlinearity of varying magnitude in all real-life chemical processes.
• The closed-loop behavior with multiple MVs changing by larger amounts is often different compared to the open-loop behavior with smaller changes and only one MV changing at a time. These differences lead to model prediction errors and is often the root cause of poor control in many Model Predictive Control (MPC) systems.
• When these problems occur, the fixing and the solution are not easy. COLUMBO was designed to address these problems and help the control engineer come up with improve the Model Predictive Control (MPC) models and thereby the Model Predictive Control (MPC) quality and performance.
• If after a few months or years, if an Model Predictive Control (MPC) is not performing well and if you turn off the Model Predictive Control (MPC) and resume step tests again in an effort to re-identify the Model Predictive Control (MPC) open-loop models because things have changed the problems caused by the differences explained in the A and B bullets above will still cause a potential error in the real Model Predictive Control (MPC) models when the Model Predictive Control (MPC) is running and moving several MVs simultaneously with larger step sizes.

How does COLUMBO work? What are details about the COLUMBO algorithm and internals? The COLUMBO algorithm is a PiControl invention and PiControl corporate intellectual property and cannot be disclosed in open literature. All that can be revealed is that COLUMBO tries to identify new models based on the measured data (CVs) and the MV trajectory calculated by the Model Predictive Control (MPC) system or the Advanced Process Control (APC) system. Internally, COLUMBO is equipped with a powerful and new optimizer designed for handling multiple inputs, both analog and digital subject to constraint limits and nonlinear math processing capabilities. Written in C++, COLUMBO code is super-fast and super-compact allowing the optimization calculations to complete and converge in an amazingly short time. See below an overview of how COLUMBO works: