Skip to Content

Register Login

Industry Online Support

Technical Forum

Navigation

Model Predictive Control - Introduction

Created by: RonaldNijssen at: 9/12/2008 12:03 AM (11 Replies)

Rating (9)

Thanks 4

Result pages:: |; 1 | 2; |

12 Entries

Entries per page: 10 | all

9/12/2008 12:03 AM	Rate (1)
RonaldNijssen Advanced Member Joined: 1/21/2008 Last visit: 7/12/2024 Posts: 39 Rating: (14)	Dear, I have been working on exploring the new SIMATIC PCS 7 Model Predictive Controller I was thinking that it may be interesting to share some experiences, ideas and answer some questions. You may not be working on an application but learning something through this medium for development of your skills may be interesting. At the same time I get to learn more about questions and use-cases that exist. So, this is a thread with the objective to share information and have the readers learn about this exciting topic Please let me know if you are interested and I will do my first postings, in which i explain what MPC is about, what you can expect and do with it Kind regards Ronald Nijssen

Suggestion To thank Quote Answer
This contribution was helpful to 1 thankful Users CHINA_PCS7

9/12/2008 2:49 AM	Rate (0)
Phantom75 Posts: 2661 Rating: (279)	Sure Ronald! This is a very interesting topic and so far Siemens is lagging behind the competition (Honeywell for example). Any info is very welcome!!
Phantom75 Posts: 2661 Rating: (279)	Last edited by: Phantom75 at: 9/12/2008 2:50 AM
Suggestion To thank Quote Answer

9/12/2008 3:02 AM	Rate (0)
RonaldNijssen Advanced Member Joined: 1/21/2008 Last visit: 7/12/2024 Posts: 39 Rating: (14)	Hi Phantom! I will get something going, when you tell me in a private conversation what Honeywell has to offer as a standard built-in product feature. I will do my first posting to this thread soon Kind regards Ronald Nijssen

Suggestion To thank Quote Answer

9/12/2008 7:17 AM	Rate (0)
Pandiani Bronze Member Joined: 10/31/2005 Last visit: 9/4/2024 Posts: 366 Rating: (31)	Ronald, I'm looking forward to it...

Suggestion To thank Quote Answer

9/12/2008 11:28 PM	Rate (3)
RonaldNijssen Advanced Member Joined: 1/21/2008 Last visit: 7/12/2024 Posts: 39 Rating: (14)	Model Predicitive Control (1) I started to look at this technology with the introduction of it in PCS 7 V7.0 SP1. We would show its functionality with the example project that is installed automatically with this version of PCS 7. It seemed straight forward, until I talked with the first customers. The first thing that is required to be understood is the role of a Control loop. We have used PID control so often that we seemed to forget what we are actually doing. The objective of a control loop is to create a response to a process error that will cause the process value (or control value) to return to a desired level (setpoint). The way a PID loop does this is relatively simple, it takes the error, multiplies it with a Gain and sends the resulting value to a manipulating device (for example a control valve). This mechanism (the P action)will leave an error in place as the result, simply because the error is the source of the response (no error = no response). So an I action has been added to the PID loop, this action takes the error, multiplies it (applies a time constant) and increments or decrements the manipulated value each time the algorithm is executed, the level of increment or decrement is based on the error and the time (Ti). Tuning tolls (like the PCS 7 PID Tuner) will calculate the optimal time (Ti) based on the process time constant and the optimal Gain (P) based on the process its responsiveness to a change in manipulated value. The D action has been added as well to allow the PID control loop to take a pre-emptive action based on change of the Control Value. If you look at the PID algorithm then there are several conclusions you can draw: 1. Its response, although timed nicely using a tuning tool, is momentary (it reacts to a given situation without anticipation of the future) 2. Every single PID loop does its own thing, it doesn't care about other Control values, unless you implement some sort of override control 3. As PID loops orchestrate their own response, several PID loops maye create instability among each other if their Control values are somehow related in a process. An example of "3" is a Distillation column, when the reflux, the return flow in the Top, is increased to control the top-temperature, then, after some time, the reboiler will be hit with more liquid and will be required to respond. This will generate more vapor, that will have an impact on the top-temperature. This will result in a change of reflux that ..... As a result a temporary offset may cause some upswing. Another sample is a process where one control value is not allowed to exceed a specific threshold, but while it is not close to that level, a second control value can be the main control value. To make things more complicated, the first control value has a considerable time delay before it starts to reach its excess value. How do make the second loop aware of what may happen in the future with the first control value? Momentary override control will not be sufficient. As you may expect, the answer that was developed many years ago was control, based on an algorithm, that looks ahead (predicts) and can compromise its output to the manipulating devices based on the error of multiple control values (multi- variable) The SIMATIC PCS 7 APC library offers a MPC block that does this. Next time I will write about the timing and response aspects of a process, and how a MPC controller deals with them. This will be a good introduction to the first step of designing a MPC controller, which is the modelling phase. The basis for the model is the first word in "Process Control", the Process! Ronald Nijssen

Suggestion To thank Quote Answer
This contribution was helpful to 1 thankful Users PSW2586

9/13/2008 2:39 AM	Rate (0)
Dorftrottel? Silver Member Joined: 8/10/2006 Last visit: 8/13/2024 Posts: 298 Rating: (29)	Ronald, I sincerely appreciate this primer. Can the moderators make it a sticky for a while at least? Jim

Suggestion To thank Quote Answer

9/16/2008 5:13 AM	Rate (1)
RonaldNijssen Advanced Member Joined: 1/21/2008 Last visit: 7/12/2024 Posts: 39 Rating: (14)	Model Predictive Control (2) This is the second posting regarding this subject. And as promised, this posting will address the Process. In PID control the process typically is boiled down to a single Process Variable (or Control Variable, a CV) and one Output to the process (or Manipulated variable). A disturbance (Disturbance Variable, DV) can be introduced to add a pre-emptive function to the algorithm. In Model Predictive Control the bandwidth of the Process that is being controlled typically needs to be expanded, the built-in MPC controller that you find in PCS 7 will be able to control 4 CV's, using 4 MV's and it will support one DV. What does this mean (beside the numerical presence). This implies that a single control algorithm will control ofmultiple Temperatures, and a Pressure in a distillation column. (the CV's) It does so by manipulating up to 4 signals that in any way have impact on the process. (reboiler energy, reflux flow, pressure control) It also allows for one disturbance (uncontrolled variable). (feed flow) The number of CV's and MV's can be unrelated, you can for example control 3 CV's with 2 MV's, all MV's work together to control all CV's (there is not a one-on-one relationship like in PID control) The MV's and DV are variables that have impact on the process, the result of this impact is the CV's their value(the measurements). This is the core of the concept. Because we capture the impact of the combined activity of the MV's and DV on the process, and the resulting effects on the CV's, we can have an algorithm use that understanding to come up with a balanced control action. To be able to identify the impact of the MV's and DV, we need some mathematical tools to extract that impact and qualify/quantify it. The impact is defined for each individual MV and DV as it works on each CV. This means that, if all MV's, DV's and CV's are used in the PCS 7 MPC controller, the impact consists of 20 (5 x 4) relationships. The impact is defined as a time-based coefficient for each of these relationships. (a number that changes over time) The exercise of identifying this impact is the Modelling process, SIMATIC PCS 7 provides a tool (Modprecon) that performs this exercise, this tool is included in PCS 7 V7.0 SP1. Also if you do not intend to implement a MPC controller, the information that the tool extracts from a process response, can be very helpfull to define initial settings for a PID control algorithm. The produced data will give direction to the process gain (how strong does the process responds to a MV change) and process timing (when does the process show to be steady again). I have attached a zip file with screenshots and exports of the following process: One CV, two disturbances and one MV. Note that a MV has the same effect as a disturbance, it has impact on the process. A MV seems to be different, because it is being manipulated, but in regard to what it does to the process its impact is not qualifieddifferent from the impact of a disturbance variable. The following picturesare in the zip file 1. Processresponse.jpg, this is a screenshot of data recorded in the Trend tool of the CFC editor. It shows the CV, DV, DV2 and MV (OUT) As you can see, the CV changes somewhat, the OUT changes continuously and the DV's change stepwise. The OUT is controlled by a normal PID loop 2. The first step that the modelling tool will take is to extract the different DV's and MV's from the data. The results can be found in CV1. jpg, MV1. jpg, DV.jpg and DV2.jpg Note that MV1, DV and DV2 are impacting the process and CV1 is the result of that impact. 3. The relationship between DV and CV1, DV2 and CV1, MV1 and CV1 is being identified and displayed in MV1-CV1.jpg, DV-CV1.jpg and DV2-CV1.jpg. Note that these relationships show the positive or negative response and time (secs). A positive response implies that an increase of the DV or MV results in an increase of the CV, and that a decrease will result in a decrease. A negative response reverses this (increase results in decrease and decrease results in increase). The model is a process with two heating sources (DV and DV2) and one cooling source (MV1), CV1 is the temperature. After the relationships have been identified, the tool will use these to "Model" the response by performing a calculation and in CV1.jpg you can actually see the % match between the reality (recorded value of CV1) and the Model (calculated CV1). Once the Model (in fact the relationships) have been identified, we are half-way to a working MPC controller, however, we can see now, what gain/time response each DV and MV has on our process, as mentioned before, this is usefull for any type of control strategy. Next time we will talk about how this information is turned into a control algorithm and how that control algorithm uses it to drive up to 4 MV's. After we complete this thread, I can provide additional information to have you generate your models yourself and also sample projects with simulation. Ronald Nijssen Attachment MPC-Model.ZIP (713 Downloads)

Suggestion To thank Quote Answer
This contribution was helpful to 2 thankful Users CarlosVieira PSW2586

9/18/2008 10:12 PM	Rate (0)
Pacanins Posts: 23 Rating: (1)	Questions: 1) are you using the MCP example that is included in PCS 7 V7.0 SP1 and 2) If it is not, could you please include your Project or multiproject with the simulation. 3) I suppose you are running a simulation inside a Controller and not using PLCSIM. (Timing issues) Request: Could you put all these explanations in a power point presentation? Thank you very much for writing this explanations!
Pacanins Posts: 23 Rating: (1)
Suggestion To thank Quote Answer

9/30/2008 6:10 AM	Rate (0)
RonaldNijssen Advanced Member Joined: 1/21/2008 Last visit: 7/12/2024 Posts: 39 Rating: (14)	Sorry for not responding earlier. To your question, I am using a demo I created myself, it has three models in it, and I will post this demo soon so you will be able to drive it yourself. Part of the demo is an unsupported (under the hood) function I built that reads the prediction from the Instance Data Block that goes with the MPC block. I display this prediction in a Function Trend on the OS. A pre-requisite is the use of a real controller (instead of PLC-SIM) since I am using BSEND to send that array of reals to the OS Ronald

Suggestion To thank Quote Answer

9/30/2008 6:21 AM	Rate (3)
RonaldNijssen Advanced Member Joined: 1/21/2008 Last visit: 7/12/2024 Posts: 39 Rating: (14)	Model Predictive Control (3) Last week I visited two customers in the Chemical Industry sector. I had the opportunity of discussing MPC with one of them, their application consisted of reactors that are charged with reactant chemicals. Typically the feed flow/exothermal reaction and resulting energy-balance are metrics that are driving the charging steps in these processes. Controlling these metrics will generate more yield. This customer did look at MPC in the past but had stepped away from it for lack of value compared to the cost associated with it. The discussion went down the path that we discussed the MPC tools and Functions in PCS 7 as being relevant in three ways, where the user can decide to stay with one step and not proceed. the steps are: 1. Format recorded (already available?) trend data and feed it into the ModPreCon tool, let the tool generate the Process response information and use that to optimize existing control or educate operators 2. Generate a MPC controller data set and have it predict, still driving all MV's yourself, by means of logic, manual operations or PID control. if the prediction is read from the MPC block, which as mentioned requires some dedicated custom code, then this prediction can be presented. The customer made the funny remark that he then can call the operator telling him that he will have a problem in 30 minutes! Actually this MPC function could run on a separate controller, for example an AS RTX, and get the MV, CV and DV's through a network connection or OPC interface from any other system. 3. Let the MPC controller do the control and use the MV's it calculates to drive your process. There is clearly effort, and as such costs, involved in going through these steps, the amount of effort is small for step 1 and increases in Step 3 (this step requires more functionality to be added like overrides, manual switches, operator interface). However, the license cost is relatively low and when an AS RTX is required, these come at a low price as well. Ronald

Suggestion To thank Quote Answer

|

Share this page:

Share this page on...

Follow us on