MPC performance assessment

Hi, I am new recruit in a refinery company. We have MPC on distillation column and we are using step response model for MPC. The model we are
using was identified 2 yrs back and so it is felt that model should be re-identified.
As re-identification requires huge cost, I am asked to do economic assessment for it (i.e. will re-identified model give enough performance improvement to compensate for re-identification cost)
I don't know where to start? I think I should start with validating the available model using recent data available. But i m confused, whether i can use the data, collected in closed loop, for validation of step-model?
Also, once I have done validation (and if i find that there is some mismatch between model and plant), how can i evaluate the performance improvement by using updated model.
Thank you for your help.
Regards, Suraj
Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload
One easy method might be to look at the original validation results especially some step change, then run the same single experiment on the column today. One step change, measure the response, and see how it compares to the response two years ago. If it is still fairly close, then you are ok. That isn't going to be so precise, but it will give you a little confidence. Next, examine how much the delayed response is worth in bad product, excess energy, etc. If it is over some threshold, then consider doing the rest of the project.
The performance improvement should be based upon sound business principles, the most important of which is the quest for positive economics. What is the value of off spec product or time lost during a step transition? What is the cost of control on each model?
Why do you feel that the model should be changed after 2 years? Did you make some process change? Bigger or smaller pumps? What has happened that makes the plant operate differently? If these can be identified, then you may need to just accomodate those changes in the model, and again get away with a smaller project.
Michael

Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload

Here are some ideas: First of all, MPC is still a -ve-feedback technology and so you still reap the benefits of a -ve feedback control system, one of which benefits is a certain tolerance of plant modelling errors. Without this, of course, MPC would not be of much use since the models it relies on are inherently crude and simple. Nevertheless, given that there is always the possibility of excessive modelling errors lead to a significant degradation of performance, then you need to know how MPC performance relates to modelling error. For that information: go back to the MPC supplier for advice and information. Onec you have a handle on what constitutes a significant modelling error, you next need to find out whether or not your plant & model have a significant mismatch. Ask your MPC supplier what things you could observe from the MPC and plant performance that are indicative of potentially significant modelling errors. You are now entering the realm of control-system performance monitoring. Has there, for example, been a significant change in controlled plant performance since the system was first installed? If you can establish that there is (a) potential for performance improvement and (b) that the improvement could be achieved by means of an improved platn model, then try to quantify that with the help of the MPC experts, and put a money figure against it. Then compare that with the cost of carrying out the re-modelling exercise.
Kelvin B. Hales Kelvin Hales Associates Limited Consulting Process Control Engineers Web: www.khace.com
Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload
Probably the first step would be to see how well the controller is performing. The easy way to do this is to see how well the measurements are following the setpoints, but this doesn't really tell you how well your controller is doing as the usual reason why a measurement isn't on setpoint is because there's disturbances entering the process (and if there weren't disturbance, you wouldn't really need a controller). But of course you don't know the size of the disturbance as it changes with time, which means the error variance doesn't really tell you if you need a new controller.
To answer one of your questions above, you cannot use the closed loop data for validation of your models - many have tried, all have failed, more or less. And it's because generally the only reason the plant is moving is because of the disturbances, and you don't know what these are. If you did, they'd be feedforward variables. And don't think you can just wait for a setpoint change - they're done by a higher level controls or operators, and these guys are often making their changes in response to a disturbance.
What to do? There are packages available that do measure controller performance without plant tests - all you need is an estimate of the deadtime, but you probably have that from the previous plant tests. The number given by these packages is scaled from 0 to 1, and just indicates how close you are to an "optimal" controller. So if your current control is bad, as indicated by these packages, then it's time to remodel (assuming the economics are there as well). And to answer your last question, they will tell you how much improvement you can expect to get with better models.
Ok, I should point out the I work for a company that sells this stuff. You can find more info at www.ControlArtsInc.com
Add pictures here
<% if( /^image/.test(type) ){ %>
<% } %>
<%-name%>
Add image file
Upload

Polytechforum.com is a website by engineers for engineers. It is not affiliated with any of manufacturers or vendors discussed here. All logos and trade names are the property of their respective owners.