Re: DMC vs RMPCT

[ Follow Ups ] [ Post Followup ] [ Control Consulting ] [ FAQ ]

Posted by Anonymous on August 15, 2000 at 20:01:45:

In Reply to: Re: DMC vs RMPCT posted by Don Snowden on August 03, 2000 at 09:11:32:

I think that one way that the modeling packages of Honeywell and AspenTech differ is in the use of the model form - Honeywell uses mainly Laplace Transform models, AspenTech uses step or impulse weights.(Actually, I haven't used either package for some time, but the web sites of both companies indicate that they still do the modeling this way.)

Is there any difference to the end user? The esteemed Dr. MacGregor at McMaster University did a study on this very question (see "Identification of Finite Impulse Models: Methods & Robustness Issues, Ind. Eng. Chem Res, 1996, p 4078) and found that the {AspenTech} step/impulse weight models "provide much worse results or require much more data" than {Honeywell} Laplace Transform models. This study would seem to indicate that perhaps the Honeywell product is fundamentally better than the AspenTech one.

Besides getting better models, the Laplace transform methods give better indications of model accuracy (did AspenTech ever get around to adding confidence intervals to their model parameters?), and are more versatile (ever try to get PID tuning parameters from a step-weight model?). The disadvantage to Laplace Transforms is that you can't model very odd responses well (although some would say that the odd-shaped responses from the AspenTech modelling software are often an artifact of too many parameters).

Obviously, there is more to a modeling package than the academic issues that Dr. MacGregor raises, and in other areas AspenTech's package may be superior. Perhaps the statistical efficiency of the model is not an issue - if the data is good, you're going to get good models from either package. Also, Don mentioned that the analysis time is cut down by half using AspenTech; I think Dr. MacGregor would respond that using Laplace Transform models can easily cut the data gathering time down by half (which is usually much longer than the analysis time, and requires perturbing an open-loop plant).

The issue of model form appears to be important academically, but overlooked by the industrial community; perhaps it’s not too important to the users after all. It would be interesting to see if any industrial users have comments in this area.

Follow Ups:

Re: DMC vs RMPCT Don Snowden 15:08:32 8/17/00 (0)
Re: DMC vs RMPCT Dan O'Connor 14:30:11 8/17100 (3)
- Re: (2)

Post a Followup

Name :
E-Mail :
Subject :

Comments:
: I think that one way that the modeling packages of Honeywell and AspenTech differ is in the use of the model form - Honeywell uses mainly Laplace Transform models, AspenTech uses step or impulse weights.(Actually, I haven't used either package for some time, but the web sites of both companies indicate that they still do the modeling this way.) : Is there any difference to the end user? The esteemed Dr. MacGregor at McMaster University did a study on this very question (see "Identification of Finite Impulse Models: Methods & Robustness Issues, Ind. Eng. Chem Res, 1996, p 4078) and found that the {AspenTech} step/impulse weight models "provide much worse results or require much more data" than {Honeywell} Laplace Transform models. This study would seem to indicate that perhaps the Honeywell product is fundamentally better than the AspenTech one. : Besides getting better models, the Laplace transform methods give better indications of model accuracy (did AspenTech ever get around to adding confidence intervals to their model parameters?), and are more versatile (ever try to get PID tuning parameters from a step-weight model?). The disadvantage to Laplace Transforms is that you can't model very odd responses well (although some would say that the odd-shaped responses from the AspenTech modelling software are often an artifact of too many parameters). : Obviously, there is more to a modeling package than the academic issues that Dr. MacGregor raises, and in other areas AspenTech's package may be superior. Perhaps the statistical efficiency of the model is not an issue - if the data is good, you're going to get good models from either package. Also, Don mentioned that the analysis time is cut down by half using AspenTech; I think Dr. MacGregor would respond that using Laplace Transform models can easily cut the data gathering time down by half (which is usually much longer than the analysis time, and requires perturbing an open-loop plant). : The issue of model form appears to be important academically, but overlooked by the industrial community; perhaps it’s not too important to the users after all. It would be interesting to see if any industrial users have comments in this area.
Optional Link URL:
Link Title:
Optional Image URL:

[ Follow Ups ] [ Post Followup ] [ Control Consulting ] [ FAQ ]