May 3rd, 1996

Q: How long did it take you to create the DuPont model?
A: About nine months

Q: Had you been using VisSim before the DuPont project?
A: Yes. About two months.

Q: What modeling tools had you used prior to VisSim.
A: Tutsim and Labview, though Labview isn't really simulation software.

Q: Could you have done this project with those tools?
A: No way!

Q: What was the most interesting part of the project?
A: Good question. I'd have to say when we verified the model with actual plant operation. I was totally surprised at how close the model was to the actual plant output. Many of the unit operations were dead on. I didn't have to retune at all! There was one case where the subsystem had six inputs and the actual output looked to be swamped with noise. I figured it was all stochastic. Then the model mapped exactly onto the crazy output. It was very gratifying.

Q: What was the worst part?
A: Converting our 16-bit DLLs to 32-bit was a pain, though we did rely on Bill at VSI for of lot of that. The other pain was getting the model to run numerically stable with a 0.5 second step size. We had to create a custom first-order filter block to solve flow/pressure loops.

Q: What would you do differently?
A: On the next job, I'd use predefined approaches based on our DuPont experience. I'd say the next job should take about half the time this one did.

Q: Is the approach you took typical in the industry today?
A: No, I don't think so. A lot of folks take a pure PDE (partial differential equation) approach, and only try to solve a small piece of the overall system. Nobody has really attempted to model the whole system from first principles, nonlinear warts and all, from an ODE (ordinary differential equation) standpoint, and get the excellent results in real-time that we have seen.

Q: So we give you a platform for easy creation of complex systems?
A: Absolutely!

Q: How long did you spend validating the model?
A: We're not done yet, but I'd say we spent about 20% of our time validating the model against measured data sets. It was very time consuming. We use multiple channels of 16,000 points per channel. On my old 486, with 16-bit VisSim, it took quite a while to adjust a parameter and rerun the simulation. Things improved dramatically with VisSim/32 and my new Pentium.

Q: How did you go about modeling something like valve?
A: I had the choice of using the manufacturer's specs or measuring flow and pressure drop, and calculating the CV curve myself. When I used the latter method, I used a VisSim map block to do a look-up to get the curve right. I did have a little trouble getting enough points in the curve so the piecewise linear look-up would give an accurate answer.

Q: Do you find that actual CV curves differ from the manufacturer's spec?
A: Yes. Sometimes it would be off as much as 20-30%.

Q: How applicable is the large scale system modeling to the process industry?
A: Very. It gives process control engineers the ability to design controls and evaluate control changes to a process prior to implementation. It works both on-line and off-line. Off-line, you can tune up your DCSs and PLCs with a high degree of confidence with a model like this one. On-line, if you see unexpected variance between plant output and model output, you know you have a problem in the plant, and model will tell you where to look. From there, you just work backward from the output display that shows the variance and see where in the model the variance begins. You can then easily find the plugged line or clogged screen on the plant floor.

Q: What about using "soft sensors" in your models?
A: Sure. A soft sensor is something that a high fidelity model like ours can give you. Let's say you have a VisSim model that realistically represents plant behavior, but there is a part of the process that you just can't physically get a sensor to. Just "right-click" on the part of the model that corresponds to the piece you want to measure. As long as the model is running in real-time and is being fed actual plant inputs, you can use any of VisSim's displays to give you the values that a real sensor would normally provide.

Q: What conclusions can you draw from this experience?
A: High fidelity modeling can be done on a large scale! And, from a financial standpoint, it's extremely worthwhile to the process control community. There are tremendous cost savings in reduced downtime due to off-line tuning and control design, as well as operator training.