When implementing a new Model Predictive Controller (MPC) project or maintaining an existing MPC application, model identification is the most difficult, time-consuming and labor-intensive part of the process. Traditional testing, modeling, integration and commissioning procedures typically take three to four months of around-the-clock work to complete.
BOC Gases (now part of the Linde Group), which wanted to replace the poorly performing MPC on the air separation unit at its Hartford, Ill., site, sought a faster and lower cost method to deploy a new system while maintaining safe and effective plant operation. So, in 2006 the company partnered with Matrikon, Inc., Edmonton, a veteran of 49 conventional MPC implementations at other BOC sites, to apply a radical new methodology using Tai Ji technology in testing and modeling the Hartford plant.
BOC’s Hartford air separation unit makes both liquid and gas products. On average, the plant produces 600 metric tons of gaseous oxygen and 650 metric tons of liquid nitrogen and oxygen. The site also has gaseous oxygen pipelines that feed nearby customers and a gaseous argon/nitrogen (GAN) pipeline.
BOC Gases monitors most of its plants from remote operation centers (ROC); few personnel actually work at the plants. Operators at the ROC are watching five-to-six plants at any given time. To ease the load on operators, BOC decided in October, 2001, to implement an MPC for the plant.
In five years of operation, the performance of this original MPC consistently fell short of specifications. Operators complained of constant controller issues that regularly led to the MPC being switched off and the plant running under operator control. Worse yet, the operators called the plant one of the hardest for them to run.
The process at Hartford is very nonlinear, especially the low pressure column (LPC) purity control, which is extremely critical. Running the plant under automatic (regulatory) control is difficult and often can require operator intervention to maintain purities; even with manual intervention, some products, especially GAN and cryogenic liquid argon (CLAR), can fall below specifications.
With operators burdened with monitoring five or six other sites, reliance on automatic control needed to be reduced to an absolute minimum. It was crucial that MPC be revamped as soon as possible. And, of course, reliable MPC was a must.
The Tai Ji implementation method
BOC and Matrikon decided to use Matrikon Control Performance Monitor Tai Ji to greatly accelerate the commissioning process. Based on Dr. Yucai Zhu’s industry-proven Tai Ji identification technology, it provides an alternative to traditional step testing and model identification methods. BOC saw three clear advantages in the method:
- It’s automatic. Testing and model identification are done automatically rather than manually by engineering personnel.
- It’s multivariable. Multiple manipulated variables (MV) are tested simultaneously, making test time much shorter than for single variable tests.
- It’s closed loop. Tests can be performed closed loop (MPC or PID), resulting in fewer disturbances and operator interventions.
In short, this new method of MPC implementation allows the modeling, step testing, integration and commissioning phases to overlap — at the procedure’s peak all four phases, in fact, are underway simultaneously — while requiring fewer personnel resources, at a greatly reduced risk of process upsets. Step-by-step, the methodology ran as follows:
Closed-loop operating data were obtained from the plant — approximately one month’s worth of data from when there was a lot of movement in the plant. The data were put through Control Performance Monitor Tai Ji to obtain initial models. These then were incorporated into the MPC and the controller was brought online.
The plant was commissioned for two days to ensure that everything was running correctly and that the MV and controlled variables (CV) were functioning as they should. Once the initial commissioning was complete, a Tai Ji step test was performed for three to four days. This test involved moving all the MV simultaneously by perturbing CV target values every minute.