(151c) Overcoming Numerical Ill-Conditioning in a Dmcplus® Composite Controller at the Baton Rouge Chemical Plant

ExxonMobil's Baton Rouge Chemical Plant (BRCP) employs a large DMCplus® Composite controller on its olefins plant to coordinate the movement of cracking furnaces with downstream constraints. Approximately 300 manipulated variables (MVs) and 1000 controlled variables (CVs) in the DMCplus® Composite controller successfully push the plant to new levels of production while maintaining multiple constraints.

The DMCplus® Composite controller was designed by several Advanced Process Control (APC) experts internal and external to ExxonMobil, but the maintenance of the controller continues to require the majority of two ExxonMobil Applications Engineers' time. The amount of continuing maintenance for the BRCP DMCplus® Composite controller can be justified by its size, but is also related to numerical solution problems, e.g. failure to converge, that occurred during the application's lifecycle.

Because the olefins cracking furnaces at BRCP process a wide variety of feed-types, the DMCplus® models of the furnaces and the downstream controlled variables must change depending on feed-type. The controller configuration file (ccf) switching utility is sometimes used to switch models/ccfs during run time to account for changes in process models. BRCP, however, would have had to maintain many different models/ccfs for many different furnaces to represent all process models for the different feed-types. Instead BRCP used feed-type-dependent gain multipliers (GMULTs) to gain schedule and change models -- precluding the need for distinct models/ccfs per feed-type. This was a more attractive solution compared to ccf switching.

Despite conditioning the model gain matrix against small condition numbers and near singularities, the DMCplus® Composite controller frequently failed to converge from the beginning of its commissioning. It was soon discovered that the optimization search engine only recognizes single-precision GMULTs. Model gains and GMULTs that had been conditioned to 15 decimal places were actually being replaced with numbers of much lower resolution. Rather than converting all base models to single-precision, ExxonMobil made other compromises: the GMULTs were simplified to one and two significant figures and some aliasing (the connections between upstream MVs and downstream CVs) was eliminated in the DMCplus® Composite controller.

While the broken alias connections simplified the optimization problem and reduced the frequency of convergence failures, they also caused the compromised DMCplus® Composite controller to behave incorrectly. For example, one type of compromise forced the DMCplus® Composite controller to treat distinct feed-types as the same, thereby failing to exploit the difference in yields to most effectively control multiple downstream constraints like drum levels and distillation tower delta pressures. Considerable manual intervention was required to keep the DMCplus® Composite controller running without convergence errors or timeouts while also having appropriate behavior. ExxonMobil engineers experimented with different combinations and complexities in alias connections, GMULTs, and DMCplus® Composite controller tuning to get quickly converged and correct solutions. The experiments were less than successful - the convergence errors and timeouts continued.

ExxonMobil engineers worked with Aspen Technology, Inc. to try to reduce convergence errors and timeouts so the compromise could be taken out of the DMCplus® Composite controller. The most significant improvement came with Aspen's change to an interior point optimization search engine, a linear programming method that achieves optimization by going through the middle of the solid defined by the problem rather than around its surface (as does the Simplex method).

The service factor of the BRCP DMCplus® Composite controller has improved, climbing many percentage points, and with the help from Aspen, and ExxonMobil will move back to the original design of the DMCplus® Composite controller.

[©2007 Exxon Mobil Corporation. To the extent the user is entitled to disclose and distribute this document, the user may forward, distribute, and/or photocopy this copyrighted document only if unaltered and complete, including all of its headers, disclaimers, and other information. You may not copy this document to a Web site without approval from ExxonMobil. ExxonMobil does not guarantee the typical (or other) values. The information in this document relates only to the named product, materials or processes when not in combination with any other product, materials or processes. We based the information on data believed to be reliable on the date compiled, but we do not represent, warrant, or otherwise guarantee, expressly or impliedly, the merchantability, fitness for a particular purpose, suitability, accuracy, reliability, or completeness of this information or the products, materials, or processes described. The user is solely responsible for all determinations regarding any use of material, product or process in its territories of interest. We expressly disclaim liability for any loss, damage, or injury directly or indirectly suffered or incurred as a result of or related to anyone using or relying on any of the information in this document. There is no endorsement of any product, material or process, and we expressly disclaim any contrary implication. The terms, ?we? and "ExxonMobil" are used for convenience, and may include any one or more of ExxonMobil Chemical Company, Exxon Mobil Corporation, or any affiliates they directly or indirectly steward. ExxonMobil is a trademark of Exxon Mobil Corporation.

DMCplus is a multivariable control product developed by Aspen Technology following its merger with Dynamic Matrix Control Corporation and Setpoint, Inc. DMCplus is a trademark of Aspen Technology, Inc.]