(54c) Improving the Efficiency and Reliability of Distillation Model-Predictive Control Projects

This is an outline of a proposed paper to present a summary / overview of what has been learned in the art of implementing model-predictive control (MPC) projects for large-scale process plants. Traditionally, practitioners have performed extensive and invasive plant tests and used the dynamic response data from the plant to ?identify? the required models for the MPC application. In this paper, we will describe the experience gained in four recent projects that have been commissioned using an adaptive multivariable controller. Here, we used first-principles simulation models to eliminate the need for step testing entirely. Such tests are known to be vexatious and expensive, risking violations of process, equipment, safety, and environmental limits.

Earlier in 2003, we had reported experience with use of dynamic simulation for developing the dynamic models required for using an MPC package. These applications were implemented for very slow moving distillation columns (notoriously non-linear and difficult to step-test) with excellent results. Since then, the situation has simplified even further, since the adaptive controller (STAR) used by us in subsequent projects requires the user to provide only the steady-state gains for all input/output variable pairs. The user is not required to specify process dynamics ? these are developed adaptively, on-line. Hence, the need for step tests can be eliminated or at least minimized.

The moral of the story is simply this: If you know how to match an Aspen Plus, Hysys, PRO II, or equivalent simulation model to "steady-state" plant data, you can obtain the required steady-state gains quickly by perturbing the model systematically. This enables building an MPC controller in a very straightforward manner. The adaptive controller we used provides one-button CV tuning, which makes commissioning quick and efficient.

All of the MPC projects described in this paper (gas plants and fractionators) were highly successful; two of them were done for BP in the USA, and two the UK. In this paper we will describe the pertinent methodological details about these applications and draw out the common threads, the recommended project execution procedure, along with a list of "dos and don'ts" for ensuring success. Our objective is to show that MPC projects can be done using first-principles knowledge, while minimizing, if not eliminating entirely, the need for empirical dynamic models based on step tests. This should serve as an interesting departure from conventional practice and thus should provide a useful counterpoint to those who hitherto have only seen step tests to build fixed dynamic models.

By focusing on a case-study oriented discussion of what does and does not work, what potential pitfalls exist, and the lessons we have learned, we hope to interest the audience in this approach. Our understanding is that such first-hand accounts are not commonplace in industry, so such discussions should be valuable for those who are practitioners themselves or are responsible for assigning such project work.