(84c) Improved Closed-Loop Subspace Identification Technology for Adaptive Modeling and APC Sustained Value | AIChE

(84c) Improved Closed-Loop Subspace Identification Technology for Adaptive Modeling and APC Sustained Value


Harmse, M. - Presenter, Aspen Technology, Inc.
Zheng, Q. - Presenter, Aspen Technology, Inc.
Zhao, H. - Presenter, Aspen Technology, Inc.
Campbell, J. - Presenter, Aspen Technology, Inc.

Sustaining the value of Advanced Process Control (APC) applications has been a focus for many refinery and petrochemical companies over the past decade.  Changes in plant operation such as equipment aging, lost bubble caps, fouling of packing in heat exchangers, process modifications and feedstock changes will deteriorate the performance of any APC applications and reduce the financial benefits of the controller.  One of the necessary steps is to maintain the process model and keep it up to date.  Frequent precision revamps of running controllers via a partial plant test and re-identification of a specific sub-model of the process is typically the most efficient and practical approach.  New tools for model accuracy monitoring and model adaptation have been developed by exploiting the latest technology, such as automated closed-loop step testing (e.g. using the SmartStepTM algorithm), online model quality monitoring and model re-identification.  These new tools are designed to help APC users to diagnose and repair model-mismatch as often as is needed.  The main benefit is to prevent the long delays occurring between model mismatch and model repair.  Among the new tools that are utilized for this purpose, subspace identification plays a critical role.  

Subspace identification emerged in the 1980s and eventually attracted a lot of academic attentions.  AspenTech has first implemented it in DMCplusTM (AMS 3.0) and this software was released in October 2000. In contrast to what is reported in most case studies found in academic papers, in industrial APC projects the users have to deal with many challenging problems that were not considered in the academic papers, such as the very large number of inputs and outputs.  For example, the largest reported DMCplus controller has 283 manipulated variables (model inputs) and 603 controlled variables (model outputs), with thousands of native collinear subsystems, as well as extensive sparsity in the model.  The main objective is to use cost-efficient ways to test the process unit in order to get more accurate models more quickly, since that will ensure better performance of the DMCplus controller.  Control engineers are concerned with the test duration because the unit operates sub-optimally during this time period, and product give-away is often more expensive than the man-hours involved in the test.  They are also concerned about the accuracy of the model’s gains and gain-ratios (colinearity aspects).   After more than 10 years of practical use in the process industry, the major features of subspace identification such as true multi-input multi-output (MIMO) identification and faster model convergence have been evident, resulting in substantial improvements in project execution efficiency.  Recently, the properties of subspace identification algorithms in closed-loop environment have also been well studied and documented (e.g. Qin), and some improved algorithms have emerged to address these issues.  With these developments, modern subspace identification technology has now matured to the point where it works in a synergistic way with innovative multivariable step testing technology like SmartStep.  The combination of subspace identification and the multivariable step testing algorithms ensures that online adaptive modeling is becoming a reality.  Control engineers doing APC maintenance work will be able to take full advantage of these technologies, ensuring the sustained value goal is met.

In this paper the recent progress of subspace identification technology is reviewed and several algorithms dealing with closed-loop data will be summarized in a unified Least-Squares framework developed by Qin (2006).  Special issues in industrial MPC practice will be discussed and how that can be addressed will be demonstrated with actual project examples.