(772a) Automated Real-Time Diagnostics of Model Predictive Control Actions

Aditya Kumar and Michael Nikolaou, University of Houston, Houston, TX

Model predictive control (MPC) is widely used in process industries.  The design and closed-loop behavior of MPC systems can now be supported by well developed theoretical tools.  However, it is well documented that many MPC installations may not function in the automatic mode for a substantial fraction of time.  In many instances, this is caused by human operators switching MPC to the manual mode, because of poor understanding of MPC operation.  In such case, all benefits from MPC are obviously lost.  Questions that operators frequently raise about MPC include the following.

-          Why is MPC taking such action?

-          What will happen if I allow it to take such action?

-          Do I need to intervene?

-          How far can MPC move before I have to intervene?

While experienced operators or engineers may be able to answer some of the above questions, particularly if they spend enough time analyzing data, there is a need for an automated tool that would provide insight into MPC actions, by answering in real time questions such as raised above.  The purpose of this presentation is to provide a diagnostic framework and specific tools for answering some of these questions and related variants.

            The approach taken is algorithmic in nature, namely diagnostic information is generated by applying specific computational tools.  The results of these computations can be easily presented in natural language that can be understood by operators.

            The proposed approach is applied on data taken from a real MPC system with 24 controlled and 12 manipulated variables.  The results demonstrate that a number of questions raised by operators can be answered satisfactorily.  Some of the answers generated agree with the results produced by extensive off-line analysis of the same data by experienced engineers.

            In future work, the proposed framework will be extended to include additional elements, and integration with natural language processing tools that will make it user-friendly will be undertaken.