New Approach to Multivariable Process Control (MPC)

End-users of traditional model-based predictive Multivariable Process Control technology (MPC) are increasingly looking for alternative technologies to fill the gaps that have emerged in the wake of industry’s aggressive adoption and deployment of MPC technology throughout the 1990s and 2000s.

Early on, MPC proved itself as a powerful multivariable control and optimization tool for many applications.  But long-term experience has also shown that MPC is not appropriate for all multivariable applications, especially industry’s numerous smaller applications, where agility, operability, and unattended reliability, are especially important.  Moreover, MPC technology continues to struggle in several basic areas, including cost, maintenance, support, reliability and performance.

Current State of Multivariable Process Control

The present MPC applications available have the following well-known features:

• Emphasis on detailed models
• Embedded optimization
• Large matrix design practice
• Limited performance metrics

The majority of expense and complexity associated with MPC derives directly from its dependence on detailed models, but experience now shows that model accuracy is a dynamically moving target, so that even automated re-modeling will not overcome this structural conflict.  This is the same lesson that ultimately dashed the hopes of single-loop “auto-tuning”. 

Embedding optimizers within MPC applications add further cost and complexity, and may not be necessary in many applications.  For many applications, the operating team, by the nature of their jobs, already knows where the optimum operating point resides, as does the business planning side of modern operating facilities.  The “big envelope” matrix design practice for certain industries and applications is not appropriate for others.  While the overall “service factor” metric is widely used and acceptable to many, there is a need for more precise performance metrics.

Observations such as these cast a new light on the old paradigm, highlighting the need to take a fresh look at the lessons that have been learned, make appropriate course corrections at today’s crossroads, and lay out an updated roadmap to move important MPC technology forward again.

Opportunities for Improvements ahead

Based on this experience, new approaches should consider the following features and changes:

• A gain-direction matrix, without detailed models
• Control-layer automation, availed of business-layer optimization
• MPC core-competency and agility at the plant level
• Appropriate “operational” performance criteria
• Effective transparent metrics

Probably the most important and enduring contribution of the model-based era to process automation will prove to be the matrix (not models).  The matrix – without detailed models – has emerged as the foundational concept for multivariable process control, constraint management and optimization.  The matrix puts all members of the operating team – process engineers, operations personnel, and control engineers – onto the same page, to align their efforts for more effective results.

The optimum operating point in a process is typically well-known by the operating team.  In addition, it is known with even deeper insight by business planning functions, who are exposed to numerous factors not exposed in the control system layer.  This makes it unnecessary for many processes. The essential role of MPC is to honor constraint limits and optimization targets in the live process environment, where the related process values – not the optimization results themselves – are subject to change in real-time.

Essentially all processes, large and small, are multivariable – they have one set of available “handles” and another set of variables to be managed and optimized by effective use of those handles.   Therefore, in the modern age, multivariable MPC must be a core-competency of all process operating facilities.  This means costs that fall within operating budgets, schedules that fall within manufacturing plans, support that falls within the purview of resident control engineers, and technology that falls within native control system capabilities (think function blocks).

One of the most difficult challenges for MPC and all control has been defining appropriate performance criteria.  Classical “error-minimization” has always been the default criteria, because that is the basis on which control algorithms are designed in the first place.  But experience has revealed that industrial process operation places higher priority on “operational” performance, which is conceptually like a set of safe driving habits, such as obeying speed limits and not over-shooting stop signs.  For example, a self-driving car needs to get from NY to LA without incident, not in record time.  For process automation to reach new levels of performance and progress – both single-loop and multivariable – industry must come to grips with the concept of “operational” performance criteria.  Operational performance has the fortuitous virtues of being robust with regard to dynamically changing process models and providing an intuitive “speed limit” tuning parameter.

Today, industry needs to address the current gaps that have emerged and consider alternatives going forward on a priority basis, so that the important business and technology of advanced process control can move forward and avoid becoming stalled or misdirected.  To the extent the journey is just beginning.  Read more at APCperformance.com.