A vision for MPC performance maintenance

[摘要] Model predictive control (MPC) is an advanced control that has found widespread use in industries, particularly in process industries like oil refining and petrochemicals. Although the basic premise behind MPC is easy to comprehend, its inner workings are still generally viewed or regarded as too advanced for actual plant operator understanding. This lack of understanding is exposed when MPC performance deteriorates sometime after commissioning, as is often the case in some commercially operated process plants. Currently operators might have difficulty over reasoning about MPC performance degradation and formulating steps to investigate the cause.A tool is described that aims to make MPC more transparent to the operators. Commonly reported causes of MPC performance degradation are discussed and ways in which the operator can recognise them when they occur are outlined. Issues that are addressed include: making the set of controlled variables to be used for a given set of manipulated variables simpler and clearer; ways to recognise when a MPC controller is performing poorly and to identify the source of performance deterioration. An aim is to determine under what instances the operator can return the MPC performance to previous levels or request for specialist support or simply switch the MPC off. A goal is to avoid the kind of often reported situation where the operator gets worried that the controller is deteriorating and ends up taking knee jerk actions that cause further problems in MPC.At the top of the maintenance tool hierarchy is the trends comparison group, where MPC reference graphical performance trends are compared with actual graphical performance trends counterpart. If any abnormality is observed in these trends, the operator is encouraged to choose an option from a list of preliminary diagnostic questions contained in a group below trends comparison group, which best describes the observed abnormality. Each abnormality is associated with a list of suspected causes. When a suspected cause is chosen from the associated list, the operator is led to the symptoms investigation window, which contains scripts detailing steps for systematic examination of each symptom, with a view to either rejecting or confirming the suspicion. Assisted in the investigation are four background information windows: the virtual plant without MPC window, the virtual plant with MPC window, the transfer function matrix window and steady state gain, relative gain array (RGA) and relative weight array (RWA) window. The windows contain information and guidance that the operator might refer to from time to time during symptom investigation.Development of the maintenance tool is still at the design stage. The key components described have been research implementing MPC on three nonlinear process models, a continuous stirred tank reactor (CSTR), an evaporator process and a fluid catalytic cracking unit (FCCU). Case studies representing different MPC degradation scenarios are simulated, followed by a systematic procedure for diagnosing, isolating and recovering from such degradation, based on assumed operator’s perspective and expert’s technical reasoning. The knowledge gained from the case studies is used to develop an outline of a vision for a data-driven model predictive maintenance tool to help the operator make sensible judgements about performance degradation, the form and direction of diagnosis and fault isolation, and possibly, the recovery procedure.

[发布机构] University:University of Glasgow;Department:School of Engineering

[关键词] Model Predictive Control, MPC Performance Maintenance, MPC Maintenance Tool