GOALI: Performance Monitoring Principles for Nonlinear and Linear Model Predictive Control

点击翻译按钮获取中文摘要

Current estimates of the economic impact in the process industries of model predictive control (MPC) are in the billions of dollars per year. But the process industries currently have no accepted methods tailored to the problem of assessing the performance of the thousands of loops currently running under model predictive control.The goal of this project is to establish principles for performing such assessment for installed MPC technology. Establishing these principles will enable the development of software tools to automatically quantify the key performance index (KPI) for each MPC control system under study. Rank ordering of the different loops' KPIs enables comparison, prioritization, and pinpoints those systems requiring further attention. Comparisons of each loop's actual KPI to three other calculated KPIs allow further diagnosis including: impact of process constraints, poor process models, inappropriate noise filtering, and large disturbances. Longitudinal analysis allows the user to select appropriate time periods for benchmarking performance and spotting seasonal trends and variations. Feedback from this assessment analysis to the control system technology vendors creates incentives and points to promising areas for further improving the control technology.There is little established theory for assessing performance of MPC based on nonlinear models, so this part of the research ventures into a completely new area. The industrial partner in this GOALI proposal, ExxonMobil, will provide the relevant industrial data sets and supervise the supported Ph.D. student for two summers at ExxonMobil.Intellectual Merit:The problems to be addressed in this research are fundamental and far reaching. The required theory draws upon: optimization; probability, statistics and statistical inference; random variables and Markov chains; nonlinear control theory; numerical simulation and time-averaging of numerical simulations to determine statistical properties (Monte Carlo methods). The theory draws from current techniques developed in these fundamental disciplines and will expose new problems that cannot be addressed with the currenttools provided by these fields.Broader Impact:Advance technology has been acquired and implemented, but users have no clearinghouse of objective data and information telling them where these systems are working well, and perhaps more importantly, where they are not working well. Developing new systems to provide this information will enable: (i) users to systematically manage larger installations, (ii) identify and prioritize the troublesome parts of their systems, and (iii) spur development of new control theory tailored to addressing the identified, unmet practical needs. The people responsible for implementing and maintaining advanced technology require management tools to assess the technology. This is true of almost any advanced technology.The systems theory and algorithms developed in this research are completely general and can be applied to any class of dynamic manufacturing processes for which a reasonable predictive model can be developed.