GOALI: Optimization Based Design and Plantwide Control Systems

全厂控制涉及到哪些变量应该被测量、控制和操纵，以及这些变量应该如何联系在一起的问题。对于一些工厂，分散的比例-积分-导数（PID）方法产生可接受的结果，而在其他工厂则不然。目前还没有一种系统的方法可以让设计人员确定一个特定的工厂控制体系结构的预期性能。本研究解决的基本问题是：给定动态过程模型和预期的设定值变化/干扰，分散架构是否足以实现可接受的性能，或者应该对工厂的全部或部分使用交互模型预测控制（MPC）架构？PI开发了两种基于优化的方法来回答这个问题。其中一个涉及最优控制（OC），这是本文计划研究的重点，而另一个使用混合整数线性规划（MILP）公式。所研究的方法假设一个动态过程模型是可用的，因为这种模型在实践中变得越来越普遍。该方法的目标是从流程模型中提取有关控制模型所代表的工厂所需的体系结构类型的信息。PI的OC研究使用初始状态强制，全状态反馈，并且需要为每个候选分散架构调整PID反馈和P前馈控制器。这种OC方法成功地应用于两个杜邦实例。在这项资助下进行的研究使用输出反馈，对设定值或负载进行步进强迫，并且在架构筛选过程中不需要进行PID调整。PI计划将新方法应用于杜邦的两个例子，以及其他几个更大的工厂范围的例子。这些其他的例子包括田纳西伊士曼工艺，醋酸乙烯工艺，可能还有HAD工艺。杜邦公司和阿斯彭技术公司提供的动态仿真软件将用于本研究。杜邦合作者将为所研究的过程提供开发模型预测控制器的协助。

Abstract - McAvoy - 0071320Plantwide control involves the issue of which variables should be measured, controlled, and manipulated, and how these variables should be linked together. For some plants a decentralized proportional-integral-derivative (PID) approach produces acceptable results, while in others it does not. At present there is no systematic methodology that allows a designer to determine how well a particular plantwide control architecture can be expected to perform. The basic question addressed in this research is: given a dynamic process model and expected set point changes/disturbances, is a decentralized architecture sufficient for acceptable performance, or should an interacting model predictive control (MPC) architecture be used for all or part of the plant?The PI has developed two optimization-based approaches to answering this question. One involves optimal control (OC) and it is the focus of the research planned here, while the other makes use of a mixed integer linear programming (MILP) formulation. The approach to be studied assumes that a dynamic process model is available, since such models are becoming more and more common in practice. The objective of the approach is to extract information from the process model about what type of architecture is required to control the plant that the model represents. The PI's OC research used initial state forcing, full state feedback, and it required tuning PID feedback and P feedforward controllers for each candidate decentralized architecture. This OC approach was applied successfully to two DuPont examples. The research to be done under this grant uses output feedback, step forcing for either set points or loads, and no PID tuning is required during the architecture screening process. The PI plans to apply the new methodology to the two DuPont examples, as well as to several other, larger plantwide examples. These other examples include the Tennessee Eastman process, a vinyl acetate process, and possibly an HAD process. Dynamic simulation software has been provided by DuPont and by Aspen Technology and will be used in this research. The DuPont collaborators will provide assistance in developing model predictive controllers for the processes studied.