Towards Reliable and Efficient Real-Time Optimization of Processing Plants

随着全球竞争的加剧，流程工业面临着提高生产效率、产品质量和工艺安全的巨大压力。因此，实时优化(RTO)被广泛用于工厂(无论是单个机组还是更大的工厂的一部分)的运行优化。RTO是一种基于模型的上层控制系统，它反复向下层控制系统提供设定值，目的是使过程运行尽可能接近经济最优。RTO系统提供了高层计划和调度与监管控制之间的联系。RTO通常是在严格的、非线性过程模型的基础上进行的。然而，模型永远不会准确地反映过程的真实行为，因此通常收敛于模型最优的优化对于实际对象来说不是最优的；此外，计算的运行点可能违反约束。在过去的十年中，已经发展了几种方法来解决这一问题，其中包括参数自适应、梯度校正(称为修改器自适应)或仅使用观察到的植物行为进行直接搜索。这些方法都有一定的缺陷和局限性，特别是不能保证每次迭代都满足约束条件或收敛速度慢。该项目研究了修改和组合不同方法的方法，以提出改进的基于数据和模型的RTO方案。目标是开发一种RTO方案，通过使用实际精度的模型(即，不需要过多的建模工作)和测量数据来实现快速收敛到真正的工厂最优。

With increasing global competition, the process industry faces intense pressure to improve production efficiency, product quality and process safety. As a result, real-time optimization (RTO) is used extensively for the operational optimization of the plant (either a single unit or part of a larger plant). RTO is a model-based upper-level control system that repeatedly provides set points to the lower-level control system with the objective to maintain process operation as close as possible to the economic optimum. The RTO system provides a link between high-level planning and scheduling and regulatory control. RTO is usually performed on the basis of a rigorous, nonlinear process model. However, the model will never represent the true behavior of the process exactly, and so the optimization, which typically converges to the model optimum, will not be optimal for the real plant; in addition, the computed operating point may violate the constraints. Several approaches have been developed in the past decade to cope with this problem, which include parameter adaptation, gradient correction (called modifier adaptation) or direct search using only the observed plant behavior. These approaches all have certain drawbacks and limitations, in particular it cannot be guaranteed that the constraints are met at each iteration or convergence is slow. This project investigates ways of modifying and combining different approaches to come up with an improved data- and model-based RTO scheme. The goal is the development of an RTO scheme that implements fast convergence to the true plant optimum through the use of a model of realistic accuracy (i.e. without excessive effort for model building) and measured data.