Mnemonic Enhancement Optimization For Process Systems

Real-time optimization (RTO) for process systems is important to industrial energy economization and profit increase. The real-time performance is a key criterion to evaluate RTO implementation. Thus, the improvement of the performance attracts plenty of research interests. There are fruitful productions concerning large-scale real-time optimization algorithms. However, as the system scale keeps growing and the process models tends to include more meso-scale and micro-scale details, the real-time performance remains a bottleneck of RTO. The purpose of this work aims at improving the real-time performance based on the repetitive feature of RTO problem sequence.This work gives a detailed analysis of RTO systems, and proposes the mnemonic enhancement optimization (MEO) method on the basis of the similarity among the RTO problems. This method views the operational conditions and the various disturbances as the parameters of optimization problems. It preserves previous successful solutions as experience, and then uses the experience to approximate the optimum to be solved. The approximation is then sent to the optimization algorithm as the starting point to locate the real optimum precisely. Theoretical analysis shows that, generally, the optimum is a continuous function of the parameters, based on which the MEO approximation is proved to converge to the real optimum with probability 1 as RTO continues to run.In this work, the MEO framework is set up. It mainly consists of two parts: the management of the experience database and the multivariate approximation methods of MEO. The threshold of the experience database is proposed to limit the space complexity. At the same time, an incremental Delaunay triangulation for arbitrary dimensional space is constructed. Sequentially, the nested node selection method is introduced, which builds a solid foundation for the node selection for MEO approximation.The MEO method with zero-order approximation is proposed first. Based on it, the barycentric interpolation method is introduced into MEO, which brings the MEO with the first-order composite approximation method. The natural expansions of the first-order composite approximation lead to the MEO methods with full-space and partial-space multivariate Lagrange interpolations of any order. Finally, as a salutary attempt besides interpolation methods, the linear fitting method is introduced into MEO and the feasibility of higher-order fitting methods is discussed. Numerical experiments show that all proposed MEO methods outperform the traditional method in RTO. And these MEO methods have different characteristics and are suitable for different situations.