Unsupervised Learning-based Robust Optimization for Ethylene Cracking Furnace Scheduling

Abstract

Machine learning technologies have received great attention in the field of optimization under uncertainty, which learn effective information unsupervised from uncertain data. This work proposes a novel data-driven uncertainty set that uses two machine learning methods and a typical uncertainty set: partial least squares is adopted to decompose the dataset into two subspaces by extracting the relation between data, and then the uncertainties within the divided subspaces are further described by support vector clustering-based and polyhedral uncertainty sets. The proposed data-driven uncertainty set-induced robust optimization framework not only preserve the tractability similar to the classical ones, but also tradeoffs between robustness and optimality well. The final real-world example of cracking furnace scheduling demonstrates the applicability and validity of the proposed framework.