SURROGATE MODELING FOR NONLINEAR BLENDING OPERATIONS USING DATA-DRIVEN MIP-BASED MACHINE LEARNING TECHNIQUES

Abstract

The crude oil supply chain stretches widely from production/purchasing of the oil to final goods used by customers in power generation. Importantly, plenty of decisions should be made along with this supply chain management, such as the availabilities of the crude oil mix, the capacity and type of production, and the distribution and storage mechanism. Typically, the supply chain comprises three planning levels: 1) strategic, 2) tactical, and 3) operational planning. Integrating optimization decisions between these levels elevates a nonlinear model with intensive, intractable, large-scale computational characteristics. Therefore, the most practical way to address these decisions is to focus on different planning levels and then integrate them; hence, the motivation for addressing different blending operations is to tackle the operational planning level. In response, high-quality solutions involving process system engineering (PSE) are proposed for recipe optimization and production scheduling. Recent modeling and solving algorithms (MSA) advances in collaboration with high-performance computing capabilities (HPC) are utilized to innovate solutions that could fit into industrial-size problems. The MSA is represented in the utilization of surrogate modeling for predicting the functional behavior of a system using analytical formulations as an alternative to complex models that often lead to non-convergence issues and not sufficiently accurate solutions. The HPC is achieved by the software of IMPL (Industrial Modeling and Programming Language) from Industrial Algorithms Limited. The surrogate model construction and validation procedure addressed in this thesis consist of five major steps applied in nonlinear blending operations. The first is the input (x) data sets generation performed using the technique of Latin Hypercube Sampling (LHS), accompanied with a rescaling strategy, and used for evaluating the output (y) data using the formulation of the first principles. Secondly, the generated data are improved with a normalization procedure to mitigate numerical issues and avoid biased surrogates. Thirdly, mixed-integer quadratic programming (MIQP) is employed, relying on the least-squares regression, to build an optimizable surrogate function for each variable of concern able to substitute the rigorous nonlinear and nonconvex formulation of the first principles. Fourthly, smaller and simpler surrogates are established and selected to be potentially employed. Fifthly, the rigorous and surrogate models both are embedded in a nonlinear (NLP) optimization environment of a single-mode single-period blending operation to validate and compare their performance in terms of convergence, optimality, and computational time. Finally, a base for future work is placed; it compromises two stages. The first stage introduces the linear programming (LP) approximation technique using factor flow for linearizing NLP optimization of a multi-mode multi-period blending operation into a mixed-integer linear (MILP) model. The second stage is fixing the binary solution setup of the optimized MILP.

Date of Award	2022
Original language	American English
Awarding Institution	HBKU College of Science and Engineering