(175a) A Framework for Optimization-Based Design of Heat-Integrated Crude Oil Distillation Units Using a Surrogate Model and Support Vector Machine

Global demand for petroleum-based fuels for industrial, transportation and domestic use has been increasing since inception. These fuels (gasoline, kerosene, diesel, liquefied petroleum gas, fuel oil, etc.) are produced in refineries via crude oil distillation. Crude oil distillation is a complex and capital-intensive process; the crude oil distillation system comprises distillation columns, pump-around loops, side-strippers and a heat recovery system, known as the preheat train. The large number of degrees of freedom (column structure and operating conditions) and complex interactions between individual units makes the design of crude oil distillation unit a highly challenging task. Rigorous tray-by-tray column models are useful for designing crude oil distillation columns, but the approach requires trial and error, or, if optimization is applied, good initial estimates to obtain convergence and significant computational time and poses numerical difficulties.

Surrogate models, including those based on artificial neural networks (ANNs), have enabled significant capabilities in modelling and optimization of various chemical processes, including crude oil distillation. Researchers have shown that artificial neural networks can provide simple, accurate and robust simulation models. Artificial neural network-based distillation models have been implemented in optimization frameworks to perform operational optimization of existing crude oil distillation unit. To date, however, surrogate models for crude oil distillation reported in the literature have not considered design of the distillation column, i.e. optimization-based design, especially considering discrete variables representing the number of trays in each column section.

This paper presents a novel systematic approach for the design of crude oil distillation units that combines artificial neural network-based distillation models and a support vector machine (SVM). The support vector machine aims to ensure that only feasible design alternatives – i.e. those that meet constraints related to product specifications and hydraulic performance – are considered during the search process. This filtering helps to produce an optimal solution that is feasible; it also reduces computational effort by limiting the search to points within the demarcated feasible region.

The proposed framework comprises five main steps. First, an initial feasible design of a crude oil distillation unit is built using a rigorous tray-by-tray model in Aspen HYSYS. In step 2, the most important degrees of freedom are identified and represented using probability distribution functions. Latin hypercube sampling is then applied on these distributions to generate samples (design inputs) that will be used in the calculations. In step 3, all the samples are simulated and the corresponding responses (design outputs) are collected and automatically separated into feasible and infeasible solutions. In step 4, feasible samples are used to build a surrogate (ANN) distillation column model, while the entire set of samples is used to build a support vector machine. In the final step, the resulting models are implemented in an optimization framework, in which a genetic algorithm is used to search for design alternatives that minimize the total annualized cost (the sum of annualized capital cost and annual operating cost). Pinch analysis is used within the design procedure to determine minimum utility targets and estimate the utility costs (i.e. operating costs). The capabilities of the approach are illustrated through its application to an industrially-relevant case study.