(248k) Black and Grey Box Modeling of Hydrocarbon Physiochemical Properties

This research explores advanced modeling to predict physical properties of hydrocarbon solutions. The underlying context is the utilization of gas-to-liquid (GTL) technologies in global markets. This research is specifically interested in GTL synthetic kerosene. Synthetic kerosene fit modern agendas, such as reduced environmental impact and sheer abundance of the resource. That being said, GTL kerosene has distinct deficiencies in physical properties that aviation fuels must exhibit (ASTM D1655 standards). Currently, its best utilization is the blending of GTL and crude oil kerosene, in order to meet these standards.

This endeavor seeks to elucidate blends that maximize the usage of synthetic kerosene, while still meeting ASTM D1655 standards. This is accomplished through the study of hydrocarbon solutions on a physiochemical level. Techniques are used to find models with predictive power for physical properties, based on hydrocarbon composition. But the physical property regression trends are highly nonlinear; they transcend the efficacy of traditional regression methods.  For this reason, advanced algorithms are utilized.

Three robust modeling techniques were performed and analyzed: artificial neural networks (ANN), support vector machine (SVM) and Kriging interpolation. These powerful methods were chosen because of their respective and distinct strengths. Models were constructed in R, the statistical programming language, with the packages neuralnet, e1071 and, DiceKriging, respectively.

The experimental training data, used for regression techniques, are solutions of: cyclo-paraffins, n-paraffins, iso-paraffins and, aromatic hydrocarbons. Specifically, these solutions are blends of n-decane, solT, decalin, and toluene. solT is a proprietary blend of iso-paraffins, distributed by Shell^® corporation. Physical properties measured were freezing point, density, heat content, and flashpoint.

Each technique gives relatively accurate results, but they can be further optimized. ANN and SVM, both within the realm of machine learning, are intrinsically strong regression techniques. For this reason, Monte Carlo optimization was implemented. This technique provides increased performance, albeit with long run times.

Kriging modeling is unique to ANN and SVM from a fundamental standpoint; it is an interpolation method. The optimization of Kriging models do not require Monte Carlo methods of the same scale, which drastically reduces run times with virtually no compromises in accuracy.

With the optimized predictive models established, it was applied to the overarching context: maximizing the presence of synthetic fuel in a crude oil blend that still meets ASTM standards. Each physical property was also visualized with a three dimensional ternary diagram. These results show physical property trends of hydrocarbons, establishes algorithmic code to calculate them, and elucidates maximized synthetic-conventional kerosene blends for its future applicability in the aviation industry.