Molecular-level kinetic modeling of conventional and unconventional hydroprocessing feedstocks
Date
2019
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Hydroprocessing is a catalytic upgrading process in a hydrogen-rich environment that is commonly used to remove impurities, saturate carbon-carbon bonds, and sometimes break carbon-carbon bonds. Kinetic models are essential in the optimization of hydroprocessing reactor systems to meet strict environmental and product yield specifications. Traditional lumped kinetic models are often feedstock dependent and limited in their prediction capability. Molecular-level kinetic models can address those drawbacks by considering the fundamental chemistry and kinetics of the process. In this work, molecular-level kinetic models were developed for triglyceride and vacuum gas oil hydroprocessing. Model development parallelly also facilitated the improvement of the model building tools to accurately capture the real process and improve the user experience. ☐ First, a molecular-level kinetic model was developed for triglyceride hydroprocessing. Triglycerides representative of the types present in coconut oil and soybean oil were defined using 8 to 22 carbon fatty acids. The triglycerides acted as seeds to a reaction network detailing three parallel deoxygenation pathways: hydrodeoxygenation, decarboxylation, and decarbonylation. Isomerization, cyclization, aromatization, and cracking reactions were iteratively added to the reaction network. The final network contained 476 species and 1709 reactions. Using the network, material balances were written for the kinetic model. The kinetic parameters were optimized based on experimental data over a range of temperatures, pressures, and catalyst contact times. The final kinetic model simulated properties had good agreement with experimental values. To evaluate the end-use value of the diesel product, cetane number and cloud point property models were constructed and optimized based on experimental data. These property models were used to study the product diesel cetane number versus cloud point tradeoff to determine the end-use properties of the product fuel. ☐ Then, a molecular-level kinetic model was constructed for a vacuum gas oil hydroprocessing unit. The experimental data were from an operating refinery unit over the two-year catalyst life. Feedstock molecules containing up to 45 carbons of paraffinic, olefinic, naphthenic, and aromatic type were selected to represent the molecular composition. On those molecules, the typical desulfurization, denitrogenation, saturation, cracking, ring opening, and isomerization reactions were applied. The final network included 5747 reactions and 1532 species. The species feedstock concentrations were determined by sampling the probabilities of the presence of different structural attributes using probability density functions (PDFs). PDF parameters were optimized using a simulated annealing algorithm. To minimize the optimization burden of the PDF parameters, a library containing 21 sets of PDF parameters was created and used to determine the starting point of optimization for each new dataset. For the kinetic model, the reactor system was represented as a series of 19 pseudo-PFRs. The pseudo-PFRs were individual catalyst layers with side-by-side reaction and vapor-liquid equilibrium. Quantitative structure/reactivity correlations and linear free-energy relationships (LFERs) were used to reduce the number of kinetic parameters. The activity of each type of catalyst was modeled independently using the catalyst LFER concept. After optimization using a simulated annealing algorithm, the model showed good agreement with the experimental measurements. ☐ Next, a strategy to generate data-driven models from molecular-level kinetic models was evaluated to greatly reduce the time required for a prediction. The triglyceride hydroprocessing model was used to generate 20,000 datasets for a small ranges of input parameters. For each dataset, the calculated cetane number, cloud point, and yield were recorded. As an initial approach, multilinear regression, decision tree regression, gradient boosting regression, and artificial neural network models were generated from the data. All data-driven models were able to predict results accurately and very quickly (<<1 second), but they only worked in the small ranges of the underlying data. As soon as the inputs exceeded the input parameter ranges, data-driven model predictions greatly diverged from the kinetic model results. In terms of data requirements, multilinear regression models needed much less data than the decision tree regression and artificial neural network models at the cost of some accuracy if the model input parameter range was too wide. ☐ Finally, a feed concentration modeling tool, the Initial Condition Generator (ICG) was developed. The tool was designed to allow users to import a list of molecules, define the PDF structures, input the experimental data, and optimize the kinetic parameters. Special attention was paid to reducing the user requirement of computer science or kinetic modeling expertise. The tool was a direct outcome of the identified need of a simplified feedstock composition model compared to the previous modeling framework. Thus, the need for parallel model development and model builder development to improve future model development can be seen.
Description
Keywords
Hydroprocessing, Kinetic modeling, Machine learning, Molecular-level