Nonlinear Models for Crude Distillation Units and their Application to Refinery Planning Optimization

This thesis is concerned with the study and development of nonlinear models for refinery production planning. The emphasis is on the front end of refineries, namely the CDU (crude distillation unit) for developing nonlinear process models suitable for optimization problems. We start with the established optimization approaches for refinery production planning, namely the fixed-yield model and swing cut models, which are based on linear programming (LP). After describing the problem statement and various information and inputs into the planning models, we present these LP models and describe the benefits of the improvement the swing cut model provides over the fixed yield model. The comparison provides the clear incentive to capture the nonlinearities of the process units.;We then present the first nonlinear programming (NLP) model, based on the fractionation index (FI) model for the CDU. We develop the model as a series of fractionation units. In contrast to current nonlinear CDU models, the Fl model does not require site or simulation data to generate the equations for yield and properties calculations. The FI model depends on the characteristics of the column. We illustrate the robustness, simplicity and speed of the FI model with several examples. We then show the suitability of the Fl model for the NLP refinery planning model in a series of runs, comparing the results of the LP and NLP models in terms of economics, feed purchase decision and impact on the refinery products and economics. It is shown that the NLP model can predict significant economic improvements, while being comparable in computational effort and robustness to the LP models.;Following the successful development and implementation of the FI model, we expand the scope and application of the model to address multiperiod planning problems. We propose a mixed-integer nonlinear programming (MINLP) model that selects the crudes to be produced in each time period, their sequence and the amounts of products to be produced. The proposed model uses the traveling salesman constraints to generate the sequence at each time period. It is shown that the MINLP model is robust and computationally efficient, while having the capability of addressing complex planning problems.;Finally, we propose an alternative nonlinear model for the CDU that is more rigorous. The semi-rigorous aggregate model provides simplifying approaches that are less demanding than the detailed tray-by-tray calculations. The complexity of the CDU requires decomposing the column into a series of mixed-type columns. Our proposed model uses some of the CPM (column profile map) variables, but can handle conventional and steam distillation columns and their cascade. We compare the aggregate model to the rigorous methods showing that the accuracy is similar to the tray by tray models. However, the corresponding NLP model was less robust and slower than the FI model. We propose further studies of the challenges of applying the aggregate model.;We conclude the thesis with a list of findings and contributions.