378797 Refinery Optimization Under Uncertainties

Tuesday, November 18, 2014: 2:05 PM
403 (Hilton Atlanta)
Yu Yang and Paul I. Barton, Department of Chemical Engineering, Massachusetts Institute of Technology, Cambridge, MA

In the refining industry, crude oil procurement is the largest expenditure and these purchases have a large impact on refinery profitability. Moreover, once a specific crude oil is selected and purchased, the refining process must be operated to extract the most value possible from the crude based on the characteristics of the crude and the constraints of the refinery, while ensuring that all refined products meet quality specifications. However, a major challenge in optimizing this system is that the crude qualities are not known accurately at the time of purchase and the refined product market prices may change markedly between the date of crude purchase and the date when that crude is refined. Hence, the goal of this research is to provide a systematic approach to maximize refining profitability in the presence of uncertainties.

  Towards this end, a simplified refinery scheme [1] including crude distillation unit, reformer, cracker, desulfurization and isomerization is studied and modeled as a scenario-based two-stage stochastic programming formulation. In stage I, the best crude oil combination is selected among several candidates and their purchase amounts are determined to maximize the expected gross margin across all scenarios. In stage II, the uncertainties are realized and the optimal operations for the plant are determined according to the real quality of crude, associated with each scenario, such that both the market demand and quality specifications are satisfied. The proposed scheme takes into account uncertainties in the crude quality, nonlinearity from the pooling and unit operation modes, and logic decisions into account, resulting in a large-scale mixed-integer non-linear programming (MINLP) problem. In order to obtain the global optimal solution for this problem efficiently, the variable discretization, feasibility and optimality-based domain reduction techniques are integrated into the non-convex generalized Benders decomposition (NGBD) methodology [2,3], which takes full advantage of the decomposable structure of the scenario-based stochastic programming problem. The results of the stochastic scheme are compared with a deterministic approach to demonstrate the benefits of expected profit maximization and variance reduction. Even though the complexity of this approach is considerably increased due to the multi-scenario and non-convex formulation, the proposed enhanced NGBD approach is still able to find and verify the global optimal solution within a couple of hours, whereas the state-of-the-art commercial software cannot find the global optimal solution in several days. Moreover, the NGBD method exhibits a linear growth in solution time with respect to the number of scenarios.


[1] J. P. Favennec, Refinery operation and management, Editions TECHNIP, Paris, 2001.

[2] X. Li and P. I. Barton, Decomposition strategy for the stochastic pooling problem, Journal of Global Optimization, vol. 54, 2012, pp. 765-790.

[3] X. Li, A. Tomasgard and P. I. Barton, Nonconvex generalized Benders decomposition  for the stochastic separable mixed-integer nonlinear problem, Journal of Optimization Theory and Applications, vol. 151, 2011, pp. 425-454.

Extended Abstract: File Not Uploaded
See more of this Session: Process Design I
See more of this Group/Topical: Computing and Systems Technology Division