(564b) Initialization Strategies for Global Optimization in Refinery Planning

Refinery planning is the problem of determining the optimal crude slate for a refinery in order to meet all product demands and specifications while satisfying all operational constraints.  Refinery planning is typically represented as a generalized pooling problem resulting in a mixed-integer, nonlinear optimization problem.  The very large number of bilinear and trilinear terms can lead to local optima, which can translate into millions or even billions of lost revenues for the refinery operators over the course of a year.  Despite recent developments in global optimization, many of the new techniques are impractical due to problem size and complexity as well as the solution time constraints present in a refinery environment.  Commonly used techniques (such as successive linear programming, SLP, and successive quadratic programming, SQP) require that all variables be initialized prior to optimization. 

In this paper we explore a convex relaxation-based generation of an initial solution for the refinery planning problem.  The algorithm uses recent enhancements in global optimization techniques applied to a large-scale industrial problem while employing specialized bounding algorithms to further improve accuracy and performance.  Combined with other techniques and heuristics, the end result is the creation of a consistent initialization strategy to help reduce the occurrence of sub-optimal solutions in refinery planning problems.