(472c) Multiobjective Gasoline Blend Planning and Scheduling

Gasoline blend planning and scheduling over extended time horizon (two or three weeks) is characterized by:

(i) Non-unique solutions.

a. Several different blend patterns can have the same total cost

b. Possibly different allocation of swing tanks for the same blend pattern.

(ii) Discontinuous operation of the blender

a. After each blend, online analyzers need to be recalibrated

(iii) Requirement to blend not less than some pre-specified amount.

(iv) Multiple objectives, e.g.:

a. Minimize total cost

b. Minimize use of the highest octane component

c. Minimize blend recipe deviation from one to the next blend of the same grade.

Our goal is to devise algorithms that enable a blend planner/scheduler to decide among the multitude of possible solutions and select those solutions that fit best with other considerations in the site operation.

Gasoline blending is decomposed into three layers:

1. Search for solutions that have the same values of the objective function(s) and are globally optimal.

a. This includes minimal total cost, minimum number of switches, etc.

2. Optimization of individual blends

a. Every blend of each grade of gasoline, over the time horizon

b. It is not mandatory to blend each grade in each time period and it is also not mandatory to blend only one grade per time period.

3. Management of swing tanks (for components and products) and of product shipping.

At the top level (search for families of solutions with global optimum), we employ differential evolution (DE), which is an evolutionary optimization algorithm capable of finding multiple solutions even when they have the same value of the objective function. Differential evolution maintains a fixed population (number of members in the population is constant) and improves the population mix from one generation to another. Each population member is a solution to the multi-time period blending problem.

Optimization at the second level is carried out by a multi-time period LP. Since each member of the population solves LP for the entire time horizon, we have parallelized DE computations, i.e. each member LP is solved on a separate CPU in parallel.

Finally, at the bottom level one deals with logical and procedural aspects which are often formulated as mixed-integer programming problems. We have chosen to forgo MIP approach and instead are using autonomous agents to solve the details at this level.

The solution methodology described above has been applied to several gasoline blending problems, each of them having different demand patterns, product qualities, and component supply patterns. In all cases, there are multiple blending patterns with different inventories and the same values of the objective function(s). Such results, if available in practice, would provide the blend planner/scheduler with more options, leading to a higher flexibility in operation.