Scheduling of multipurpose batch chemical plants

Athanasios Tsirukis, Purdue University

Abstract

The highly combinatorial and nonlinear nature of the multipurpose batch plant scheduling problem significantly limits the application of exact mathematical formulations. We introduce a solution strategy that combines approximate reasoning with reduced dimensionality mathematical formulations. The scheduling problem is decomposed into two levels: the campaign formation subproblem, and the detailed scheduling subproblem. In the first level, the products are organized in sequential groups (campaigns). Within each campaign, product members are processed in parallel. A reduced dimensionality formulation, which relaxes the integral nature of equipments, but reflects the competition among different processing tasks for the same limiting resources, is introduced to create efficient campaign structures. In the second level, the solution effort is focused on the details of each campaign. The problem of assigning the plant resources (equipments, water, electricity, etc.) to production tasks to be processed in parallel is highly nonlinear and combinatorial. A general-purpose feature extraction method was devised, that generates information about the topology of multidimensional nonlinear domains, implicitly defined by a set of nonlinear constraints. This approach evaluates the impact of selected decisions on the objective function, without requiring the complete instantiation of the decision vector, a requirement of point-oriented optimization algorithms. A formal search component directs the characterization effort towards promising regions of the decision space. The proposed approach enables the use of genetic algorithms in continuous domains. It is thus possible to exploit the attractive features of genetic search. The feature extraction algorithm proposes a reduced-volume region within which optimal scheduling decisions lie and which is further explored by a conventional optimization algorithm. The requirements for a robust mathematical programming method, able to search significantly nonlinear and combinatorial domains, led to the discovery of Generalized Hopfield Networks, which map all the important optimization algorithms on a massively parallel network of processors. Examples of large combinatorial complexity confirm the efficiency of the overall approach.

Degree

Ph.D.

Advisors

Reklaitis, Purdue University.

Subject Area

Chemical engineering|Computer science

COinS