Automated structure detection for distributed process optimization

Abstract

The design and control of large-scale engineering systems, consisting of a number of interacting subsystems, is a heavily researched topic with relevance both for industry and academia. This paper presents two methodologies for optimal model-based decomposition, where an optimization problem is decomposed into several smaller sub-problems and subsequently solved by augmented Lagrangian decomposition methods. Large-scale and highly nonlinear problems commonly arise in process optimization, and could greatly benefit from these approaches, as they reduce the storage requirements and computational costs for global optimization. The strategy presented translates the problem into a constraint graph. The first approach uses a heuristic community detection algorithm to identify highly connected clusters in the optimization problem graph representation. The second approach uses a multilevel graph bisection algorithm to find the optimal partition, given a desired number of sub-problems. The partitioned graphs are translated back into decomposed sets of sub-problems with a minimal number of coupling constraints. Results show both of these methods can be used as efficient frameworks to decompose optimization problems in linear time, in comparison to traditional methods which require polynomial time.