Generation of reduced statistical models for NLP and MINLP optimization

Abstract

A general strategy for the generation of statistical models with a reduced complexity of the mathematical formulation, in order to handle NLP and MINLP optimization problems efficiently, is proposed. Resulting from increasing complexity of structural and parameter optimization problems in chemical engineering, we are very often, especially when dealing with highly nonlinear problems, confronted with the situation that the currently available solvers are not able to cope with the problem size or require very good starting points. In many cases the utilization of reduced statistical models can help to overcome these difficulties by enabling a better insight into the optimization problem and providing sufficient initial guesses, with respect to the process structure and the unit parameters. In order to fulfill the entire optimization task it is desirable that the reduced models also consider the influence of significant parameters. The generation of such reduced statistical models on the basis of associated rigorous models is described and examined by using the commercial flowsheeting package SPEEDUP.