A mixed-integer nonlinear programming approach to structural flowsheet optimization

This thesis addresses the synthesis of chemical processes with an approach based on the use of optimization techniques. Given a flowsheet superstructure which contains a set of potentially attractive flowsheet designs, the synthesis problem is formulated as a large-scale mixed-integer nonlinear programming (MINLP) problem, whose solution yields the structure and operation of the optimal process flowsheet. Several important aspects related to the effective solution of the resulting MINLP problem are investigated.;A two-phase strategy for the OA/ER algorithm is also proposed for the solution of nonconvex MINLP problems in which no assumptions are made concerning the form of nonlinearities. Based on numerical convexity tests, a scheme is developed which systematically modifies the linearizations of nonconvex functions in an attempt to provide valid outer-approximations for the master problem. Although no guarantee of global optimality can be established, it is shown through numerical results that this strategy can often identify the global optimum.;Finally, a modelling/decomposition strategy is developed in order to exploit the special structure of the flowsheet synthesis problem. The modelling strategy minimizes the effects of nonconvexities, and the decomposition of the superstructure leads to the solution of a reduced nonlinear programming (NLP) problem corresponding to the current flowsheet. The nonexisting process units in the superstructure are suboptimized through a Lagrangian decomposition scheme. This modelling/decomposition strategy, which greatly enhances the efficiency of the OA/ER algorithm and the quality of the solutions, is illustrated with the synthesis of the toluene HDA process.;Firstly, an efficient MINLP algorithm is developed by extending the Outer-Approximation algorithm by Duran and Grossman (1986a, 1986b). The Outer-Approximation/Equality-Relaxation (OA/ER) algorithm is proposed which has the important capability of treating explicitly nonlinear equations in the MINLP problem formulation. A computer implementation, DICOPT, is developed in which the OA/ER algorithm is interfaced with state-of-the-art optimization tools and a powerful modelling language. The ability to solve large-scale MINLP problems with DICOPT is demonstrated on a variety of design and synthesis problems.