(163c) Feasibility Mapping and Optimization Using Big Data

With the advent of high-fidelity model-based computer simulations and advanced sensors, more and more data is made available for the determination of feasible operating region and the optimization of complex chemical processes. Specifically, data-driven methods can be advantageous for black-box process optimization, where the analytical forms of the objective function and constraints are either unknown or are difficult to formulate explicitly leading to lack of derivative information. However, the high-volume of data and the high-dimensionality of real industrial problems can introduce significant challenges for data-driven feasibility mapping and optimization. Current data-based approaches are mostly limited to small data sets or samples and perform well only if the problem has a few decision variables. To this end, we propose an efficient method for the approximation of convex or nonconvex, continuous or disjoint feasible regions of constrained black-box problems. Here, we approximate a feasible region as the region described by the convex hull of all feasible samples and their neighboring points while subtracting the infeasible samples and their neighboring points from the convex hull. To address high-dimensional problems, we propose a novel approach that transforms a multi-dimensional problem into a single-dimensional parametric problem over a newly introduced auxiliary variable [1]. In this presentation, we will introduce the convex-hull based feasibility mapping and single-dimensional projection methods using data obtained from high-fidelity process models and discuss the results for the intensification of an integrated carbon capture and conversion process to generate syngas utilizing power plant flue gas and natural gas within a single column.

References:

1. Bajaj, I., Hasan, M.M.F. A Novel Derivative-Free Optimization Method based on Single Dimension Projection. Accepted for publication in the Proceedings of FOCAPO/CPC 2017.