Multicomponent distillation is a workhorse of current chemical engineering, used on a large scale in the petrochemicals, petroleum and bulk chemicals sectors. Due to its widespread use, any performance improvements developed have a substantial real-life impact. In this work, we focus on reducing the heat duty requirement for multicomponent distillation through finding good configurations of distillation columns. We first explore the search space of distillation column configurations, and classify the search space into categories. We find that some categories have many more configurations than other categories, and grow much more rapidly with the number of feed components. We ask whether there are any categories that will always dominate other categories in our chosen performance measure of low heat duty requirement. We find through extensive computational experiments that indeed the hypothesis is verified, and that the so-called non-sharp configurations almost always require less heat duty than sharp-split configurations. Further, we find that configurations that use n--1 columns for an n-component feed (the socalled basic configurations) require less heat duty than configurations that have more columns (the so-called non-basic configurations), which would imply that basic configurations dominate non-basic configurations in both reduced heat duty and lower capital cost. Subsequently, we formalize the results into a mathematical framework to generate all basic configurations, and nothing but basic configurations, which we call the supernetwork formulation. We solve for minimum heat duty using conventional mathematical programming local search techniques, global search techniques, and stochastic methods. We find that local search over the continuous decisions that describe the parameters for each configuration is currently the most reliable and robust approach, beating out global search techniques in robustness and beating stochastic search in efficiency. Overall, we find that solving many relatively small non-linear programs is better than solving a single mixed-integer non-linear program for the problem of distillation configuration synthesis. We then explore ways to develop heuristics automatically from the results of the optimization through machine learning techniques. Finally, we use our methods to solve three case studies in the broad area of hydrocarbon separation. We find configurations and corresponding molar flow parameters for the recovery of ethylene from cracked naphtha which require 13 to 17% less heat duty overall and up to 17% less methane condensation duty than configurations currently used in industry. We find configurations that reduce the heat duty requirement of atmospheric crude distillation by 17% to 20% depending on the composition of the crude. We also demonstrate that some configurations when used for cyclopentane distillation can reduce the overall heat duty by 70%.
展开▼
