Multi-objective optimization of ANN-based vacuum pressure swing adsorption process for ethane and ethylene separation

A bilevel optimization methodology was developed for separating ethane and ethylene using vacuum pressure swing adsorption. Data generated through Latin hypercube sampling and normalization were employed to construct a neural network at a lower level, serving as a surrogate model for the comprehensive first-principle adsorption process. Following sensitivity analysis based on Monte Carlo simulation, optimization, data resampling, and reconciliation were performed at an upper level. Two cases were performed to optimize the ethane and ethylene separation process. In the first scenario, ethylene recovery was optimized under a purity constraint, resulting in an enhancement from 65.28 % to 87.19 %. In the second scenario, both ethylene recovery and energy consumption were simultaneously optimized with the purity constraint, leading to the generation of a Pareto front. From this Pareto front, two operating conditions were determined: one using TOPSIS and the other aimed at reducing energy consumption from a conventional distillation column to 0.733 MJ/kg-ethylene. Compared to conventional distillation, the vacuum pressure swing adsorption (VPSA) process showed 82.8 % recovery with 0.747 MJ/kg-ethylene and 72.21 % recovery with 0.683 MJ/kg-ethylene. A dynamic analysis and an economic analysis of scaling up VPSA process were performed to compare with C splitter.