Surrogate-based optimization of dual-reflux pressure swing adsorption for biogas upgrading

Abstract This study presents the optimization of a 2-bed, 6-step dual-reflux pressure swing adsorption (DR-PSA) system for biogas upgrading using artificial neural network (ANN) surrogate-assisted multi-objective approach. The ANN models were trained using datasets generated from detailed models in Aspen Adsorption, employing the Box-Behnken design of experiment based on three input variables: feed/light reflux (FE/LR) time, light reflux ratio, and bed column height. The ANN surrogate models were developed using various optimizer algorithms, activation functions, and neuron’s number in a hidden layer to identify the most accurate model for predicting DR-PSA performance indicators: CH 4 purity, CH 4 recovery, bed productivity, and cycle work. To identify optimal operating conditions, the non-dominated sorting genetic algorithm-II (NSGA-II) was employed, generating Pareto frontiers for two optimization cases: unconstrained purity/productivity and constrained productivity/cycle work. The optimization process incorporated the accurate surrogate models to evaluate the objective functions with the same three decision variables used for ANN model construction. The Pareto frontiers for maximizing CH 4 purity in LP/bed productivity revealed a sharp knee region, indicating a significant trade-off between these objectives. For constrained bed productivity/cycle work, separate Pareto frontiers were observed at different constraint levels of CH 4 purity in LP (91%, 91.5%, and 92%), with feasible solutions narrowing at higher purity constraints.