Computational-cost-efficient surrogate model of vacuum pressure swing adsorption for CO separation process optimization

• Published in: Separation and purification technology Vol. 300; p. 121827
• Main Authors: Kim, Jinsu, Son, Manwoo, Sup Han, Sang, Yoon, Young-Seek, Oh, Hyunmin
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.11.2022
• Subjects: CO Separation; Machine-learning; Optimization; Surrogate modeling; Vacuum pressure swing adsorption; SVD; ANN; CO Separation; NP; BFG; EDA; GPR; TPE; SVM; Surrogate modeling; BO; Optimization; LDG; DT; PLS; EBT; COG; DOE; MSE; PSA; ML; CSS; SHAP; POD; Vacuum pressure swing adsorption; RMSE; PSO; SLPM; MAE; NSGA; Machine-learning; PDAE
• ISSN: 1383-5866; 1873-3794
• DOI: 10.1016/j.seppur.2022.121827

•A surrogate model of the VPSA for CO separation was made for process optimization.•Purity and recovery were predicted through the full factorial design of 1.98 K data.•The optimal condition is 3.94 bar for the adsorption and 0.05 bar for the desorption.•The sensitivity result of electricity and CO price shows the efficacy of the model.•Pareto-front between productivity and purity was shown. We present a computational-cost-efficient surrogate model to optimize the separation of carbon monoxide (CO) from steel-mill off-gas when adsorbent material and the vacuum pressure swing adsorption process are used. We use a mathematical model to generate the full-factorial design data points, then use them to train, validate, and test the surrogate model, which we then use to suggest the best prediction algorithm for CO separation. The extra-tree regressor was selected by considering the test R2 score, and purity and recovery were well predicted (R2 = 0.999 for both). The optimized condition was 3.94 bar for the adsorption, and 0.05 bar for the desorption, which gave 99.99% purity and 71.3% recovery. Analysis of cost sensitivity of electricity price and CO selling price showed how the optimal conditions were moved from the base case; these results were consistent with the phenomenological knowledge which favors a large difference in pressure between the adsorption and desorption steps. Also, Pareto-front solutions of the productivity and purity implied that compared to a previous study, the productivity can be increased by 9.4% when the target purity is 90%, and by 7.5% when it is 99%.