Optimization of pressure swing adsorption in a three-layered bed for hydrogen purification using machine learning model

A novel adsorbent, UTSA-16, was combined with activated carbon and zeolite 5 A to composed a three-layered bed for hydrogen purification. The adsorption isotherms and breakthrough curves of three-layered bed for a typical steam-methane reformer (SMR) off-gas (H₂/CH₄/CO/CO₂ = 76/3/4/17 mol%) were simulated and validated. Pressure swing adsorption (PSA) cycle models were developed and analyzed to investigate the purification performances of three different combinations of adsorbents. The results showed that UTSA-16, position far from the inlet, achieved a significantly higher hydrogen purity of 99.99%, with the hydrogen recovery of 62.08% and a hydrogen productivity of 7.2209 mol/(kg·h). To explore a better optimization solution, this work introduces a heuristic algorithm, specifically the genetic algorithm (GA), to optimize the structure of the back propagation neural network (BPNN). Two machine learning models, BPNN and back propagation neural network-genetic algorithm (BPNN-GA), were used to predict hydrogen production. The results were compared above all, the BPNN model has a test error of 0.0513, which is larger than the BPNN-GA model's error of 0.0173. This indicates that the BPNN-GA model performs better for PSA cycle optimization and prediction. The correlation coefficient (R) between the targets of Aspen model and the predicted outputs are close to 1, which showed good accuracy and performance of BPNN-GA model. In conclusion, the BPNN-GA model allows for precise the determination of the optimal operational parameters for achieving optimal performance in the PSA cycle.