A 4-bed-8-step pressure swing adsorption (PSA) process has been developed to produce high-purity hydrogen from the steam methane reforming (SMR) gas mixture. The Detailed models have been established for hydrogen purification based on the experimentally determined parameters. Two surrogate models are investigated to optimize the process performance using artificial neural networks (ANN), which have been well trained by the samples, obtaining from the Detailed models using Latin hypercube sampling strategy. The results indicate that ANNs could approximate the performance and dynamic behavior of PSA process with extremely high accuracy. Herein, a robust and fast multi-objective optimization approach of PSA process using genetic algorithm on the basis of different ANN-based surrogate models has also been proposed, in which Dual- and Tri-objective optimizations are taken into account. This research shows that the method can not only find out the optimal operating conditions of the PSA process for hydrogen production with higher than 99% accuracy, namely Pareto-Optimal Fronts, but also provide a reliable reference for operational enhancement.