Metal-organic frameworks (MOFs) hold great potential for gas separation and storage, and graph neural networks have proven to be a powerful tool for exploring material structure-property relationships and discovering new materials. Unlike traditional molecular graphs, crystal graphs require consideration of periodic invariance and modes. In addition, MOF structures such as covalent bonds, functional groups, and global structures impact adsorption performance in different ways. However, redundant atomic interactions can disrupt training accuracy, potentially leading to overfitting. In this paper, we propose a multi-scale crystal graph for describing periodic crystal structures, modeling interatomic interactions at different scales while preserving periodicity invariance. We also propose a multi-head attention crystal graph network in multi-scale graphs (MHACGN-MS), which learns structural characteristics by focusing on interatomic interactions at different scales, thereby reducing interference from redundant interactions. Using MOF adsorption for gases as an example, we demonstrate that MHACGN-MS outperforms traditional graph neural networks in predicting multi-component gas adsorption. We also visualize attention scores to validate effective learning and demonstrate the model’s interpretability.