In recent years, the representation learning and embedding methods of knowledge graph have become a robust paradigm to solve link prediction problems of knowledge graph. With the continuous development of this field, hyper-relation, consisting of a main triple and several qualified key-value pairs, has become the most commonly used knowledge representation now. Most embedding models conduct representation learning and feature extraction of hyper-relations under the background of single receptive field. However, without using multiple receptive fields to cross and fuse information, they are always limited in practical application fields, and a single receptive field will greatly limit the feature extraction and link prediction ability of the model. To solve this problem, we propose the MRF3 model, a Multi-Receptive-Field Feature Fusion knowledge graph embedding model. MRF3 model uses receptive fields of different sizes to capture semantic information of different scales under the representation of triples and hyper-relations respectively, in order to extract features and complete link prediction tasks. The experimental results show that our MRF3 model has achieved good performances on multiple datasets and baselines. In entity and relation prediction, MRR index improves by 6.8% and 9.5% on average respectively, which verifies the effectiveness and superiority of the proposed model.