A new model for few-shot knowledge completion was proposed to address the problem that existing knowledge completion models fail to adequately perceive the neighborhood topological structure during knowledge representation. A structure-aware encoder based on an attention mechanism was employed to encode triples in the process of knowledge representation. The cluster partitioning of neighboring nodes was performed by the encoder based on topological structure, and the structural information from each cluster in the neighborhood was integrated to reinforce the model’s structure-aware ability. An adaptive prediction network was adopted to compute the scores of triples to be predicted, for enhancing the stability of the model’s predictions. Experiments in the NELL-One and Wiki-One datasets were carried out. Results show that, compared with the baseline models, the proposed model outperforms by 0.018, 0.021, 0.024, 0.016 and 0.019, 0.055, 0.039, 0.038 in terms of mean reciprocal rank and the hit rate of correct results ranked in the top 10, top 5, and top 1 metrics, respectively. The proposed model effectively leverages topological information from neighborhoods, thereby enhancing the accuracy of knowledge completion.