Knowledge graph embedding (KGE) is a simple and effective method for knowledge graph completion. The higher the embedding dimension, the higher the model performance. However, the increase in embedding dimension leads to a significant rise in storage and computation. In this paper, we propose CombDE, a knowledge distillation framework for KGE models to address this problem. First, CombDE introduces self-distillation, so that the student model could retain more performance of the teacher model in each distillation. Secondly, considering the influence of unreliable knowledge transfer on distillation performance, a dynamic adjustment mechanism of the soft label is proposed, which dynamically allocated weights through dynamic distillation temperature and truth value judgment to optimize distillation. Experiments show that our CombDE achieves optimal distillation performance compared to the baseline model. Meanwhile, on average, CombDE can reduce the training time by more than 70%.