Temporal Knowledge Graph Completion (TKGC) aims to discover missing factual knowledge from a given temporal knowledge graph. Previous works usually design Temporal Knowledge Graph Embedding (TKGE) models for the TKGC task. However, we have observed that the same relations can connect different timestamps and the embeddings of the same relation at different timestamps are similar. This poses challenges for TKGE in accurately predicting the entities linked by the same relations at different timestamps, especially when the given entities are similar or even the same. To tackle this challenge, we propose a Temporal Multi-grade Multivector Embeddings (TeMME) method, which contains a new score function with One-grade multivector and Two-grade multivector to learn temporal sensitive knowledge graph embeddings. Experimental results show that TeMME has significantly outperformed SOTA models in TKGC tasks. The MRR metrics in each of the four widely used benchmark datasets achieved a significant improvement: ICEWS14 (6.9%), ICEWS05-15 (6.4%), YAGO11k (4.5%), and Wikidata12k (8.3%).