Knowledge graph, which stores an extensive number of real-world facts as graph-structured data is an extremely important knowledge base for artificial intelligence. Among all kinds of knowledge graph embedding methods, distance-based models are most widely adopted in practice and show promising performance due to its effectiveness, simplicity and interpretability. However, distance-based models face two major challenges and almost all methods proposed so far fail to solve them simultaneously. One is the ability to handle complex relations and the other is to encode various relation patterns. This work proposes a new translation-based model called QuatRE, which uses a tripled vector and a learnable parameter relation radius to represent relations. The tripled vector consists of two rotation vectors and a translation vector. Two rotation vectors interact with head and tail entities to extract their semantic information in specific relations. The learnable parameter relation radius transforms the translated entities from dots into hyperspheres to enhance the ability of modeling complex relations. Experimental results show that our proposed method can both model complex relations and various relation patterns effectively.