With the rapid development of the Internet, data with time information shows an explosive growth. However, the traditional knowledge graphs don’t consider the temporal information of facts, which brings inconvenience to downstream tasks, so people proposed the temporal knowledge graph. The temporal knowledge graph models the timeliness of facts in the form of quadruples, which clearly represents the evolution process of knowledge.But it is difficult for temporal knowledge graphs to cover the full of worldwide knowledge. Therefore, predicting missing facts is necessary, including complete historical knowledge and predict future knowledge. However, the existing reasoning models only considered the one-way information transmission from the front to the back in the time dimension, and didn’t effectively use the temporal characteristics of knowledge. And the treatment of the unseen entities is too rough when predicting the future knowledge. To solve above problems, we proposed a new temporal knowledge graph reasoning model with temporal bidirectional correlation, which models the factual correlation and temporal relationship between entities in a spatial-temporal graph jointly. We also considered the influence of bidirectional information in the time dimension to further enrich the characteristic representation of entities. As for the unseen entities in future knowledge prediction, we proposed a new method to get the embedding of unseen entities, by analyzing its co-occurrence entity set to capture essential common features and generate the embedding vectors. For the above two studies, we conducted experiments on several public datasets, and got better effects.