<p>Knowledge Graph Completion (KGC) usually focuses on predicting existing relationships in the Knowledge Graph (KG), which poses a challenge in predicting previously unseen relationships. This requires models to have the capacity for generalizing from known relationships to invisible relationships. Zero-shot knowledge graph completion methods solve this challenge by predicting the missing relationships in training data. Although existing zero-shot KGC methods mainly focus on static knowledge graphs, most real-world facts show dynamic development and change with time. In addition, existing zero-shot completion models only rely on superficial information, such as entity names or relationship types, without mining and utilizing the deeper meanings and connections behind entities and relationships. In this paper, we propose a method to simulate invisible relationships using Generated Antagonistic Networks (GANs). We map the entities and relations in KG onto hyperplane of time to obtain a structured representation and employ feature encoders to enhance generalization to unseen relationships by decoupling entities from known relationships. Moreover, our framework includes a generator and discriminator to generate and validate simulated relationship vectors. Extensive experiments indicates that our model exceeds the performance of latest methods in zero-shot link prediction task.</p>