The extrapolation reasoning on Temporal Knowledge Graphs (TKGs) aims to predict unknown future facts, which is crucial to understanding the development trends of significant events. Existing researches have primarily utilized graph-based structural information and LLMs for Temporal Knowledge Graph Reasoning (TKGR), but have often neglected the potential benefits that the latent relationships between entities. To address this, we propose a Discovering Latent Relationship Model (DLRM) for TKGR. Firstly, we propose a fully connected graph architecture named entity-timestamp network (ETN) for modeling latent relationships. Secondly, we use a Fourier convolution layer (FCL) to efficiently encoder ETN and obtain entity features with global semantics. Furthermore, we augment the entity semantics by contrastive learning with the relationship-enhanced TKG embedding. Finally, the score function is used to predict the missing entities. Experiments on four benchmark datasets demonstrate that the proposed model outperforms existing state-of-the-art methods across most metrics, confirming the effectiveness of discovering latent relationships in improving the model’s performance on the task of TKGR.