Existing methods for temporal knowledge graph completion (TKGC) often fail to effectively distinguish the latent temporal associations and sequential dependencies inherent in the temporal order of facts and have largely overlooked the impact of time intervals on the interaction information. To address these limitations, we propose a TKGC model HFTS that distinguishes between Historical and Future information and is sensitive to Time Spans. This model divides temporal knowledge graph subgraphs into historical and future temporal subgraphs, aiming to capture order-dependent associations among facts. We also introduce a time-sensitive factor to measure the effectiveness of both entity and relation interactions across varying time-span. Considering that entities typically change smoothly and slowly over time, exhibiting relatively static behavior in the short term compared to the long term, we further introduce a Smooth Self-Supervised Loss (S-SSL) to enforce temporal consistency in entity representations across adjacent timestamps. Experimental results demonstrate that HFTS achieves superior MRR performance on three benchmark datasets, improving by 12.70%, 12.33%, and 6.20% over existing methods. The code of our method has been released in https://github.com/yeezyzy/HFTS.git .