<p>In the field of artificial intelligence, Knowledge Graph Embedding (KGE) has achieved remarkable progress. However, since knowledge is inherently time-sensitive, Temporal Knowledge Graph Embedding (TKGE) has been introduced to better capture the dynamic evolution of knowledge. Existing TKGE methods, nevertheless, face two major challenges: (1) Temporal variations are often simplistically associated with entities, limiting the ability to model diverse static and dynamic temporal patterns. (2) The semantic differences between entities and relations are frequently overlooked, restricting the performance of link prediction tasks. To address these challenges, this paper proposes FLS, a novel framework featuring a dual-stream architecture that decouples temporal dynamics from static semantics. In the temporal stream, relations are projected onto a helical temporal axis within a fusion space to explicitly model their temporal evolution. Meanwhile, in the semantic stream, commonsense knowledge with temporal factors removed is embedded onto a smooth Lie group to mitigate semantic discrepancies. This dedicated approach enables FLS to effectively capture complex temporal dependencies while preserving semantic consistency. Extensive experiments on multiple public datasets show that FLS achieves consistent and competitive improvements over state-of-the-art TKGE methods, particularly in capturing complex temporal evolution.</p>