Temporal Knowledge Graph Completion is essential for many real-world applications, requiring effective modeling of temporal information. However, most existing methods rely on a single time scale, limiting their ability to capture dependencies over short and long time horizons. To address this, we propose MultiTEmb, a novel multi-scale temporal completion method. By decomposing temporal information into year, quarter, month, and day, we generate feature vectors at each scale and model dependencies using a Time-aware Efficient Self-Attention Mechanism (TE-SAM) with adaptive feature weighting. To further enhance feature fusion, we introduce InfoNCE-based contrastive learning to improve temporal representation discriminability and employ an Enhanced Gated Recurrent Unit (E-GRU) to sequentially integrate multi-scale embeddings. Extensive experiments on four benchmark datasets show that MultiTEmb significantly outperforms existing knowledge graph embedding and temporal knowledge graph completion models, demonstrating its effectiveness in temporal reasoning tasks.