Effective time series prediction with hierarchical temporal dynamics is vital in fields like finance, energy, and traffic control. Recent work shows that large language models (LLMs), such as Time-LLM, can forecast by converting numerical data into text. However, current methods face key limitations: general-domain embeddings cause semantic mismatches, fixed prompts lack adaptability, and most critically, they fail to capture multi-scale temporal patterns. To address these issues, we propose a new framework based on multi-scale semantic encoding and adaptive reasoning. Our method uses a hierarchical reprogramming mechanism to decompose sequences into short-, medium-, and long-term components, better capturing multi-scale dynamics. We also introduce domain-adaptive semantic alignment to link numerical data with domain-specific text, improving coherence. Additionally, an adaptive fusion module and dynamic Chain-of-Thought (CoT) prompting enhance the use of pre-trained LLMs, increasing prediction flexibility and interpretability. Experiments show our framework outperforms state-of-the-art models in short- and long-term accuracy, cross-domain zero-shot performance, and computational efficiency.