Long-term time series forecasting remains challenging due to complex temporal dependencies, diverse data distributions, and computational inefficiencies with extended sequences. We propose CMTFormer, a novel architecture that addresses these limitations through multi-scale temporal modeling and contrastive learning. Our approach combines adaptive trend decomposition across multiple timescales with a representation learning framework that leverages self-attention mechanisms and dilated convolutions. The proposed multi-scale trend decomposition disentangles time series into interpretable components at varying resolutions, while the contrastive learning strategy enhances feature discrimination by differentiating between semantically related and unrelated temporal patterns. Extensive experiments on six real-world benchmarks spanning energy, transportation, weather, finance, and public health domains demonstrate that CMTFormer consistently outperforms state-of-the-art forecasting models.