Long-term time series forecasting (LTSF) is a crucial task with applications in weather, energy, finance, and other domains. However, real-world time series often exhibit complex temporal variations, complicating the extraction of latent patterns for accurate predictions. We observe that such series display distinct fine-scale fluctuations and coarse-scale trends across multiple sampling scales, each retaining inherent periodicity and trend components. To address this, we propose ST-Net, a novel framework that disentangles seasonal and trend representations at multiple scales. ST-Net employs a Dual-Path Feature Extractor (DPFE) to separately model scale-specific seasonal and trend components, followed by multi-scale aggregation to enhance temporal representations and uncover historical patterns. The framework integrates four core modules: a Down-sampling Block for generating multi-scale features via 1D convolutions; DPFE Blocks to decouple seasonal and trend components; a Multi-scale Mixing Block to fuse cross-scale temporal information; and a Past-Future Mapping (PFM) Block to project historical patterns into future sequences. Additionally, a Fourier-enhanced loss function is introduced to jointly optimize prediction accuracy in both time and frequency domains. Experiments on eight real-world LTSF benchmarks demonstrate that ST-Net outperforms state-of-the-art methods while maintaining a lightweight architecture. Our code is publicly available at https://github.com/LiChuan3/ST-Net .