Long-term time series forecasting (LTSF) is essential for applications requiring accurate future insights in domains such as energy management, traffic planning, and financial prediction. While Transformer architectures have recently achieved remarkable performance on LTSF tasks, they often come with high model complexity, large parameter counts, and costly inference latency, which limits their practicality in real-time or resource-constrained environments. To overcome these limitations, we propose DPC-Net, a lightweight and efficient dual-branch framework that explicitly decouples input time series into trend and seasonal branches. The trend branch captures global low-frequency patterns through simple yet effective channel-wise linear projections, enhancing interpretability and computational efficiency. The seasonal branch incorporates a novel Temporal Multi-Scale Block (TMB), which employs multi-branch convolutional filters with different receptive fields to model fine-grained seasonal structures across diverse frequencies. Furthermore, a GateCorrector adaptively fuses and refines the outputs of both branches, effectively reducing residual prediction errors. Extensive experiments on six benchmarks demonstrate that DPC-Net consistently outperforms state-of-the-art models in both accuracy and efficiency.