Long-term time series forecasting plays a crucial role across various domains such as energy management, traffic prediction, and healthcare. Accurately modeling both long-term trends and short-term seasonal fluctuations remains a key challenge. Existing methods typically treat time series as a whole without explicitly separating trend and seasonal components, leading to feature interference and reduced predictive performance. In response to these limitations, we propose DSTNet, a novel dual-branch deep learning architecture designed to disentangle and model trend and seasonal signals separately. Specifically, DSTNet utilizes a TrendNet branch to capture global low-frequency patterns and a SeasonalNet branch that incorporates a Global Pattern Extractor for multi-scale periodic features and a Local Dynamic Encoder for fine-grained local fluctuations. These branches are integrated through a lightweight fusion mechanism to generate accurate predictions. Extensive experiments conducted on nine diverse real-world datasets demonstrate that DTSNet consistently surpasses state-of-the-art methods in both predictive accuracy and robustness. Further analysis validates that explicitly separating trend and seasonal com-ponents is essential for improving forecasting performance.