<p>Long-term time series forecasting (LTSF) has important applications in scenarios such as transportation scheduling, energy management, and financial modeling. However, existing methods still face many challenges in modeling long-term dependencies and capturing trends and seasonal structure. In this paper, we propose a structured decoupled forecasting model, DTSformer, to model trends and cyclical variations in time series more efficiently. The model adopts a multi-scale moving average decomposition strategy to divide the original series into two types of components, trend and seasonality, and designs two modeling structures separately: the HACM (Hybrid Attention-Convolution Module) for modeling cyclical patterns, which takes into account the short-term dynamics and the long-period dependence by fusing the local convolution and the global attention; and the FuseMixer module for trending modeling, which combines channel aggregation and local convolution mechanisms to capture the smooth evolutionary structure in trending components. Experiments demonstrate that DTSformer achieves significant prediction accuracy improvement over existing methods on several public datasets, and especially shows stronger stability and generalization ability in long prediction tasks.</p>