<p>Time series forecasting is fundamental to both scientific research and industrial applications, yet accurately modeling long-range dependencies and disentangling multi-scale temporal structures remains a persistent challenge. Although Transformer-based architectures exhibit powerful sequence representation capabilities, they often lack explicit mechanisms to isolate and integrate heterogeneous temporal components, which can limit their robustness under complex, non-stationary conditions. To address these limitations, this study proposes a novel framework, the Dynamic Fusion Gated Patch Time Series Transformer, referred to as DFG-PatchTST. The framework integrates a hybrid time–frequency decomposition backbone to isolate trend, seasonal, and noise components; a patch-based Transformer encoder to capture localized and cross-variable interactions; and a dynamic fusion gate with residual refinement for adaptive aggregation. Comprehensive evaluations across nine benchmark datasets demonstrate that DFG-PatchTST achieves highly competitive or superior performance compared to existing baselines. While its advantages in capturing spatio-temporal dependencies are most pronounced over extended forecasting horizons, the improvement margins are more tempered on highly volatile, minute-level data. Despite these variations across temporal granularities, rigorous evaluations across five independent trials confirm that the framework maintains exceptional training stability with an MSE standard deviation below 0.015.</p>