Time series forecasting remains challenging due to the inherent non-stationarity of real-world data, characterized by complex seasonality and evolving trends. However, current methods fail to effectively model complex periodic patterns and lack explicit modeling of trend dynamics. In this paper, we propose a novel framework named D-GFAN, based on Transformer with decomposition and Fourier Analysis Network(FAN) for time series forecasting. Our model first decomposes normalized input sequences into seasonal and trend components via moving averages. The seasonal component is processed by a Gated FAN-enhanced Transformer, while a lightweight Multilayer Perceptron handles trend extrapolation. Experiments on five benchmarks demonstrate State-of-the-Art performance, with D-GFAN achieving 12% lower error reduction than PatchTST on periodic data and 18%–22% versus FEDformer/Autoformer despite using 50% fewer parameters.