<p>Frequency-domain methods for time series forecasting have garnered significant attention due to their exceptional ability to capture periodic patterns. However, prevailing approaches predominantly rely on global Fourier transforms to extract frequency features, overlooking the localized time–frequency dynamics inherent in non-stationary signals. This leads to the issue of <b>spectral confusion</b>, a phenomenon in which distinct time-domain signals yield highly similar spectral representations. This ambiguity in the frequency domain hampers both the predictive accuracy and generalization capability of forecasting models. To address this challenge, we propose F2Fnet, a forecasting architecture based on dual Fourier modeling. F2Fnet integrates a short-time Fourier transform module (STFTM) with a global periodicity extraction module (GPEM) to effectively capture localized time–frequency dynamics and global periodic trends, respectively. The STFTM employs dynamic time window analysis to achieve precise extraction of localized frequency features through fine-grained time–frequency decomposition, significantly enhancing the representation of non-stationary dynamics. Conversely, the GPEM leverages global Fourier transforms to isolate high-energy frequency components, bolstering the modeling of long-term periodic patterns. Through their synergistic interaction, these modules effectively mitigate spectral confusion, substantially improving the representation of temporal structures. Extensive experiments on multiple real-world datasets demonstrate that F2Fnet achieves forecasting accuracy on par with or surpassing state-of-the-art baselines, while exhibiting remarkable computational efficiency by substantially reducing both computational cost and memory usage. This indicates that, beyond mitigating spectral confusion, the lightweight design of F2Fnet ensures high practicality in large-scale and resource-constrained forecasting scenarios, showcasing superior performance and efficiency.</p>