SFMixer: Local Periodic Features in Frequency Domain and Global Features in Time Domain Modeling for Long-Term Time Series Forecasting
Peng Wu, Jian–wei Liu, Jiayi Han · 2024
Considering the sophisticated interplay of features within time-series data, the fusion of local and global features is essential for effective long-term forecasting. While CNNs are adept at capturing local characteristics, they often fall short in encapsulating global features within time series when using simplistic model architectures or those with reduced complexity. Additionally, existing research on modeling dependencies between time-series data using frequency domain information, especially local periodic features, is insufficient. To address these issues, we propose the SFMixer model, which utilizes an improved CNN structure and MLP network to model and integrate both local periodic variations in the frequency domain, including intra-local and inter-local periodic variations, and global features in the time domain. The SFMixer model outperforms previous methods on eight benchmark datasets. In both multivariate and univariate forecasting experiments, SFMixer achieves significant reductions in MSE and MAE. Furthermore, SFMixer demonstrates higher capabilities in capturing temporal information and training efficiency. We will make our code and model publicly available.