DLSFNet: A Modular Architecture for Long-Term Time Series Forecasting via Structural Decoupling and Multi-scale Feature Fusion

Time series forecasting plays a vital role in various domains such as financial risk management, power system scheduling, and weather prediction. However, due to the non-stationarity of real-world data and the presence of multi-scale periodic patterns, forecasting tasks still face significant challenges. To address the intertwined issues of trend variation, local fluctuations, and global periodicity, this paper proposes a modular framework for long-term time series forecasting, named Decoupling and Lightweight Scale-Fusion Network (DLSFNet). DLSFNet adopts a structural decoupling strategy that decomposes time series data into multiple components, including trends, local fluctuations, and global periodic patterns, to achieve hierarchical and specialized feature extraction. Specifically, the Trend Normalization Processing (TNP) module integrates normalization techniques with trend-specific feature extraction to enhance the representation of long-term trends. The Local Importance Feature Extractor (LIFE) module focuses on short-term disturbances and local fluctuations, capturing fine-grained and salient temporal features. The Global Periodic Extractor (GPE) module integrates time-domain and frequency-domain features to extract stable periodic patterns across the entire sequence. The outputs of these modules are subsequently integrated using a Lightweight Fusion Network (LFN), which facilitates effective feature interaction and joint representation learning. Extensive experiments conducted on multiple public datasets demonstrate that DLSFNet consistently outperforms state-of-the-art baselines in terms of forecasting accuracy, robustness, and generalization. These results confirm the effectiveness and superiority of DLSFNet in capturing and expressing complex temporal dependencies.