Multi-Scale Dual-Source Fusion Network for Long-Term Time Series Forecasting

This research addresses the limitations present in existing models for long-term time series forecasting, particularly in relation to the management of nonlinear and periodic data. Traditional methodologies, such as the Autoregressive Integrated Moving Average (ARIMA) and Exponential Smoothing, have been extensively utilized for time series predictions. However, these linear models frequently struggle to capture complex nonlinear patterns and may exhibit suboptimal performance when applied to datasets characterized by intricate seasonal variations. In contrast, Transformer-based models, including Autoformer and PatchTST, demonstrate proficiency in capturing temporal features and long-term dependencies. Nonetheless, they faced considerable challenges associated with noise, outliers, and computational complexity, which can diminish their effectiveness in practice. To address these challenges, we propose the Multi Scale Dual Source Fusion Network (MSDSFN), which is an optimized model that integrates frequency domain and time domain features. The model dynamically adjusts various features using gating mechanisms to fuse time domain and frequency domain features, enhancing robustness and prediction accuracy. In addition, it combines an efficient multi-scale attention (EMA) module that can proficiently capture short-term and long-term dependencies while maintaining channel dimensions, thus preserving basic feature details. Experimental results on multiple time series datasets demonstrate a significant improvement in performance, confirming the effectiveness and generalization ability of the model.