Scholarlink

<p>Long-term time series forecasting aims to predict extended future trends from historical data. Traditional linear models like ARIMA struggle to capture complex nonlinear patterns and intricate seasonal variations. While Transformer-based models proficiently capture long-term dependencies, they face considerable challenges with noise, outliers, and computational complexity. To address these limitations, we propose the Multi-Scale Dual-Source Fusion Network (MSDSFN), an optimized model integrating frequency and time domain features. The model dynamically aggregates these dual-source features using a Cross-Modal Evidential Fusion mechanism grounded in Dirichlet expectation and Dempster-Shafer (DS) theory. By explicitly quantifying epistemic uncertainty, this theoretically bounded approach strictly maximizes the Signal-to-Noise Ratio (SNR), significantly enhancing model robustness and prediction accuracy. Additionally, an efficient multi-scale attention (EMA) module captures both short- and long-term dependencies while maintaining channel dimensions to preserve essential feature details. Experimental results on multiple datasets demonstrate significant performance improvements, confirming the model’s effectiveness and generalization ability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-scale dual-source fusion network for long-term time series forecasting

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Chenyu Liao,
  • Chaoqun Hong,
  • Jialin Du,
  • Yuhang Huang

摘要

Long-term time series forecasting aims to predict extended future trends from historical data. Traditional linear models like ARIMA struggle to capture complex nonlinear patterns and intricate seasonal variations. While Transformer-based models proficiently capture long-term dependencies, they face considerable challenges with noise, outliers, and computational complexity. To address these limitations, we propose the Multi-Scale Dual-Source Fusion Network (MSDSFN), an optimized model integrating frequency and time domain features. The model dynamically aggregates these dual-source features using a Cross-Modal Evidential Fusion mechanism grounded in Dirichlet expectation and Dempster-Shafer (DS) theory. By explicitly quantifying epistemic uncertainty, this theoretically bounded approach strictly maximizes the Signal-to-Noise Ratio (SNR), significantly enhancing model robustness and prediction accuracy. Additionally, an efficient multi-scale attention (EMA) module captures both short- and long-term dependencies while maintaining channel dimensions to preserve essential feature details. Experimental results on multiple datasets demonstrate significant performance improvements, confirming the model’s effectiveness and generalization ability.