Scholarlink

To enhance the robustness and modeling capacity of time series forecasting in complex real-world scenarios, we propose a novel framework, FMVformer (Filtered Multi-View Transformer). The framework integrates two core components: (1) a Frequency Filter Block that operates as a preprocessing module to suppress noise and refine input signals through frequency-domain analysis; (2) a Multi-View Patch Transformer Block (MVPB) that extracts multi-scale temporal patterns and multi-view representations via hierarchical decomposition. MVPB comprises two sub-modules: Within-Patch Cyclic Attention (WPCA) and Across-Patch Residual Attention (APRA). The WPCA adopts a temporal cyclic panning strategy to model intra-patch dynamics from diverse perspectives, while APRA leverages residual connections between adjacent patches to mitigate the influence of abrupt values and enhance cross-patch consistency. Comprehensive evaluations on seven benchmark datasets demonstrate FMVformer’s consistent superiority over state-of-the-art baselines across multiple metrics, confirming its enhanced capabilities in noise suppression, structural pattern recognition, and prediction accuracy.

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

FMVformer: A Filtered Multi-view Transformer for Time Series Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hengshuai Fan,
  • Mingliang Zhang,
  • Zixin Hao,
  • Jingyi Sun,
  • Haibin Zhang,
  • Bin Li,
  • Hailong Meng

摘要

To enhance the robustness and modeling capacity of time series forecasting in complex real-world scenarios, we propose a novel framework, FMVformer (Filtered Multi-View Transformer). The framework integrates two core components: (1) a Frequency Filter Block that operates as a preprocessing module to suppress noise and refine input signals through frequency-domain analysis; (2) a Multi-View Patch Transformer Block (MVPB) that extracts multi-scale temporal patterns and multi-view representations via hierarchical decomposition. MVPB comprises two sub-modules: Within-Patch Cyclic Attention (WPCA) and Across-Patch Residual Attention (APRA). The WPCA adopts a temporal cyclic panning strategy to model intra-patch dynamics from diverse perspectives, while APRA leverages residual connections between adjacent patches to mitigate the influence of abrupt values and enhance cross-patch consistency. Comprehensive evaluations on seven benchmark datasets demonstrate FMVformer’s consistent superiority over state-of-the-art baselines across multiple metrics, confirming its enhanced capabilities in noise suppression, structural pattern recognition, and prediction accuracy.