Scholarlink

Accurate forecasting of multivariate time series (MTS) is crucial across various domains. However, existing models often fail to effectively capture the complex periodic patterns across different variables, and struggle with utilizing periodic features in attention mechanism, which leads to uniformly distributed attention weights. To address these limitations, we introduce a novel Transformer model equipped with learnable period detection and periodic attention mechanisms. The combination of wavelet decomposition and convolution design can better extract the dominant period of each variable. The period-based Gaussian decay strategy can guide attention to relationships within the same position of each period. Experimental results show that our model significantly outperforms state-of-the-art attention-based models, particularly on datasets exhibiting notable periodic behavior.

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

PAformer: Transformer with Learnable Period Detection and Periodic Attention for Multivariate Time Series

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Meng Wan,
  • Huan Hao,
  • Yuxuan Bi,
  • Jue Wang,
  • Qi Su,
  • Fang Liu,
  • Peng Shi,
  • Xueyan Wei,
  • Yangang Wang,
  • Shulong Wang,
  • Helian Wu

摘要

Accurate forecasting of multivariate time series (MTS) is crucial across various domains. However, existing models often fail to effectively capture the complex periodic patterns across different variables, and struggle with utilizing periodic features in attention mechanism, which leads to uniformly distributed attention weights. To address these limitations, we introduce a novel Transformer model equipped with learnable period detection and periodic attention mechanisms. The combination of wavelet decomposition and convolution design can better extract the dominant period of each variable. The period-based Gaussian decay strategy can guide attention to relationships within the same position of each period. Experimental results show that our model significantly outperforms state-of-the-art attention-based models, particularly on datasets exhibiting notable periodic behavior.