Scholarlink

<p>Multivariate long-sequence time series forecasting (MLTSF) poses significant challenges. Existing Transformer-based methods mostly focus on the time domain, neglecting the global attributes and intricate frequency patterns in the data. To address this limitation, we propose a novel Patchwise SpectroTemporal Analytic Network (PSTANet) for MLTSF. Our model synergizes frequency domain attention with time domain convolution. Specifically, we introduce the Construction of Multi-Scale Patches (CMSP) module, which provides detailed contextual information for subsequent spectro-temporal analysis. We further design two parallel modules: Frequency-Based Cross-Patch Attention (FCPA) and Intra-Patch Convolution Operation (IPCO). The FCPA module leverages Fourier transformation to highlight key frequency components, capturing essential frequency domain features and mitigating noise impact. Meanwhile, the IPCO module extracts temporal details within patches, complementing the frequency domain analysis. By combining these strategies, PSTANet provides a robust framework for comprehensive spectro-temporal analysis. Extensive experiments on eight real datasets demonstrate that our model achieves improvements of 4.50% in MSE and 3.81% in MAE over benchmark models, underscoring its innovative contribution to time series forecasting.</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

Pstanet: patchwise spectro-temporal analytic network for multivariate long-sequence time series forecasting

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Chunru Dong,
  • Wei Luo,
  • Qiang Hua,
  • Yong Zhang,
  • Feng Zhang

摘要

Multivariate long-sequence time series forecasting (MLTSF) poses significant challenges. Existing Transformer-based methods mostly focus on the time domain, neglecting the global attributes and intricate frequency patterns in the data. To address this limitation, we propose a novel Patchwise SpectroTemporal Analytic Network (PSTANet) for MLTSF. Our model synergizes frequency domain attention with time domain convolution. Specifically, we introduce the Construction of Multi-Scale Patches (CMSP) module, which provides detailed contextual information for subsequent spectro-temporal analysis. We further design two parallel modules: Frequency-Based Cross-Patch Attention (FCPA) and Intra-Patch Convolution Operation (IPCO). The FCPA module leverages Fourier transformation to highlight key frequency components, capturing essential frequency domain features and mitigating noise impact. Meanwhile, the IPCO module extracts temporal details within patches, complementing the frequency domain analysis. By combining these strategies, PSTANet provides a robust framework for comprehensive spectro-temporal analysis. Extensive experiments on eight real datasets demonstrate that our model achieves improvements of 4.50% in MSE and 3.81% in MAE over benchmark models, underscoring its innovative contribution to time series forecasting.