Scholarlink

<p>Accurately capturing inter-series and intra-series variations is crucial for multivariate long-term time-series forecasting. Existing channel-independent and channel-mixing approaches struggle with complex inter-series relationships, while RNN-based models face challenges in capturing long-term intra-series dependencies. Additionally, current decomposition methods struggle with complex trends in the series, further hindering intra-series modeling. To address these, we propose GDRNet, which consists of four components: the channel grouping block (CGB), the channel group multi-mixer block (CGMB), the time-slice dilated residual GRU (SDRGRU), and the multi-trend decomposition block (MTDB). CGB groups channels with similar distributions for inter-series learning, while CGMB captures complex dependencies between series across various granularities and perspectives. SDRGRU expands the receptive field and incorporates residual learning to capture long-term intra-series&#xa0;dependencies, while MTDB enhances trend-seasonal decomposition, further facilitating precise intra-series modeling. GDRNet achieving 9.12% and 22.30% improvements in multivariate and univariate forecasting tasks, respectively, showcases its effectiveness in 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

GDRNet: a channel grouping based time-slice dilated residual network for long-term time-series forecasting

  • 期刊论文

  • 发表日期：

    2025年2月14日

查看全文
  • Qingda Bao,
  • Shengfa Miao,
  • Yulin Tian,
  • Xin Jin,
  • Puming Wang,
  • Qian Jiang,
  • Shaowen Yao,
  • Da Hu,
  • Ruoshu Wang

摘要

Accurately capturing inter-series and intra-series variations is crucial for multivariate long-term time-series forecasting. Existing channel-independent and channel-mixing approaches struggle with complex inter-series relationships, while RNN-based models face challenges in capturing long-term intra-series dependencies. Additionally, current decomposition methods struggle with complex trends in the series, further hindering intra-series modeling. To address these, we propose GDRNet, which consists of four components: the channel grouping block (CGB), the channel group multi-mixer block (CGMB), the time-slice dilated residual GRU (SDRGRU), and the multi-trend decomposition block (MTDB). CGB groups channels with similar distributions for inter-series learning, while CGMB captures complex dependencies between series across various granularities and perspectives. SDRGRU expands the receptive field and incorporates residual learning to capture long-term intra-series dependencies, while MTDB enhances trend-seasonal decomposition, further facilitating precise intra-series modeling. GDRNet achieving 9.12% and 22.30% improvements in multivariate and univariate forecasting tasks, respectively, showcases its effectiveness in time-series forecasting.