Scholarlink

<p>Multivariate time series forecasting is critical in domains such as energy management, traffic prediction, and weather monitoring. Existing Transformer-based approaches face a trade-off between efficiently modeling inter-variable dependencies and adaptively capturing diverse temporal patterns. To address this, we propose a Dynamic Ring-Star Transformer (DRSFormer), a structure-driven Transformer that represents each variable as a node in a virtual topology and combines sparse local routing with multi-center global fusion. The architecture jointly models inter-variable dependencies by combining sparse local interactions with global semantic fusion, while capturing diverse temporal dynamics through adaptive multi-scale temporal modeling. Extensive experiments are conducted on multiple public benchmarks, including ETT, Weather, ECL, Traffic, Exchange, and PEMS. The results demonstrate that DRSFormer achieves robust and competitive performance. On PEMS04, it attains a mean squared error (MSE) of 0.089, representing a 19.8% relative improvement over iTransformer and highlighting its effectiveness in modeling complex temporal and structural dependencies.</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

DRSFormer: A transformer with ring-star topology for multivariate time series forecasting

  • 期刊论文

  • 发表日期：

    2026年5月13日

查看全文
  • Xing Wu,
  • Ying Zhou,
  • Xinyi Duan,
  • Quan Qian,
  • Jun Song,
  • Jianbiao Dai

摘要

Multivariate time series forecasting is critical in domains such as energy management, traffic prediction, and weather monitoring. Existing Transformer-based approaches face a trade-off between efficiently modeling inter-variable dependencies and adaptively capturing diverse temporal patterns. To address this, we propose a Dynamic Ring-Star Transformer (DRSFormer), a structure-driven Transformer that represents each variable as a node in a virtual topology and combines sparse local routing with multi-center global fusion. The architecture jointly models inter-variable dependencies by combining sparse local interactions with global semantic fusion, while capturing diverse temporal dynamics through adaptive multi-scale temporal modeling. Extensive experiments are conducted on multiple public benchmarks, including ETT, Weather, ECL, Traffic, Exchange, and PEMS. The results demonstrate that DRSFormer achieves robust and competitive performance. On PEMS04, it attains a mean squared error (MSE) of 0.089, representing a 19.8% relative improvement over iTransformer and highlighting its effectiveness in modeling complex temporal and structural dependencies.