Scholarlink

Recently, there has been a surge in linear-based solutions for the long-term time series forecasting (LTSF) task, driven by their effectiveness and efficiency. Building on these models, this paper introduces dynamic sparse training to explore the potential of creating lightweight models for LTSF in scenarios with extremely limited computational resources, without the need for cumbersome architectural design. To achieve this, we extensively explore the search space during sparse training to achieve a very high sparsity ratio for linear-based models. This approach results in models with fewer than 1k parameters and saves up to 33  \(\times \) training computational cost with only a slight performance loss. Experimental results on six real-life multivariate and univariate time series datasets demonstrate the effectiveness of our approach, achieving a better trade-off between computational efficiency and performance. We hope this finding opens up new research directions for the LTSF task, enabling the development of extremely lightweight models that can operate effectively in resource-constrained environments. Our code is publicly available at: https://github.com/QiaoXiao7282/LTSF-DST .

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

Achieving Long-Term Time Series Forecasting Models with Fewer Than 1k Parameters Through Dynamic Sparse Training

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Qiao Xiao,
  • Boqian Wu,
  • Mykola Pechenizkiy,
  • Decebal Constantin Mocanu

摘要

Recently, there has been a surge in linear-based solutions for the long-term time series forecasting (LTSF) task, driven by their effectiveness and efficiency. Building on these models, this paper introduces dynamic sparse training to explore the potential of creating lightweight models for LTSF in scenarios with extremely limited computational resources, without the need for cumbersome architectural design. To achieve this, we extensively explore the search space during sparse training to achieve a very high sparsity ratio for linear-based models. This approach results in models with fewer than 1k parameters and saves up to 33  \(\times \) training computational cost with only a slight performance loss. Experimental results on six real-life multivariate and univariate time series datasets demonstrate the effectiveness of our approach, achieving a better trade-off between computational efficiency and performance. We hope this finding opens up new research directions for the LTSF task, enabling the development of extremely lightweight models that can operate effectively in resource-constrained environments. Our code is publicly available at: https://github.com/QiaoXiao7282/LTSF-DST .