Scholarlink

<p>Time series forecasting plays a crucial role in numerous applications. Views at different scale may offer complementary information and contain temporal coherence of a time series, but are largely neglected by existing studies. In this paper, we aim to take advantage of the complementarity and consistency among multiple scales of a time series to enhance accuracy of forecasting models. To accomplish this, we propose a Multi-Scale forecasting Framework (MSF) that comprises two key components: the Temporal Hierarchical Information Fusion (THIF) module and the Reconciliation Network (ReconNet). The THIF module facilitates the exchange of complementary information among scales, enabling a more comprehensive representation of each scale. The ReconNet, in a top-down manner, dynamically revises predictions to ensure coherence and enhance accuracy. The framework is designed to be flexible, such that different forecasting models can be plugged in, and is capable of joint training of the components. Extensive experiments on real-world datasets indicate the effectiveness of our approach.</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

A multi-scale time series forecasting framework with temporal hierarchical information fusion and reconciliation

  • 期刊论文

  • 发表日期：

    2025年5月17日

查看全文
  • Keqin Shi,
  • Zhihua Ding,
  • Zhen Chen,
  • Hao He,
  • Weiqiang Sun,
  • Weisheng Hu

摘要

Time series forecasting plays a crucial role in numerous applications. Views at different scale may offer complementary information and contain temporal coherence of a time series, but are largely neglected by existing studies. In this paper, we aim to take advantage of the complementarity and consistency among multiple scales of a time series to enhance accuracy of forecasting models. To accomplish this, we propose a Multi-Scale forecasting Framework (MSF) that comprises two key components: the Temporal Hierarchical Information Fusion (THIF) module and the Reconciliation Network (ReconNet). The THIF module facilitates the exchange of complementary information among scales, enabling a more comprehensive representation of each scale. The ReconNet, in a top-down manner, dynamically revises predictions to ensure coherence and enhance accuracy. The framework is designed to be flexible, such that different forecasting models can be plugged in, and is capable of joint training of the components. Extensive experiments on real-world datasets indicate the effectiveness of our approach.