Scholarlink

<p>The task of multivariate time series (MTS) forecasting aims to utilize historical time series data to predict future trends and variations. MTS data possesses numerous latent features, with spatiotemporal correlations and periodicity being particularly significant. Most existing methods focused on either spatiotemporal correlations or periodicity, limiting their effectiveness. The paper proposes a novel STP-cube model by simultaneously capturing both, offering a more robust solution for MTS forecasting. MTS data is first transformed into a three-dimensional structure—the <i>cube</i>, through special stacking. Then graph convolution is utilized to capture spatiotemporal dependencies among sensors and Conv2Ds in another dimension to extract periodic features within sensors. By incorporating both features, the forecasting effectiveness is significantly enhanced. STP-cube model shows robust performance across six public MTS datasets, particularly excelling on the ETTh1 dataset where it reduces MSE by 0.177 and MAE by 0.084 comparing to the SOTA model.</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

STP-Cube—Multivariate Time Series Forecasting Based on Spatiotemporal Correlations and Periodicity

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月28日

查看全文
  • Yidong Zhang,
  • Jie Jing,
  • Luqi Liu,
  • Chengming Lan,
  • Peng Shi

摘要

The task of multivariate time series (MTS) forecasting aims to utilize historical time series data to predict future trends and variations. MTS data possesses numerous latent features, with spatiotemporal correlations and periodicity being particularly significant. Most existing methods focused on either spatiotemporal correlations or periodicity, limiting their effectiveness. The paper proposes a novel STP-cube model by simultaneously capturing both, offering a more robust solution for MTS forecasting. MTS data is first transformed into a three-dimensional structure—the cube, through special stacking. Then graph convolution is utilized to capture spatiotemporal dependencies among sensors and Conv2Ds in another dimension to extract periodic features within sensors. By incorporating both features, the forecasting effectiveness is significantly enhanced. STP-cube model shows robust performance across six public MTS datasets, particularly excelling on the ETTh1 dataset where it reduces MSE by 0.177 and MAE by 0.084 comparing to the SOTA model.