Scholarlink

<p>Multivariate Long Series Time-series Forecasting (MLSTF) plays an important role in enterprise production and operation. Previous MLSTF algorithms extract sequence features by adding correlation computation methods in time and variables, but still have the following problems: 1) Lack of multivariate correlation modeling in complex time patterns. 2) Lack of effective multi-scale feature alignment fusion of the data. In order to solve the above problems, this manuscript proposes the MSTCA network, which is a multi-horizon spatio-temporal correlation aggregation method. First, a Multi Scale Grid (MSG) batch sliding feature extraction strategy is proposed to converge multivariate information in different temporal modes by different size windows. Meanwhile, a gridding method is used to separate the features of different variables, which avoids the problem of losing the original information caused by unified multivariate aggregation. Second, in order to effectively extract the fragmented features of MSG, the channel Multiscale Graph Convolutional Network (MGCN) and Multiscale Hierarchical Attention (MHA) structure are designed to perform affinity discovery in multiscale spatial and temporal dimensions, respectively. Finally, in order to effectively fuse the multilevel features of MGCN and MHA, a feature-aligned Time Series Bidirectional Feature Pyramid Networks (TS-BiFPN) approach is proposed for robust characterization of multiscale advanced information patterns. Compared with the model based on graph convolution with self-attention structure, the accuracy of our proposed MSTCA model is significantly improved in long series time series prediction. Extensive experiments were conducted across multiple datasets under different configurations. The results demonstrate that the MSTCA model outperforms baseline methods, achieving the highest average ranking scores of 1.61 (MSE) and 1.82 (MAE) for MSE and MAE metrics, respectively.</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

MSTCA: Multi-horizon Spatiotemporal Correlation Aggregation for Multivariate Long-term Time Series Forecasting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月21日

查看全文
  • Dunhuang Shi,
  • Tao Zhang,
  • Yuntao Duan,
  • Lei Sun

摘要

Multivariate Long Series Time-series Forecasting (MLSTF) plays an important role in enterprise production and operation. Previous MLSTF algorithms extract sequence features by adding correlation computation methods in time and variables, but still have the following problems: 1) Lack of multivariate correlation modeling in complex time patterns. 2) Lack of effective multi-scale feature alignment fusion of the data. In order to solve the above problems, this manuscript proposes the MSTCA network, which is a multi-horizon spatio-temporal correlation aggregation method. First, a Multi Scale Grid (MSG) batch sliding feature extraction strategy is proposed to converge multivariate information in different temporal modes by different size windows. Meanwhile, a gridding method is used to separate the features of different variables, which avoids the problem of losing the original information caused by unified multivariate aggregation. Second, in order to effectively extract the fragmented features of MSG, the channel Multiscale Graph Convolutional Network (MGCN) and Multiscale Hierarchical Attention (MHA) structure are designed to perform affinity discovery in multiscale spatial and temporal dimensions, respectively. Finally, in order to effectively fuse the multilevel features of MGCN and MHA, a feature-aligned Time Series Bidirectional Feature Pyramid Networks (TS-BiFPN) approach is proposed for robust characterization of multiscale advanced information patterns. Compared with the model based on graph convolution with self-attention structure, the accuracy of our proposed MSTCA model is significantly improved in long series time series prediction. Extensive experiments were conducted across multiple datasets under different configurations. The results demonstrate that the MSTCA model outperforms baseline methods, achieving the highest average ranking scores of 1.61 (MSE) and 1.82 (MAE) for MSE and MAE metrics, respectively.