Scholarlink

<p>Nonlinear time series data, characterized by complex patterns and dynamic correlations, are increasingly important in fields like banking, meteorology, and engineering. However, their complexity and uncertainty pose significant challenges to traditional prediction methods. To tackle these issues, a new approach, referred to as the rolling variational multi-attention bidirectional encoder-decoder framework (R-VABED), has been introduced. Rolling Variational Mode Decomposition (R-VMD) integrates a rolling sliding window strategy with Variational Mode Decomposition (VMD), enabling continuous updates to the observation window and effective capture of transient features in time series data. By effectively mitigating data noise and minimizing information leakage, R-VMD enhances data integrity and boosts the model’s prediction accuracy in real-world applications. To rigorously evaluate the performance of the proposed model, a series of experiments was executed utilizing authentic and unaltered datasets. The empirical results substantiate that R-VABED exhibits a substantial enhancement over Crossformer. Specifically, it achieves a ~ 50% reduction in mean absolute error for the water quality dataset, a ~ 90% decline in mean squared error for the air quality dataset, and a 50% decrease in MSE for the electricity load dataset. These findings demonstrate that R-VABED achieves competitive accuracy and stability in handling complex time series forecasting assignments. Our implementation code is publicly available via the link: <a href="https://github.com/TimeSeriesLab/R-VABED.git">https://github.com/TimeSeriesLab/R-VABED.git</a>.</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

R-VABED: a nonlinear time series prediction model based on rolling decomposition and bidirectional LSTM

  • 期刊论文

  • 发表日期：

    2025年10月18日

查看全文
  • Hongxing Peng,
  • Shuxia Jiang,
  • Jianji Ren,
  • Haiqing Liu,
  • Yongliang Yuan,
  • Yun Xin,
  • Yanan Li,
  • Yunfeng Chen

摘要

Nonlinear time series data, characterized by complex patterns and dynamic correlations, are increasingly important in fields like banking, meteorology, and engineering. However, their complexity and uncertainty pose significant challenges to traditional prediction methods. To tackle these issues, a new approach, referred to as the rolling variational multi-attention bidirectional encoder-decoder framework (R-VABED), has been introduced. Rolling Variational Mode Decomposition (R-VMD) integrates a rolling sliding window strategy with Variational Mode Decomposition (VMD), enabling continuous updates to the observation window and effective capture of transient features in time series data. By effectively mitigating data noise and minimizing information leakage, R-VMD enhances data integrity and boosts the model’s prediction accuracy in real-world applications. To rigorously evaluate the performance of the proposed model, a series of experiments was executed utilizing authentic and unaltered datasets. The empirical results substantiate that R-VABED exhibits a substantial enhancement over Crossformer. Specifically, it achieves a ~ 50% reduction in mean absolute error for the water quality dataset, a ~ 90% decline in mean squared error for the air quality dataset, and a 50% decrease in MSE for the electricity load dataset. These findings demonstrate that R-VABED achieves competitive accuracy and stability in handling complex time series forecasting assignments. Our implementation code is publicly available via the link: https://github.com/TimeSeriesLab/R-VABED.git.