Scholarlink

<p>The recent progress in computing, networking, and sensor technology has dramatically expanded the vast amount of Time Series (TS) data available. TS forecasting is crucial for monitoring and controlling complex systems in the real world. Reducing and predicting errors is a critical challenge across various fields, creating a demand for innovative models. This paper presents the Multi-view Gradient Residual Reductive (MvGR2) model, implemented within a deep learning framework. MvGR2 employs a unique approach that prioritizes the reduction of forecasting errors through a novel multi-view perspective and a strategy to minimize gradient residuals. The hypothesis is that this integrated approach can more effectively capture complex temporal patterns and improve prediction accuracy. The methodology involves empirical evaluation across eight diverse datasets, encompassing different domains and characteristics. MvGR2 is compared to conventional deep learning models such as Convolutional Neural Network (CNN), Bidirectional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU), Recurrent Neural Network (RNN) as well as recent state-of-the-art architectures, including Informer, iTransformer, PatchTCN, TCN, and TimesNet. Results demonstrate that MvGR2 outperforms conventional models, achieving substantial average improvements. In terms of RMSE, improvements range from 1.19% to 38.04%, with notable gains observed for MvGR2-RNN (38.04%), MvGR2-BiLSTM (35.08%), and MvGR2-PatchTCN (33.26%). For MAPE, improvements range from 9.55% to 41.09%, with the highest gains achieved by MvGR2-iTransformer (41.09%), MvGR2-Informer (34.55%), and MvGR2-PatchTCN (18.61%). Non-parametric post-hoc statistical analysis confirms the significance of these improvements, with all relevant comparisons yielding p-values below 0.05. The findings highlight MvGR2’s robust capability to accurately capture complex Time Series patterns, providing detailed insights through predictive pattern analysis. The unique contributions of the MvGR2 model are emphasized, demonstrating its distinctiveness compared to other existing models. Additionally, the paper examines potential applications of MvGR2 across various domains and datasets, illustrating its versatility and efficiency in minimizing forecasting errors. The MvGR2 model presents a highly effective and adaptable approach for enhancing forecasting accuracy in diverse scenarios.</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

Multi-view gradient residual reductive (MvGR2) model for univariate time series predictions

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Vipin Kumar,
  • Ritika Singh,
  • Subham Kumar

摘要

The recent progress in computing, networking, and sensor technology has dramatically expanded the vast amount of Time Series (TS) data available. TS forecasting is crucial for monitoring and controlling complex systems in the real world. Reducing and predicting errors is a critical challenge across various fields, creating a demand for innovative models. This paper presents the Multi-view Gradient Residual Reductive (MvGR2) model, implemented within a deep learning framework. MvGR2 employs a unique approach that prioritizes the reduction of forecasting errors through a novel multi-view perspective and a strategy to minimize gradient residuals. The hypothesis is that this integrated approach can more effectively capture complex temporal patterns and improve prediction accuracy. The methodology involves empirical evaluation across eight diverse datasets, encompassing different domains and characteristics. MvGR2 is compared to conventional deep learning models such as Convolutional Neural Network (CNN), Bidirectional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU), Recurrent Neural Network (RNN) as well as recent state-of-the-art architectures, including Informer, iTransformer, PatchTCN, TCN, and TimesNet. Results demonstrate that MvGR2 outperforms conventional models, achieving substantial average improvements. In terms of RMSE, improvements range from 1.19% to 38.04%, with notable gains observed for MvGR2-RNN (38.04%), MvGR2-BiLSTM (35.08%), and MvGR2-PatchTCN (33.26%). For MAPE, improvements range from 9.55% to 41.09%, with the highest gains achieved by MvGR2-iTransformer (41.09%), MvGR2-Informer (34.55%), and MvGR2-PatchTCN (18.61%). Non-parametric post-hoc statistical analysis confirms the significance of these improvements, with all relevant comparisons yielding p-values below 0.05. The findings highlight MvGR2’s robust capability to accurately capture complex Time Series patterns, providing detailed insights through predictive pattern analysis. The unique contributions of the MvGR2 model are emphasized, demonstrating its distinctiveness compared to other existing models. Additionally, the paper examines potential applications of MvGR2 across various domains and datasets, illustrating its versatility and efficiency in minimizing forecasting errors. The MvGR2 model presents a highly effective and adaptable approach for enhancing forecasting accuracy in diverse scenarios.