Scholarlink

<p>Accurate and generalizable prediction of high-dimensional multivariate nonlinear time series data is vital to fields such as energy, meteorology, and social sentiment analysis, yet remains highly challenging. Although deep learning models can effectively capture complex temporal patterns, their prediction performance is highly sensitive to model architecture and hyperparameter choices, and there is still insufficient research on the optimization mechanism for such complex models. To address this challenge, a novel evolutionary hyperparameter optimization framework, named Chaotic Regulated Marine Predators Algorithm–TimesNet (CRMPA-TimesNet), is proposed for application to multivariate time series prediction tasks. Within the CRMPA-TimesNet framework, firstly, TimesNet is adopted as the backbone network to construct a modeling architecture for multivariate time series, and key hyperparameters that significantly affect its performance are identified. Secondly, an enhanced marine predators algorithm, referred to as CRMPA, is introduced to optimize these hyperparameters and the corresponding model structure configuration. Comparative experiments on three real-world datasets in different domains demonstrate that CRMPA-TimesNet outperforms six state-of-the-art baselines, achieving average reductions of 23.85%, 20.33%, and 23.18% in MAE; 11.96%, 34.61%, and 34.92% in MSE; and average improvements of 8.31%, 7.77%, and 9.14% in Directional Accuracy. These results not only prove the effectiveness of the CRMPA hyperparameter optimization method but also highlight the superiority of the proposed CRMPA-TimesNet optimization framework in multivariate time series forecasting and its potential for parallel applications.</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

Crmpa-timesnet: hyperparameter optimization with a modified metaheuristic for time series forecasting

  • 期刊论文

  • 发表日期：

    2026年5月27日

查看全文
  • Li Zhao,
  • Chundong Wang,
  • Hao Lin,
  • Rui Huang,
  • Hongjing Ma,
  • Chaoyang Guo

摘要

Accurate and generalizable prediction of high-dimensional multivariate nonlinear time series data is vital to fields such as energy, meteorology, and social sentiment analysis, yet remains highly challenging. Although deep learning models can effectively capture complex temporal patterns, their prediction performance is highly sensitive to model architecture and hyperparameter choices, and there is still insufficient research on the optimization mechanism for such complex models. To address this challenge, a novel evolutionary hyperparameter optimization framework, named Chaotic Regulated Marine Predators Algorithm–TimesNet (CRMPA-TimesNet), is proposed for application to multivariate time series prediction tasks. Within the CRMPA-TimesNet framework, firstly, TimesNet is adopted as the backbone network to construct a modeling architecture for multivariate time series, and key hyperparameters that significantly affect its performance are identified. Secondly, an enhanced marine predators algorithm, referred to as CRMPA, is introduced to optimize these hyperparameters and the corresponding model structure configuration. Comparative experiments on three real-world datasets in different domains demonstrate that CRMPA-TimesNet outperforms six state-of-the-art baselines, achieving average reductions of 23.85%, 20.33%, and 23.18% in MAE; 11.96%, 34.61%, and 34.92% in MSE; and average improvements of 8.31%, 7.77%, and 9.14% in Directional Accuracy. These results not only prove the effectiveness of the CRMPA hyperparameter optimization method but also highlight the superiority of the proposed CRMPA-TimesNet optimization framework in multivariate time series forecasting and its potential for parallel applications.