Scholarlink

<p>Time series forecasting is a critical task in various domains, such as finance, energy, transportation, and meteorology, which involves predicting future trends based on historical data. However, existing time series forecasting models often struggle to simultaneously capture short-term volatility and long-term dependencies in complex, nonlinear data. These limitations are directly addressed in this paper through the integration of dynamic regularization and attention mechanisms. We propose a novel hybrid model that integrates dynamic regularization (DR) and attention mechanisms into the random vector functional link long short-term memory (RVFL-LSTM) framework, named dynamic attention-enhanced RVFL-LSTM (DARVFL-LSTM), to effectively handle nonlinear and non-stationary time series data. DR dynamically adjusts the regularization strength based on the data’s volatility, effectively balancing the learning of short-term and long-term patterns. The attention mechanism assigns dynamic weights to sliding window observations, emphasizing key data points to enhance short-term feature extraction. Comprehensive experiments on eight public time series datasets demonstrate that DARVFL-LSTM significantly outperforms baseline models, including RVFL-LSTM, long short-term memory (LSTM), and online sequential extreme learning machine (OS-ELM), in terms of accuracy and adaptability. Moreover, ablation and efficiency analyses confirm the complementary roles of DR and attention mechanisms. The results establish DARVFL-LSTM as an efficient and precise solution for complex time series forecasting tasks, particularly in application domains such as energy load prediction, traffic forecasting, weather analysis, and financial trend modeling.</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

DARVFL-LSTM: a time series prediction model integrating dynamic regularization and attention mechanism

  • 期刊论文

  • 发表日期：

    2025年5月28日

查看全文
  • Gaofeng Liu

摘要

Time series forecasting is a critical task in various domains, such as finance, energy, transportation, and meteorology, which involves predicting future trends based on historical data. However, existing time series forecasting models often struggle to simultaneously capture short-term volatility and long-term dependencies in complex, nonlinear data. These limitations are directly addressed in this paper through the integration of dynamic regularization and attention mechanisms. We propose a novel hybrid model that integrates dynamic regularization (DR) and attention mechanisms into the random vector functional link long short-term memory (RVFL-LSTM) framework, named dynamic attention-enhanced RVFL-LSTM (DARVFL-LSTM), to effectively handle nonlinear and non-stationary time series data. DR dynamically adjusts the regularization strength based on the data’s volatility, effectively balancing the learning of short-term and long-term patterns. The attention mechanism assigns dynamic weights to sliding window observations, emphasizing key data points to enhance short-term feature extraction. Comprehensive experiments on eight public time series datasets demonstrate that DARVFL-LSTM significantly outperforms baseline models, including RVFL-LSTM, long short-term memory (LSTM), and online sequential extreme learning machine (OS-ELM), in terms of accuracy and adaptability. Moreover, ablation and efficiency analyses confirm the complementary roles of DR and attention mechanisms. The results establish DARVFL-LSTM as an efficient and precise solution for complex time series forecasting tasks, particularly in application domains such as energy load prediction, traffic forecasting, weather analysis, and financial trend modeling.