Scholarlink

<p>Long Short-Term Memory (LSTM) is widely used in time-series analysis and prediction. Combining LSTM with mathematical transformer models such as Fourier and Wavelet or with statistical methods like Lomb-Scargle, Harmonic Regression, and Seasonal-Trend decomposition can effectively capture time-series behavior and enhance the efficiency and accuracy of prediction models. However, this approach heavily relies on the characteristics of the data. Depending on whether the time series data is periodic, chaotic, or quasi-periodic, the selection of appropriate tools varies—an issue often overlooked in current research that combines LSTM with mathematical transformers or statistical methods. In this study, we aim to demonstrate that using nonlinear dynamic systems techniques, which can capture the characteristics of time series data, leads to the selection of the best transformer. Moreover, this approach explains why LSTM and its combination models may not always perform well. This insight is vital for designing new models based on the dynamic behavior of data. Our experimental findings on two real-world datasets indicate that a combination of LSTM with transformers or decomposition tools does not perform well on highly sensitive and chaotic time series. In other words, we present a model that explains the inefficiencies of LSTM-based models in certain 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

Investigating LSTM-based time series prediction using dynamic systems measures

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月4日

查看全文
  • Amin Mahmoudi

摘要

Long Short-Term Memory (LSTM) is widely used in time-series analysis and prediction. Combining LSTM with mathematical transformer models such as Fourier and Wavelet or with statistical methods like Lomb-Scargle, Harmonic Regression, and Seasonal-Trend decomposition can effectively capture time-series behavior and enhance the efficiency and accuracy of prediction models. However, this approach heavily relies on the characteristics of the data. Depending on whether the time series data is periodic, chaotic, or quasi-periodic, the selection of appropriate tools varies—an issue often overlooked in current research that combines LSTM with mathematical transformers or statistical methods. In this study, we aim to demonstrate that using nonlinear dynamic systems techniques, which can capture the characteristics of time series data, leads to the selection of the best transformer. Moreover, this approach explains why LSTM and its combination models may not always perform well. This insight is vital for designing new models based on the dynamic behavior of data. Our experimental findings on two real-world datasets indicate that a combination of LSTM with transformers or decomposition tools does not perform well on highly sensitive and chaotic time series. In other words, we present a model that explains the inefficiencies of LSTM-based models in certain scenarios.