Scholarlink

<p>Solar flares have a major influence on space weather and technological systems. Predicting flares remains a difficult task because of the non-stationary and imbalanced nature of flare time series. This study introduces an enhanced deep LSTM network that integrates Seasonal-Trend decomposition using Loess (STL) within an encoder–decoder architecture. The proposed approach reduces data complexity by separating trend, seasonal, and residual components, facilitating the model’s ability to capture long-term temporal dependencies in flare activity. Experiments on the GOES X-ray flare catalog (2003–2023). The model achieves TSS of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(0.924\pm 0.008\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0.924</mn> <mo>±</mo> <mn>0.008</mn> </mrow> </math></EquationSource> </InlineEquation> and an accuracy of 0.946&#xa0;±&#xa0;0.008, demonstrating its robustness in handling complex, non-stationary data. These results confirm the effectiveness of combining decomposition-based preprocessing with deep learning architectures for forecasting flares. The findings underscore the importance of incorporating trend, seasonal, and residual components, together with adaptive resampling, to enhance flare prediction, thereby highlighting their potential for deployment in large-scale, high-performance, real-time forecasting environments.</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

Non-stationary flare time series prediction using bidirectional long short-term memory based on STL decomposition

  • 期刊论文

  • 发表日期：

    2026年4月7日

查看全文
  • Zeinab Hassani,
  • Davud Mohammadpur,
  • Hossein Safari

摘要

Solar flares have a major influence on space weather and technological systems. Predicting flares remains a difficult task because of the non-stationary and imbalanced nature of flare time series. This study introduces an enhanced deep LSTM network that integrates Seasonal-Trend decomposition using Loess (STL) within an encoder–decoder architecture. The proposed approach reduces data complexity by separating trend, seasonal, and residual components, facilitating the model’s ability to capture long-term temporal dependencies in flare activity. Experiments on the GOES X-ray flare catalog (2003–2023). The model achieves TSS of \(0.924\pm 0.008\) 0.924 ± 0.008 and an accuracy of 0.946 ± 0.008, demonstrating its robustness in handling complex, non-stationary data. These results confirm the effectiveness of combining decomposition-based preprocessing with deep learning architectures for forecasting flares. The findings underscore the importance of incorporating trend, seasonal, and residual components, together with adaptive resampling, to enhance flare prediction, thereby highlighting their potential for deployment in large-scale, high-performance, real-time forecasting environments.