Scholarlink

<p>Earth system infrastructures that utilize satellite-based informatics such as Global Positioning System (GPS) communications are regularly impacted by Ionospheric Total Electron Content (TEC) gradients. Unfortunately, modeling these TEC gradients under physical laws to guide mitigation strategies is challenging due to their dynamic and sudden occurrence likelihoods. For context, a multi-layer perceptron machine learning (ML) model could predict hourly TEC values to about 80% accuracy of a GPS receiver’s estimates under solar wind and flux constraints. However, it is unclear whether these models could capture and preserve the TEC gradient irregularities within the predicted signals due to their output resolutions. To bridge this gap, we introduce, for the first time, a 30-second-resolution TEC prediction model based on the Light Gradient Boosting Machine algorithm. This cadence enables the derivation of Rate of TEC changes (ROT), and its index, ROTI, as diagnostic irregularity indicators to assess model robustness. The current implementation employs GPS observations across solar cycle 24, over a station located at 5.49 °S, 47.49 °W. A multi-metric evaluation framework, including dynamic time warping, is used to assess the model while SHAP (SHapley Additive exPlanations) enhances the interpretability of feature attributions. For the 30s predictions, a test accuracy of 91% (MAE = 4.38 TECU) is observed over the high solar activity period (2015), demonstrating the model’s applicability for accurate high-rosolution TEC estimations. We benchmark TEC predictions against the International Reference Ionosphere (IRI-2020) and showed that the hourly variant of our model outperforms IRI by about 35% in accuracy, 57% drop in absolute errors and 54% gain in prediction skill. More importantly, we captured significant irregularity (ROTI) dynamics and morphologies with the 30s TEC model under three geomagnetic storm intensity levels in contrast to an attention-based Long Short-Term Memory model subjected to the same sets of experiment. This study demonstrates the possibility of achieving scalable TEC irregularities detections with a single TEC-based model without explicitly training redundant models for individual transients.</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

t-STEP: An interpretable model for total electron content predictions and irregularities estimations

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Stephen Tete,
  • Carl Shneider,
  • Maxime Cordy,
  • Claudio Cesaroni,
  • Andreas Hein,
  • Vasily Petrov

摘要

Earth system infrastructures that utilize satellite-based informatics such as Global Positioning System (GPS) communications are regularly impacted by Ionospheric Total Electron Content (TEC) gradients. Unfortunately, modeling these TEC gradients under physical laws to guide mitigation strategies is challenging due to their dynamic and sudden occurrence likelihoods. For context, a multi-layer perceptron machine learning (ML) model could predict hourly TEC values to about 80% accuracy of a GPS receiver’s estimates under solar wind and flux constraints. However, it is unclear whether these models could capture and preserve the TEC gradient irregularities within the predicted signals due to their output resolutions. To bridge this gap, we introduce, for the first time, a 30-second-resolution TEC prediction model based on the Light Gradient Boosting Machine algorithm. This cadence enables the derivation of Rate of TEC changes (ROT), and its index, ROTI, as diagnostic irregularity indicators to assess model robustness. The current implementation employs GPS observations across solar cycle 24, over a station located at 5.49 °S, 47.49 °W. A multi-metric evaluation framework, including dynamic time warping, is used to assess the model while SHAP (SHapley Additive exPlanations) enhances the interpretability of feature attributions. For the 30s predictions, a test accuracy of 91% (MAE = 4.38 TECU) is observed over the high solar activity period (2015), demonstrating the model’s applicability for accurate high-rosolution TEC estimations. We benchmark TEC predictions against the International Reference Ionosphere (IRI-2020) and showed that the hourly variant of our model outperforms IRI by about 35% in accuracy, 57% drop in absolute errors and 54% gain in prediction skill. More importantly, we captured significant irregularity (ROTI) dynamics and morphologies with the 30s TEC model under three geomagnetic storm intensity levels in contrast to an attention-based Long Short-Term Memory model subjected to the same sets of experiment. This study demonstrates the possibility of achieving scalable TEC irregularities detections with a single TEC-based model without explicitly training redundant models for individual transients.