Scholarlink

<p>Forecasting ionospheric Total Electron Content (TEC) remains a challenging task for space weather applications due to strong variability across latitudes and solar cycle phases. This study presents a comparative analysis of neural network models, namely Long Short-Term Memory (LSTM) and Backpropagation (BP), and physics-based models, including the Thermosphere-Ionosphere-Electrodynamics General Circulation Model (TIE-GCM) and NeQuick. The evaluation is performed at three GPS stations: Novosibirsk (mid-latitude), Varanasi (low-latitude), and Port Blair (equatorial latitude), during the solar maximum year of 2014 and the solar minimum year of 2018 during Solar Cycle 24. The results reveal station- and solar cycle phase-dependent variations in model performance. In Novosibirsk, LSTM consistently achieved the highest correlation (R ≈ 0.85) and the lowest RMSE, particularly during the solar maximum phase, successfully capturing the variability of storm-time TEC. Varanasi exhibited strong seasonal variability, with neural models maintaining stable performance across both solar phases, whereas physics-based models struggled with equatorial anomaly effects, particularly during the solar minimum. Port Blair showed the largest model differences: LSTM maintained stable performance with reduced error distributions, while NeQuick consistently underestimated TEC, with errors increasing during the solar minimum phase. Seasonal analysis indicates that neural approaches maintain stable performance across all seasons, whereas physics-based models degrade during transitional seasons. At local time 14:00, comparative analysis shows that LSTM exhibits lower RMSE across all stations and phases, achieving a relative accuracy share greater than 40%, while physics-based models remain below 20% accuracy share. All statistical metrics, including the correlation coefficient (R), RMSE, and MRE, were computed by averaging results over stations, seasons, and solar cycle phases under both quiet and disturbed conditions. This station- and phase-resolved study demonstrates the effectiveness of deep learning architectures in ionospheric forecasting. The study enhances TEC prediction by demonstrating improved performance across diverse latitudinal regimes and solar cycle phases, providing comparative insights for space-weather forecasting.</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

Comparative evaluation of neural and physics-based models for forecasting ionospheric Total Electron Content at Low and Mid-latitude Stations during maximum and minimum phases of Solar Cycle 24

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Sunil Kumar Chaurasiya,
  • Kalpana Patel,
  • Abhay Kumar Singh

摘要

Forecasting ionospheric Total Electron Content (TEC) remains a challenging task for space weather applications due to strong variability across latitudes and solar cycle phases. This study presents a comparative analysis of neural network models, namely Long Short-Term Memory (LSTM) and Backpropagation (BP), and physics-based models, including the Thermosphere-Ionosphere-Electrodynamics General Circulation Model (TIE-GCM) and NeQuick. The evaluation is performed at three GPS stations: Novosibirsk (mid-latitude), Varanasi (low-latitude), and Port Blair (equatorial latitude), during the solar maximum year of 2014 and the solar minimum year of 2018 during Solar Cycle 24. The results reveal station- and solar cycle phase-dependent variations in model performance. In Novosibirsk, LSTM consistently achieved the highest correlation (R ≈ 0.85) and the lowest RMSE, particularly during the solar maximum phase, successfully capturing the variability of storm-time TEC. Varanasi exhibited strong seasonal variability, with neural models maintaining stable performance across both solar phases, whereas physics-based models struggled with equatorial anomaly effects, particularly during the solar minimum. Port Blair showed the largest model differences: LSTM maintained stable performance with reduced error distributions, while NeQuick consistently underestimated TEC, with errors increasing during the solar minimum phase. Seasonal analysis indicates that neural approaches maintain stable performance across all seasons, whereas physics-based models degrade during transitional seasons. At local time 14:00, comparative analysis shows that LSTM exhibits lower RMSE across all stations and phases, achieving a relative accuracy share greater than 40%, while physics-based models remain below 20% accuracy share. All statistical metrics, including the correlation coefficient (R), RMSE, and MRE, were computed by averaging results over stations, seasons, and solar cycle phases under both quiet and disturbed conditions. This station- and phase-resolved study demonstrates the effectiveness of deep learning architectures in ionospheric forecasting. The study enhances TEC prediction by demonstrating improved performance across diverse latitudinal regimes and solar cycle phases, providing comparative insights for space-weather forecasting.