Ionospheric TEC prediction in low-latitude Indian region during geomagnetic storm periods based on XGBoost with optuna framework and comparison with IRI-Plas 2020

Abstract

This paper introduces a strong 24-h TEC forecast model based on the Extreme Gradient Boosting (XGBoost) algorithm, implemented for GPS-based TEC data at both the IISC and Chum stations. For a better tuning of the model's performance, two hyperparameter tuning approaches—Optuna and Hyperopt—were utilized. The assessment pertains to five primary GS events: the Halloween Storm (2003), St. Patrick's Day Storm (2015), the February 2022 Mother's Day Storm (2024), and yet another GS in 2024. The AI model forecasts were compared against the IRI-PLAS 2020 empirical model on the basis of standard performance measures such as Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Mean Squared Logarithmic Error (MSLE). Results show that the Optuna-tuned XGBoost model always performed better than the IRI-PLAS 2020 model for all GS events. RMSE values for Optuna were considerably lower: 4.7254 for the Halloween Storm versus 8.7335 using IRI-PLAS, 2.8389 versus 17.4692 for St. Patrick's Day, 4.4167 versus 9.9611 for the 2022 storm, 10.8861 versus 16.4781 for the Mother's Day event, and 12.8778 versus 29.8309 for the second 2024 storm. For comparison, the XGBoost model optimized using Hyperopt, tested over the same events, had RMSEs of 14.52, 7.01, 5.19, 6.80 and 14.97, respectively. While Hyperopt outperformed the IRI-PLAS model in each scenario, Optuna produced the best results overall. These results confirm the effectiveness of machine learning, specifically the XGBoost model optimized using Optuna tuning, in observing intricate TEC dynamics during geomagnetic disturbances. The research points out how AI-driven forecasting can dramatically enhance space weather resilience by limiting error in prediction and providing more trustworthy GNSS and communication system functionality under high-impact solar activity.