Abstract
<jats:p>Представлен комплексный подход к обработке данных полного электронного содержания (ПЭС) ионосферы, включающий восстановление пропусков и прогнозирование вариаций ПЭС. Для восстановления пропусков разработан алгоритм на основе метода машинного обучения Random Forest с динамическими сдвиговыми предикторами на основе данных смежных ГНСС-станций, обеспечивающий высокую точность (R² > 98,4 %, MAPE < 4,3 %) при длительности пропусков до суток. Для прогнозирования предложена методика автоматизированного синтеза архитектур LSTM-сетей с подбором гиперпараметров. Исследования на базе данных ГНСС за 2021–2023 гг. показали, что расширение обучающей выборки снижает ошибку прогноза в четыре раза и обеспечивает устойчивость моделей на горизонтах до семи суток (R² ≈ 98 %). Основные погрешности обоих методов связаны с периодами повышенной геомагнитной активности.</jats:p> <jats:p>Introduction. Ionospheric distortions are a key factor limiting the accuracy and reliability of wideband satellite communication systems. The Total Electron Content (TEC) of the ionosphere serves as the primary parameter for assessing and correcting such distortions. However, the practical application of TEC data is hindered by gaps in measurements and the lack of predictive forecasting tools. Objective. Development of a comprehensive approach to TEC data processing that enables the recovery of gaps in experimental time series and the prediction of future TEC variations to implement predictive correction of ionospheric distortions within prediction interval. Methods. For missing data recovery, an algorithm based on the ensemble Random Forest method is proposed, utilizing dynamic shift predictors and data from adjacent Global Navigation Satellite Systems (GNSS) stations. For predicting future TEC variations, a methodology for the automated synthesis of Long Short-Term Memory (LSTM) recurrent neural network architectures with hyperparameter optimization was developed. Experimental studies were conducted on data from a network of nine reference GNSS stations in the Samara region for the period 2021–2023, employing seasonal cross-validation. Results. The recovery algorithm ensures accuracy with R² > 99.4% and MAPE < 4.3% for gaps lasting up to one day, and demonstrates robustness to seasonal variations. Increasing the volume and spatiotemporal representativeness of the training sample reduces the prediction error by a factor of four and ensures model stability for prediction horizons of up to seven days, with a coefficient of determination around 98%. Conclusions. The main errors of the gap recovery algorithm and the TEC variation prediction method are associated with periods of anomalous geomagnetic activity. This highlights the promise of incorporating geophysical indices of solar and geomagnetic activity into the feature space of the models as a priority direction for future research.</jats:p>