ИСТИНА

Different models for forecasting relativistic electron fluences at geostationary orbit were developed. The input parameters included 11 variables: geomagnetic activity indices, solar wind parameters, interplanetary magnetic field, and two energetic channels of relativistic electrons (>2 MeV and >0.8 MeV). The electron flux data were obtained from the GOES-15 satellite, covering nearly the entire 24th solar cycle from early 2011 to late 2019. A first group of models based on gradient boosting techniques and LSTM were developed to forecast daily electron fluences three days ahead. The dataset was split into training, validation, and testing subsets in a ratio of 0.7/0.15/0.15. Results showed that the CatBoost algorithm outperformed other gradient boosting methods, while LSTM performed best among the recurrent models. For the test set from the second half of 2018 to the end of 2019, the R² scores for three-day-ahead forecasts were 0.89, 0.69, and 0.49, respectively. Since these results were obtained during the solar cycle minimum, to assess forecast quality across different years, the main dataset was divided into overlapping four-year samples using a sliding window approach. Each sample consisted of two years for training, one year for validation, and the subsequent year for testing. The forecast accuracy was found to depend on the electron flux intensity: higher flux levels led to lower R² values and higher RMSE. Consequently, forecast accuracy deteriorates as the solar cycle progresses through its declining phase, with this effect becoming more pronounced at longer forecast horizons. The contribution of each input parameter was also analyzed. It was found that the most significant variables are the Kp and SYM-H indices, ion temperature, and solar wind speed.