Time-Series Forecasting of 10.7 cm Solar Radio Flux using ResNeXt-iTransformer Hybrid Model

This work presents a hybrid deep learning framework for forecasting 10.7 cm Solar Radio Flux (F10.7), a key driver of thermospheric density, satellite drag, and space weather impacts. The proposed ResNeXt-iTransformer model combines a one-dimensional convolutional (ResNeXt-101) backbone with an inverted Transformer (iTransformer) encoder to capture both local and long-range temporal dependencies. The input feature set comprises four solar flux indices (Rotation, Obs, Adj, URSI-D), twelve Fourier seasonal and solar-cycle harmonics, Kp and ap geomagnetic indices, and sunspot numbers. Trained on daily data from 1952 to 2007 with rolling-window cross-validation, the model achieves an RMSE of 0.65, yielding a 34–52% reduction in error relative to ARIMA, LSTM, and vanilla Transformer baselines, while maintaining correlation coefficients (p > 0.9) and 22-year recursive forecasts within 10% of peak flux. The framework supports a conditional stochastic model for the geomagnetic ap index that the NRLMSIS-2.0 atmospheric model can utilize to generate multi-year thermospheric density scenarios. These results demonstrate the effectiveness of integrating convolutional feature extraction with inverted attention mechanisms for reliable solar radio flux prediction and downstream atmospheric drag modeling.