Comparative Analysis of Deep Learning Models for Daily Solar Indices Forecasting in Solar Cycle 25

Abstract

Accurate forecasting of solar activity indices, particularly the Sunspot Number (SSN) and the F10.7 solar radio flux index (F10.7), is essential for effective space weather monitoring, as severe solar and ionospheric disturbances can significantly impact satellite operations, radio communications, and navigation systems. This paper presents a comparative analysis of deep learning models—Long Short-Term Memory (LSTM), Temporal Convolutional Networks (TCN), and encoder-only Transformer architectures—for daily forecasting of SSN and F10.7 up to 14 days ahead based on past 27 days. Considering relatively simple model structures, both single-step and multi-step prediction strategies are explored to evaluate the models’ capability in handling short-and long-term dependencies in time series data. Daily solar activity data spanning seven solar cycles (Cycles 19–25), obtained from the GFZ Helmholtz Centre for Geosciences, are used for model training and evaluation. Experimental results show that LSTM consistently achieves the best performance across most forecast horizons, particularly in short-to medium-term predictions. The Transformer model delivers competitive and stable results, while TCN performs relatively less effectively, indicating the need for more complex architecture and optimization strategies. These findings highlight the strengths and limitations of each architecture for solar activity forecasting applications.