Forecasting the Onset of Solar Cycle 26 using Deep Learning and Multi-Indicator Solar Activity Data

抄録

Accurate forecasting of solar cycle onsets is pivotal for space weather preparedness and for advancing our understanding of solar dynamo mechanisms. This study addresses this challenge by employing the optimized Long Short-Term Memory network (LSTM+), tailored for iterative, multi-step ahead prediction, to forecast the commencement of Solar Cycle 26 (SC26). Leveraging historical data from SC18 through SC24, the LSTM+ framework was trained and rigorously validated using three key solar activity indicators - sunspot number (SSN), sunspot area (SSA), and the solar flare index (SFI). The validation across SC23 and SC24 demonstrated the model’s robust capability to capture fundamental cycle dynamics with high fidelity, consistently outperforming baseline statistical and simpler neural network models. Critically, the application of the validated LSTM+ model to the ongoing SC25 yields a convergent prediction across all three solar indices, projecting the onset of SC26 to occur around May 2030. This timing implies a duration of approximately 10.4 years for SC25. This work thus provides a novel, data-driven, and multi-indicator-supported estimate for the next solar minimum, contributing a significant input for long-range solar activity forecasting.