Prediction of solar cycles 26 and 27 based on LSTM-FCN

Abstract

Predicting solar activity changes is crucial for Earth’s climate, communication systems, and aerospace technology. This study employs the Long Short-Term Memory Fully Convolutional Network (LSTM-FCN) deep learning method to predict the amplitudes and peak times of Solar Cycles (SCs) 26 and 27 for both the entire solar disk and the northern and southern hemispheres. The experimental data comprises the monthly mean total Sunspot Number (SSN) data and the monthly mean northern and southern Hemispheric Sunspot Number (HSN) data, provided by the World Data Center — Sunspot Index and Long-term Solar Observations (WDC-SILSO). The experimental process tested the Input–Output Window Ratio (IOWR) from 4:1 to 14:1, and the results indicate that when the IOWR is 10:1, the normalized Relative RMSE (RRMSE) is minimized at 0.078. According to the prediction, SC 26 is expected to peak in June 2034 with an amplitude of 194.4, and SC 27 is expected to peak in July 2045 with an amplitude of 244.2. It was also found that SC 26 and SC 27 have northern and southern hemisphere asymmetry. This study demonstrates the potential application of the LSTM-FCN deep learning method in forecasting SCs, providing a new tool and approach for solar physics research.