Sunspot Forecast Using Temporal Convolutional Neural (TCN) Network Based on Phase Space Reconstruction

The long-term prediction of sunspot is of great significance to spaceflight, satellite and communication. Based on the chaotic characteristics of sunspot, a prediction model of TCN network based on phase space reconstruction (PSR) is established. PSR is a technology that can map a one-dimensional sunspot sequence to high-dimensional space and reconstruct the original sunspot system. Embedding dimension
$m$
and delay time τ are the most important parameters in PSR. In order to verify the effectiveness of the proposed model, four models with different τ and
$m$
are studied in the experiment. The results show that the PSR-TCN model with m = 8,
$T$
= 37 is superior to all other competitors in terms of RMSE. This study predicts 13-month smoothed monthly sunspot number of the Solar Cycle 25 using this model, making the forecast of sunspot number from January 2020 to December 2030. The maximum sunspot number is 139.55 that will occur in April 2024. This indicates that the cycle would be slightly stronger than Solar Cycle 24.