An Improved Prediction of Solar Cycle 25 Using Deep Learning Based Neural Network

A deep-learning Vanilla, or single layer, Long Short-Term Memory model is proposed for improving the prediction of Solar Cycle 25. WDC-SILSO the Royal Observatory of Belgium, Brussels provides the 13-month smoothed sunspot-number data that were used to make this prediction. The root mean square error (RMSE) obtained by the proposed model, which is improved in comparison to the existing stacked LSTM model, lies within the range of 1.65 - 4.92, according to analysis on a number of temporal intervals taken into consideration in this study. The model performance has been validated by forecasting the peak amplitude of Solar Cycles 21 - 24. It is shown that for Cycles 21 and 22, the prediction error in estimating the peak is 1.159% and 0.423%, while the RMSE is estimated to be 4.149 and 3.274, respectively. For Cycle 23, the relative error and RMSE are 1.054% and 2.985, respectively, whereas for Cycle 24 they are 1.117% and 3.406, respectively. The current proposed model has exactly predicted the timing when the SSN reached its maximum for Cycle 23. While for Cycle 21, the prediction has a 1-month delay from the actual timing. For Cycles 22 and 24, the year during which the SSN reached maximum coincides with the observed year, although their month of peak occurrence showed a difference of three months and one month, respectively. The current proposed model suggests that the Cycle 25 will peak in April 2023 with an amplitude value of 136.9, which will be approximately 17.68% stronger compared with Cycle 24.