Sunspot cycle prediction using Warped Gaussian process regression

Abstract

Solar cycle prediction is a key activity in space weather research. Several techniques have been employed in recent decades in order to try to forecast the next sunspot-cycle maxima and time. In this work, the Gaussian process, a machine-learning technique, is used to make a prediction for the solar cycle 25 based on the annual sunspot number 2.0 data from 1700 to 2018. A variation known as Warped Gaussian process is employed in order to deal with the non-negativity constraint and asymmetrical data distribution. Tests using holdout data yielded a root mean square error of 10.0 within 5 years and 25.0–35.0 within 10 years. Simulations using the predictive distribution were performed to account for the uncertainty in the prediction. Cycle 25 is expected to last from 2019 to 2029, with a peak sunspot number about 117 (110 by the median) occurring most likely in 2024. Thus our method predicts that solar Cycle 25 will be weaker than previous ones, implying a continuing trend of declining solar activity as observed in the past two cycles.