Multi-step probabilistic forecasting for sunspot numbers based on LightGBM

Abstract

As a comprehensive indicator of solar activity encompassing the entire visible disk, sunspot number (SSN) plays a pivotal role in both space weather forecasting and anomaly event monitoring. In this study, we propose a straightforward yet effective probabilistic model based on light gradient boosting machine (LightGBM) for multi-step ahead prediction of sunspot numbers. To achieve this, by leveraging the seasonal-trend decomposition using locally estimated scatterplot smoothing (Loess) method known as STL, we decompose the trend, seasonal variations, and residual components from the time series of sunspots, and these components are used as input features. The stepwise optimization algorithm in Optuna is employed to fine-tune the model’s hyperparameters. We conduct a comprehensive performance analysis of the proposed model using the SSN dataset, covering the period of 1755/02-2024/10. Our experimental results demonstrate that the proposed model outperforms existing methods by reducing errors. Furthermore, through quantitative uncertainty and probabilistic analysis, we establish the reliability of prediction intervals in most cases, enabling effective anomaly detection during anomalous events.