Sunspot cycle prediction: exploring data-driven and machine learning approaches

Abstract

The prediction of solar phenomena, such as sunspot activity, plays a critical role in understanding the solar cycle and its impact on space weather. This thesis explores the application of time series analysis techniques to forecast sunspot numbers during the solar minima of the 25th solar cycle, using data from the Sunspot Index and Long-term Solar Observations (SILSO). Five predictive models were evaluated: ARIMA, SARIMA, Random Forest, XGBoost, and LSTM with a focus on assessing their accuracy and robustness in predicting sunspot counts.The classical models, ARIMA and SARIMA, exhibited higher error values compared to machine learning approaches, with ARIMA recording a Mean Absolute Error (MAE) of 57.60 and a Root Mean Squared Error (RMSE) of 70.98, while SARIMA showed similar performance. These results suggest that classical models struggle to capture the underlying patterns of sunspot data. In contrast, machine learning models, particularly Random Forest, significantly outperformed the classical methods. The Random Forest model with a lag of 7 produced the lowest MAE (15.04) and RMSE (19.35), making it the most effective model in this comparison. Although XGBoost also demonstrated strong performance, increasing the lag from 7 to 12 led to a slight increase in errors, possibly due to overfitting. Additionally, a deep learning model, LSTM, was tested. The LSTM model with a lag of 7 yielded an MAE of 15.82 and RMSE of 20.91, but its performance worsened when the lag was increased to 12, highlighting the need for careful tuning and optimization in deep learning models.Overall, the study demonstrates that machine learning models, especially Random Forest with a lag of 7, provide superior predictive performance over traditional approaches. This work highlights the importance of data preprocessing, model selection, and optimization in time series forecasting, contributing to the broader field of solar physics and space weather prediction.