SAM-LSTM: Prediction of the Sunspot Cycles Based on LSTM-Attention Model with Soft Attention Mechanism
Xinhua Miao, Binhong Li · 2024
This paper aims to develop a reliable mathematical model for predicting sunspot characteristics and their impact on various fields. Previous studies have used methods such as time series analysis, spectral analysis and neural networks to predict sunspots, but the results are not ideal. This study proposes a new mathematical model designed to more accurately predict sunspot cycles and explain their reliability. We will use publicly available observational data to provide improved methods and results for sunspot predictions. We collected the required data on the number of sunspots, magnetic field strength, etc. from relevant websites. In the data preprocessing stage, we filled missing values, handled outliers, and performed spectral analysis on the data through exploratory data analysis methods such as scatter plot visualization and trend exploration. We use Fourier transforms to explore periodicity and integrate relevant annual, monthly, and daily data into period-based data. Because of the relatively poor performance of ARIMA, we introduce the LSTM-Attention model based on soft attention mechanism to fit and predict the original and the integrated data. By comparing the results with NAOC (National Astronomical Observatories, CAS) and NASA (National Aeronautics and Space Administration), we achieved relatively accurate predictions of the start and duration of the current and next sunspot cycles.