Deep Learning of Solar Wind Time–Frequency Representations for Predicting Local Ground Horizontal Magnetic Component

Abstract:

The Sun emits a stream of charged particles from its upper atmosphere known as the solar wind. Explosive events known as coronal mass ejections (CME) are also emitted. The solar wind has a variety of parameters such as magnetic field, pressure, temperature, and speed, which change over time. Fluctuations in these parameters affect the Earth's local space environment, particularly during CME events, which are known as geomagnetic storms. The earth’s magnetic field can experience strong perturbations during these storms. The local ground magnetometer measurements are sensitive to these perturbations. In this study, we observe how the local ground horizontal magnetic component can be predicted using solar wind data by using a Long Short-Term Memory (LSTM)-based deep learning network. LSTM based networks are widely used for time series data due to their sequence learning capability. The input data were obtained from Omni solar-wind data for the years 2001 to 2018. The ground horizontal magnetic components for the corresponding time, the target data, are calculated using SuperMag data for the Ottawa ground Magnetometer Station for the years 2001 to 2018. The input dataset is used to extract both time domain and frequency domain features. The optimal set of features from both the time domain and frequency domain are chosen using feature selection methods. Our study finds that a combination of both time-domain and frequency-domain features performs better than using the feature sets separately when used to train deep learning networks of similar depths.