Solar Wind Speed Time-Series Forecasting Based on Long Short-Term Memory (LSTM) Neural Network Model

Abstract

AbstractChanges in solar wind speed is one of the earliest manifestations of space weather disturbances driven by solar flare, Coronal Mass Ejection (CME), or coronal hole. An increase in solar wind speed can cause geomagnetic disturbances indicated by either Dst or Kp indices. This study proposed a Long Short-Term Memory (LSTM) Neural Network model to produce hourly forecasts of solar wind speed for the next 24 h during solar minimum. The model uses input from solar wind speed observed by Advanced Composition Explorer (ACE) satellite in Lagrangian-1 (L1) orbit as internal input and the area of the geoeffective coronal hole from Solar Dynamic Observatory (SDO)/Atmospheric Imaging Assembly (AIA) at 193 Angstrom satellite detected by DeLuna as an external input. The model has an average RMSE (Root Mean Square Error) value of 40.04 km/s. We also conducted validation analysis using the forecast product and compared it with solar wind speed data observed by Deep Space Climate Observatory (DSCOVR) satellite from October 28, 2019, to May 9, 2021. We found the forecast results 77% correct, 11% overestimate, and 12% underestimate, with a best-fit correlation of 0.85.KeywordsSolar wind speedGeoeffective coronal holeLong short-term memory