LSTM Model for Ring Current Proton Fluxes

The widely used Artificial Neural Network (ANN) have been used to predict space plasmas that have a strong correlation with geomagnetic indices. While the ~10 keV protons have strong correlation with geomagnetic indices, the ~100 keV protons have a long decay timescale, and their fluxes mainly response to moderate-large storms. The Long-short term memory (LSTM) neural network, which is a recurrent neural network (RNN) that keeps both “long-term memory” and “short-term memory”, is ideal for predicting a time sequence of a quantify. We built a LSTM model for predicting ring current proton fluxes. This LSTM model slightly outperform the ANN model on predicting the low-energy 10s keV proton fluxes. On predicting the high-energy 100s keV proton fluxes, however, the LSTM model can stably produce main features including the long-term decay pattern, while the ANN model are typically too sensitive to the “short-term memory”, i.e., the geomagnetic indices in the near history. Our study suggests that the LSTM is ideal and efficient in modeling the ring current proton dynamics, especially in producing the time sequence of the particle distribution.