Exploring Three Recurrent Neural Network Architectures for Geomagnetic Predictions

Three different recurrent neural network (RNN) architectures are studied for the prediction of geomagnetic activity. The RNNs Elman, gated unit (GRU), and long short-term memory (LSTM). take solar wind data as inputs to predict Dst index. index summarizes complex processes into a single time series. models trained tested using five-fold cross-validation based on hourly resolution OMNI dataset from years 1995–2015. plasma (particle density speed), vector magnetic fields, year, day. regularized early stopping dropout. We find that both perform better than Elman model; however, we see no significant difference in performance between GRU LSTM. with dropout require more weights reach same validation error networks without However, gap training becomes smaller when is applied, reducing over-fitting improving generalization. Another advantage it can be applied during provide confidence limits predictions. increase increasing magnitude: consequence less populated input-target space events large values, thereby uncertainty estimates. best have test set RMSE 8.8 nT, bias close zero, linear correlation 0.90.