Auroral electrojet predictions with dynamic neural networks

Neural networks with internal feedback from the hidden nodes to the input [Elman, 1990] are developed for prediction of the auroral electrojet index AE from solar wind data. Unlike linear and nonlinear autoregressive moving-average (ARMA) models, such networks are free to develop their own internal representation of the recurrent state variables. Further, they do not incorporate an explicit memory for past states; the memory is implicitly given by the feedback structure of the networks. It is shown that an Elman recurrent network can predict around 70% of the observed AE variance using a single sample of solar wind density, velocity, and magnetic field as input. A neural network with identical solar wind input, but without a feedback mechanism, only predicts around 45% of the AE variance. It is also shown that four recurrent state variables are optimal: the use of more than four hidden nodes does not improve the predictions, but with less than that the prediction accuracy drops. This provides all indication that the global-scale auroral electrojet dynamics can be characterized by a small number of degrees of freedom.