Improving forecasts of geomagnetic storms with evolved recurrent neural networks

Recurrent neural networks (RNNs) have been used for modeling the dynamics of the Dst index. Researchers have experimented with various inputs to the model, and have found improvements in prediction accuracy using measurements of the interplanetary magnetic field (IMF) taken from the Advanced Composition Explorer satellite. The output of the model is the one hour ahead forecasted Dst index. Previous models have used gradient information, usually gradient descent, for optimization of RNN parameters. This paper uses the IMF inputs (that have been found to work well) to the RNN and uses a Genetic algorithm for training the RNN. The proposed model is compared to a model used in operational forecasts which relies on solar wind data and IMF parameters, as well as a model which uses IMF data only. Both of the comparison models were trained with gradient descent. A series of geomagnetic storms that so far have been difficult to forecast are used to evaluate model performance. It is shown that the proposed evolutionary method of training the RNN outperforms both models which were trained by gradient descent.