The use of sequential recurrent neural filters in forecasting the Dst index for the strong magnetic storm of autumn 2003

Neural based geomagnetic forecasting literature has heavily relied upon non-sequential algorithms for estimation of model parameters. This paper proposes sequential Bayesian recurrent neural filters for online forecasting of the Dst index. Online updating of the RNN parameters allows for newly arrived observations to be included into the model. The online RNN filters are compared to two (non-sequentially trained) models on a severe double storm that has so far been difficult to forecast. It is shown that the proposed models can significantly reduce forecast errors over non-sequentially trained recurrent neural models.