Sequential Modeling of D_st Dynamics with SEEk Trained Recurrent Neural Networks

Abstract

A sequential framework for modeling magnetospheric plasma interactions with a SEEK trained recurrent neural network is proposed. An overview of the state-space modeling framework is provided, along with a review of previous Kalman trained neural models. The proposed algorithm is described and is evaluated against an EKF trained RNN and a gradient based model. The exogenous inputs to the RNNs consist of three parameters, Bz, B^2, and (By)^2, where B, Bz, and By represent the magnitude, the southward and azimuthal components of the interplanetary magnetic field (IMF) respectively. It was found that the SEEK trained recurrent neural network outperforms other neural time series models trained with the Extended Kalman Filter, and gradient descent learning. The numerical simulations suggest that the SEEK filter provides superior tracking capabilities than the EKF, resulting in accurate forecast of the Dst index.