Recurrent Neural Networks based on LSTM for Predicting Geomagnetic Field

Abstract

The predicting accuracy of geomagnetic field is a major factor influencing magnetic anomaly detection, geomagnetic navigation and geomagnetism. The limitations of current methods consist of complex model, a large number of parameters, method of solving parameters with high complexity and low forecast accuracy during geomagnetic disturbed days. In this paper we explore a deep learning method for forecasting geomagnetic field that adopts structure of recurrent neural networks (RNN) based on long-short term memory (LSTM). This method of LSTM RNN includes analyzing the characteristics of geomagnetic field and training the data set of geomagnetic data with simple and robust mathematical model. Compared with current methods, the high-precision prediction of geomagnetic field based on LSTM RNN is achieved during both geomagnetic quiet and disturbed days. Furthermore, it could be found that the average error and maximum error of LSTM RNN are far smaller than those of the other methods.