Neural network model for predicting the horizontal component of Earth’s magnetic field (H) over Indian equatorial region during quiet and disturbed periods

Abstract

Artificial Neural Networks (ANNs) have proven successful in forecasting various magnetospheric and ionospheric parameters. The design of an artificial neural network (ANN) for predicting the horizontal component of Earth’s magnetic field (H) and range in H (ΔH), over Indian equatorial region, for both quiet and disturbed conditions of 23^rd solar cycle (August 1996 to December 2007) is discussed in this work. Ground magnetometer data from the stations Tirunelveli [TIR], Pondicherry [PND], Alibag [ABG], and Ujjain [UJJ] are used for training the network. Datasets from the stations Trivandrum [TRD] and Nagpur [NGP] are used for the testing procedure. The data used in this work covers the 23^rd solar cycle and it include low, moderate and high solar activity levels of both the ascending and descending phases of the solar cycle. Two sets of input parameters are used as inputs to the ANN. The first set, namely the geophysical parameters, are temporally or spatially related to the training stations. These consists of Latitude, Longitude, day of the year (DOY), local time (LT) magnetic dip angle (Inclination, I) and angle of declination (magnetic declination, D). The second set of inputs, that are driven by solar activity and affect the different stations uniformly, are Solar Flux (F10.7), Ap Index, IMF Bz, and Ion Number Density. Using these input parameters a neural network model (CCNRM) with 10 hidden neurons and 600 iterations is developed. It is found that the prediction accuracy of the model is better while training with original time series rather than the detrended time series. Here we present the prediction of H and ΔH during the quiet, disturbed geomagnetic conditions (minor storms (Dst minimum ≤ −50 nT), major storms (Dst minimum ≤ −100 nT)) along with its seasonal variation. Different datasets of the 23^rd solar cycle including low, moderate and high solar activity levels from both the ascending and descending phases of the solar cycle are faithfully reproduced by the model. The model successfully predicted the diurnal variation, seasonal variation, minor storms and major storms within the average error limits of 11 nT, 7 nT, 27 nT and 36 nT respectively during the testing process.