Prediction of the Dst Index and Analysis of Its Dependence on Solar Wind Parameters Using Neural Network (original) (raw)

In this work, we propose an artificial neural network (ANN) with seven input parameters for the prediction of disturbance storm time (Dst) index 1 to 12 hr ahead. The ANN uses past near-Earth solar wind parameter values to forecast the Dst. The input parameters are the solar wind interplanetary magnetic field, north-south component of interplanetary magnetic field, temperature, density, speed, pressure, and electric field. The ANN was trained on the data period from 1 January 2007 to 31 December 2015, which contains 78,888 hourly data samples. While the period from 1 January 2016 to 31 May 2017 was used to test the prediction capabilities of the ANN. Several ANN structures were tested and the best results were determined using the correlation coefficient (R) during the training and prediction phases. The results indicate an adequate accuracy of R = 0.876 for prediction 2 hr in advance and R = 0.857 for prediction 12 hr in advance. The power of the proposed ANN was illustrated using the strongest six storms recorded during the prediction period. Generally, the duration and number of the input parameters significantly affect the training and prediction performance of the applied ANN. The results are outstanding in term of accuracy when considering a medium-term prediction of 12 hr in advance and in terms of timing of the Dst minimum occurrence. In addition, the results show a strong dependence on the solar wind electric current. Plain Language Summary We propose an artificial neural network for the prediction of disturbance storm time (Dst) index 1 to 12 hr ahead. The ANN uses 24 past hourly solar wind parameters values to forecast the Dst index. The input parameters are the solar wind interplanetary magnetic field, southward component of interplanetary magnetic field, temperature, density, speed, pressure, and electric field. Several ANN structures were tested and the best results were determined using the correlation coefficient (R) during the training and prediction phases. The results indicate an adequate accuracy of R = 0.876 for prediction 2 hr in advance and R = 0.857 for prediction 12 hr in advance. Generally, the duration and number of the input parameters significantly affect the training and prediction performance of the applied ANN. The results are outstanding in term of accuracy when considering a medium-term prediction of 12 hr in advance and in term of timing of the Dst minimum occurrence. In addition, the results show a strong dependence on the solar wind electric current.