Ionospheric foF2 disturbance forecast using neural network improved by a genetic algorithm

Abstract

A single station short-term ionospheric disturbance forecasting model has been developed with a genetic algorithm-based neural network (GA-NN). The genetic algorithm is used to optimize the initial weights of the neural network to avoid the local minimum during NN training. Using this model, the single station predictions of the ionospheric F2 layer critical plasma frequency, foF2, with the time-scale 1–24 h in advance in the China region have been investigated. Input parameters of the forecasting GA-NN model include the Beijing time (GMT+8), season information, solar zenith angle, day number, solar activity, geomagnetic activity, neutral winds, geographic coordinates and previous values of foF2. The training dataset in this model are obtained from the ionosonde stations in China. The data coverage is from 1990 to 2004 (more than one solar cycle) except 1995 and 2000. The data of ionospheric disturbances in 1995 (solar minimum) and 2000 (solar maximum) are used as the validation dataset. The prediction results at the different stations show that 1 h ahead prediction is more accurate than predictions of 3, 6, 12 and 24 h ahead. Comparisons between the observed and predicted values of foF2 in the low and middle latitudes during the year of solar minimum (1995) and solar maximum (2000) indicate that, the prediction accuracy at middle latitudes are generally better than that at low latitudes. The prediction root-mean-square error (RMSE) in the low solar activity is smaller than that in the high solar activity. The ionospheric disturbances prediction results manifest that the model works well even when the observed values of foF2 are far away from the monthly median value and the ionospheric storm lasts for 18 h.