Predicting foF2 in the China region using the neural networks improved by the genetic algorithm

Abstract

A regional model for the ionospheric critical frequency of the F2 layer (foF2) over China is developed by neural networks (NNs) trained by the genetic algorithm (GA). In order to avoid the ‘local minimum’ phenomena in most NN applications, GA is utilized here to optimize the initial weights of NNs. The input parameters used in this GA-NN based foF2 prediction model include Beijing time (BJT, GMT+8), day number (day of the year), seasonal information, solar cycle information, magnetic activity, magnetic declination, magnetic dip angle, angle of meridian relative to sub-solar point, solar zenith angle, and geographic coordinates. The foF2 datasets employed in this model are obtained from nine ionosonde stations located in China for the time span of 1990–2004 that covers one entire sunspot cycle. The datasets for 1996 and 2000 are selected for validation instead of for training use. Prediction results of GA-NN model, unimproved NN model, and International Reference Ionosphere 2007 (IRI2007) model (from International Radio Consultative Committee (CCIR) coefficient) are compared with the observation data for the year of 1996 and 2000 respectively. The results indicate that GA-NN model is superior to the unimproved NN model and the IRI2007 model for foF2 prediction. According to the statistical analysis of average RMSE, the GA-NN method offers an improvement of 4.89% over NN method and an improvement of 27.79% over IRI2007 model. The improvement of accuracy for one single station forecasting is validated with the data from the Wuhan ionosonde station both at 00:00 (BJT) and 12:00 (BJT) in 1996 (solar minimum) and 2000 (solar maximum).