Near-real time foF2 predictions using neural networks

Abstract

This paper describes the use of the neural network (NN) technique for the development of a near-real time global foF2 (NRTNN) empirical model. The data used are hourly daily values of foF2 from 26 worldwide ionospheric stations (based on availability) during the period 1976–1986 for training the NN and between 1977 and 1989 for verifying the prediction accuracy. The training data set includes all periods of quiet and disturbed geomagnetic conditions. Two categories of input parameters were used as inputs to the NN. The first category consists of geophysical parameters that are temporally or spatially related to the training stations. The second category, which is related to the foF2 itself, consists of three recent past observations of foF2 (i.e. real-time foF2 (F 0), 2 h (F −2) and 1 h (F −1) prior to F 0) from four control stations (i.e. Boulder (40.0°N, 254.7°E), Grahamstown (33.3°S, 26.5°E), Dourbes (50.1°N, 4.6°E) and Port Stanley (51.7°S, 302.2°E). The performance of the NRTNN was verified under both geomagnetically quiet and disturbed conditions with observed data from a few verification stations. A comparison of the root mean square error (RMSE) differences between measured values and the NRTNN predictions with our earlier standard foF2 NN empirical model is also illustrated. The results reveal that NRTNN will predict foF2 in near-real time with about 1 MHz RMSE difference anywhere on the globe, provided real time data is available at the four control stations. From the results it is also evident that in addition to the geophysical information from any geographical location, recent past observations of foF2 from these control stations could be used as inputs to a NN for near-real time foF2 predictions. Results also reveal that there is a temporal correlation between measured foF2 values at different locations.