摘要
A neural network model of the Global Navigation Satellite System - vertical total electron content （GNSS-VTEC） over Nigeria is developed. A new approach that has been utilized in this work is the consideration of the International Reference Ionosphere＇s （IRI＇s） critical plasma frequency （foF2） parameter as an additional neuron for the network＇s input layer. The work also explores the effects of using various other input layer neurons like distur- bance storm time （DST） and sunspot number. All available GNSS data from the Nigerian Permanent GNSS Network （NIGNET） were used, and these cover the period from 2011 to 2015, for 14 stations. Asides increasing the learning accuracy of the networks, the inclusion of the IRI＇s foF2 parameter as an input neuron is ideal for making the networks to learn long-term solar cycle variations. This is important especially for regions, like in this work, where the GNSS data is available for less than the period of a solar cycle. The neural network model developed in this work has been tested for time-varying and spatial per- formances. The latest 10% of the GNSS observations from each of the stations were used to test the forecasting ability of the networks, while data from 2 of the stations were entirely used for spatial performance testing. The results show that root-mean-squared-errors were generally less than 8.5 TEC units for all modes of testing performed using the optimal network. When compared to other models, the model developed in this work was observed to reduce the prediction errors to about half those of the NeQuick and the IRI model.
A neural network model of the Global Navigation Satellite System - vertical total electron content （GNSS-VTEC） over Nigeria is developed. A new approach that has been utilized in this work is the consideration of the International Reference Ionosphere＇s （IRI＇s） critical plasma frequency （foF2） parameter as an additional neuron for the network＇s input layer. The work also explores the effects of using various other input layer neurons like distur- bance storm time （DST） and sunspot number. All available GNSS data from the Nigerian Permanent GNSS Network （NIGNET） were used, and these cover the period from 2011 to 2015, for 14 stations. Asides increasing the learning accuracy of the networks, the inclusion of the IRI＇s foF2 parameter as an input neuron is ideal for making the networks to learn long-term solar cycle variations. This is important especially for regions, like in this work, where the GNSS data is available for less than the period of a solar cycle. The neural network model developed in this work has been tested for time-varying and spatial per- formances. The latest 10% of the GNSS observations from each of the stations were used to test the forecasting ability of the networks, while data from 2 of the stations were entirely used for spatial performance testing. The results show that root-mean-squared-errors were generally less than 8.5 TEC units for all modes of testing performed using the optimal network. When compared to other models, the model developed in this work was observed to reduce the prediction errors to about half those of the NeQuick and the IRI model.
