A study on machine learning for TEC Map forecast over low-latitude regions

The ionospheric total electron content is one of the main sources of error for GNSS single point positioning users. This is especially true over low-latitudes regions due to the intensified dynamics that depend on several parameters. Due to this large variability, models such as NeQuick-G (based on IRI-2016) and IONEX2 have a limited performance in describing the spatiotemporal responses over these regions. Currently, machine learning techniques have been used in data analysis and in the prediction of physical phenomena in several areas, including the ionosphere. In this work, we propose a TEC prediction model based on machine learning for low-latitude regions such as Brazil. A Multilayer Perceptron network is proposed and evaluated. TEC maps from the previous five days were used together with parameters such as solar flux, geomagnetic activity indexes, hour, day of year and spatial location (latitude and longitude). The performance of the proposed neural network was evaluated for 12 months during the maximum period of the last solar cycle. The results obtained showed that the prediction model based on machine learning was able to successfully describe the TEC distribution over the Brazilian region. A comparison with NeQuick-G and IONEX2 models reveals that the results of the proposed method have an error that can be up to 50% smaller than the other models currently used.