Machine Learning-Based Ionospheric Modelling Performance During High Ionospheric Activity

This paper presents a slant Total Electron Content (STEC) forecasting machine learning (ML) model focusing on the region of Europe and North Africa, and on the use of spatial as well as temporal information. The input parameters of the ML ionosphere model are STEC (dual frequency observations and a calibration procedure for the satellite and station instrumental biases) along with variables characterising the solar activity (F10.7 parameters) and the Earth’s geomagnetic activity (Kp/Ap indicators). We present comparison results among the ML model, Klobuchar and NeQuick ionospheric model of the STEC / Vertical Total Electron Content (VTEC). The STEC and VTEC are calculated using Global Navigation Satellite System (GNSS) observations on a super geomagnetic storm of 24th solar cycle that occurred on 17 March 2015. We compare the predictions at low latitude station ‘RABT’, mid latitude station ‘GOP6’ and a high latitude station ‘NYA2’. Our results demonstrate that the ML based modelling approach is comparatively more sensitive and robust than the existing empirical and climatological ionospheric models, in reflecting the temporal and spatial ionospheric characterisations during the high ionospheric events. When compared with the observations, ML based approach have comparatively best KPI’s (?, RMSE and R2 ) overall. Generally the presented ML model is performing better than NeQick and Klobuchar model but the NeQuick model’s forecast is better than the Klobuchar model’s forecast. In the future, this approach could be used to forecast data for non-average ionospheric scenarios. Correlation of the observations are highest to the ML models prediction, intermediate to the Nequick ionospheric model and lowest to the Klobuchar ionospheric model. In general presented ML model works fair for the mid- and high- latitude regions.