Modeling and Predicting the Ionospheric Total Electron Content Over Western China With Machine Learning

Abstract

Using the Total Electron Content (TEC) data of four grid nodes at different latitudes in western China provided by the Crustal Movement Observation Network of China (CMONOC), the empirical models were established by using BP neural network (BPNN) and random forest (RF) respectively, in which the data from 2004, 2016 and 2018 were used as test sets, and the training sets included the data from 2006 to 2019 except for those test sets. In order to improve the stability of predictions of BPNN model, an integrated neural network model based on bootstrap sampling was proposed. The TEC values from CODE's 1-day predicted global ionospheric maps (C1PG) and International Reference Ionosphere (IRI) model were used for comparison in model evaluation. The results show that the machine learning methods proposed in this paper can simulate the training set well, but the RF model performs poorly in high solar activity years, and even worse than C1PG product and IRI model. BPNN model performs well at mid-latitudes, but it is not as well as C1PG product at low-latitudes. The Inte-BP model is superior to other models in all aspects. Compared with the BPNN model that only gets weights and bias parameters after one certain training session, the Inte-BP model that integrates multiple base learners can output more stable and accurate predictions.