Features of Using Neural Network Methods to Predict the Total Electron Content of the Ionosphere in the Southern Hemisphere

Research of the behavior, as well as solving the problem of predicting the total electron content of the TEC of the ionosphere, is of great importance for studying and taking into account the influence of space weather on various technological systems. This influence is different in various parts of the globe, which is reflected in the dependence of the forecast accuracy on the position of the station. In the previous works of the authors, neural networks were developed using several groups of unidirectional and bidirectional, as well as recurrent and convolutional architectures, which provided high forecast accuracy based on data from stations along three meridians (European 30° E, Southeastern 110° E, and American 75° W). In the present paper, these models, with additional modifications, are used to determine the accuracy of the TEC forecast for stations in the Southern Hemisphere lying along the American meridian ~ 70° W. A comparison of the results obtained with previous estimates for the American meridian of the Northern Hemisphere showed that metrics such as Mean Absolute Error MAE, Root Mean Square Error RMSE, Mean Absolute Percentage Error MAPE follow the latitudinal structure of the ionosphere. MAE values are slightly higher in the Southern Hemisphere, and in both hemispheres are less than 0.4 TECU, with an average of 0.236 TECU in the Southern Hemisphere and 0.189 TECU in the Northern Hemisphere. RMSEs range from 0.5–1.5 TECU, with an average of 0.591 TECU in the Southern Hemisphere and 0.857 TECU in the Northern Hemisphere. MAPEs are less than 2%, with an average of 1.5% in the Southern Hemisphere and 1.9% in the Northern Hemisphere. An assessment of the forecast accuracy during three magnetic storms of different intensity (minimal Dst = −98 nT, −114 nT, −155 nT) showed the dependence of the accuracy of the forecast of ionospheric parameters on the latitude of the stations at which the measurements were carried out. The BiLSTM-FRF architecture reflects the nature of the disturbed state quite well, and the values for the BiTCN method actually coincide with the observed TEC. #CSOC1120.