Using Convolutional Neural Networks for TEC Prediction Accuracy Improvement

A large number of publications point to the great practical importance of predicting such an ionospheric parameter as the total electron content. At the same time, the accuracy of the forecast depends on the combination of methods used, the region and space weather conditions. Solving the problem of forecasting in dynamic conditions leads to the formation of advanced approaches to forecasting based on neural networks. The authors previously presented a paper at the 12th Computer Science Online Conference 2023 in which they developed and tested 8 architectures using the example of two Russian GPS receiver chains along the meridian 30° E and latitude ~70° N in 2015. The most widely used GIM IGS maps with a resolution of 2 h were taken as TEC data. The models used included the traditional neutral networks LSTM and GRU with proposed modifications to the RFF and FRF layer structure, and the focus was on complementing these models with bidirectional architectures. In this paper, models based on convolutional neural networks are added to the previously proposed models, and the calculations are carried out on an extended set of solar and geomagnetic activity indices, using the data for 2015, 2020 and 2022. The semi-annual statistics for the forecast accuracy with 2 h lead time showed a decrease in all error characteristics MAE, MAPE, RMSE by approximately 1.5–2 times for bidirectional architectures. It is shown that BiTCN architecture provided the highest prediction accuracy for all stations: MAE was less than 0.2–0.3 TECU, MAPE was less than 3% in 2015 and 2022, and it was less than 5.5% in 2020, RMSE was less than 0.3–0.5 TECU. The fact that the difference in accuracy in the considered ranges of latitude and longitude is negligible for this architecture is of interest. #COMESYSO1120.