The ionosphere is the environment affecting the functioning of various technological systems, which is especially important for Global Navigation Satellite Systems. The state of the ionosphere can be monitored and predicted using the total electron content (TEC). The use of machine learning has led to an avalanche-like emergence of TEC forecasting methods. However, due to the strong variability of the ionosphere, the effectiveness of forecast methods depends on the coordinates where TEC is calculated, so problems may arise in choosing an appropriate method and a compromise between accuracy and training speed. To solve these problems, the accuracy of twenty models is compared. One group (DL, RS 2023) includes three subgroups of methods: (1) classic unidirectional LSTM, GRU, TCN methods with possible modifications, (2) classic LSTM, GRU with the inclusion of a bidirectional processing, (3) BiTCN (the most efficient). The second group (SN, JASTP 2024) includes one shallow neural network architecture and 10 training algorithms. The results are given for 4 European stations and 2015 with advance times τ = 2h and 24h. The comparison was carried out using the following metrics: the mean absolute error MAE, the root-mean-square error RMSE, the mean absolute percentage error MAPE. The following results were obtained for these metrics. For deep learning models (DL), there are very large differences between the values for different τ. For shallow networks (SH), these differences are small, as well as the differences between the methods inside the group. For τ = 2h, the best SN methods are at the level of the first group of DL models (FRF and RFF). For τ = 24h, the best SN methods give results close to the second group of DL models. MAPE for all SN methods does not exceed 20%. The latitude dependence for τ = 2h shows the advantage of the DL models of the second and third groups, confirming that SN results mostly limited by the DL results of the first group. For all methods, there is a latitude dependence of the metrics: MAE and RMSE increase with decreasing latitude, MAPE decreases with decreasing latitude. For τ = 24h, the results of SN methods are nearly identical to each other. The trends in the latitude dependence coincide with the result for τ = 2h. As a whole, SN methods can provide quite worthy results and a compromise between methods is possible. #CSOC1120.