Modeling and forecasting of ionospheric total electron content using a radial basis function neural network over Lagos, Nigeria

Accurate prediction of ionospheric Total Electron Content (TEC) remains a major scientific and operational challenge over low-latitude regions due to strong nonlinear electrodynamic processes, equatorial plasma dynamics, and seasonal variability. In this study, a Radial Basis Function Neural Network (RBFNN) was developed to forecast TEC over Lagos, Nigeria using GPS-derived TEC observations from 2010 to 2016 for training and the independent year 2017 for testing, thereby ensuring a realistic out-of-sample evaluation. The RBFNN performance was critically assessed against the empirical IRI-2020 model using multiple statistical metrics including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Bias Error (MBE), Nash–Sutcliffe Efficiency (NSE), Willmott’s index of agreement (d), Kling–Gupta Efficiency (KGE), correlation analysis, and Taylor diagram metrics. Results reveal that IRI-2020 exhibits persistently large prediction errors, with MAE values reaching 23.23 TECU (April) and RMSE values as high as 27.25 TECU (May), alongside strong systematic overestimation reflected by MBE values up to 18.64 TECU (September). Across the major seasonal phases (March equinox, June solstice, September equinox, and December solstice), the models exhibited clear contrasts. Noticeably during the March equinox, GPS-TEC peaked near 55 TECU with the RBF predicting about ∼53 TECU compared to ∼44 TECU from IRI-2020, while during the June solstice GPS-TEC reduced to about 40 TECU and was closely matched by the RBF (∼38 TECU) but underestimated by IRI-2020 (∼34 TECU). More critically, IRI-2020 produced negative NSE values across all evaluated months, reaching −3.66 in September, indicating performance worse than the observational mean. In contrast, the RBFNN demonstrated excellent predictive capability, achieving MAE values as low as 1.12 TECU (September) and RMSE values as low as 1.50 TECU in the same month, representing nearly an order-of-magnitude improvement over IRI-2020. The model exhibited strong statistical reliability with NSE values up to 0.986 (September), Willmott’s index reaching 0.9965, and KGE values as high as 0.9587, confirming accurate reproduction of correlation, variability, and mean structure of TEC. The Taylor diagram analysis corroborated these findings, showing higher correlation and lower RMSD for the RBFNN relative to IRI-2020. These results demonstrated that a locally trained RBF neural network provides substantially improved TEC forecasting performance over Lagos and represents a viable operational alternative to global climatological models for GNSS reliability and space weather resilience in equatorial Africa.