A Precise Global Ionospheric Total Electron Content Forecasting Model Based on Multi-Neural Network Ensemble

Abstract The ionospheric total electron content (TEC) is a crucial parameter for studying the dynamic changes in the ionosphere. Accurate forecasting of TEC is significant for research related to space weather phenomena such as auroras and magnetic storms, as well as for long-distance communication and high-precision positioning using global navigation satellite systems (GNSS). Due to the nonlinear and highly irregular distribution of global TEC, existing forecasting models exhibit low efficiency. This study proposes a high-precision forecasting model for global TEC based on the squeeze-and-excitation (SE) attention mechanism and a combination of convolutional neural networks (CNN) and bidirectional long short-term memory (BILSTM) networks. In the data preprocessing stage, the SHAP value algorithm is employed to extract the six most significant feature parameters contributing to TEC forecasting. The model then leverages CNN and BILSTM algorithms to thoroughly explore both long and short-term dependencies in TEC data, while the SE attention mechanism is utilized to redistribute weights to the critically influential features, enabling precise forecasting of global TEC. Forecasting experiments were conducted on global TEC, and magnetic storms were categorized based on geomagnetic indices to investigate the model's accuracy across different storm levels. The experimental results indicate that the new model proposed in this study achieves an average accuracy of 2.59 TECU for ionospheric TEC forecasting, significantly outperforming similar models. When compared to the currently best-performing model, this new approach demonstrates a 24.3% improvement in accuracy, along with a marked reduction in training time. These findings suggest that the new CNNBILSTM_SE model developed in this research enhances forecasting precision, shortens model training duration, and improves the overall efficiency of forecasting models. This advancement holds significant research implications and practical value for applications in space weather prediction and high-precision GNSS positioning.