Short-term Prediction of Ionospheric TEC Based on the WNN-LSTM-Attention Combined Model

 To address the issues of insufficient extraction of nonlinear and non-stationary features in short-term forecasting of ionospheric total electron content (TEC) and weak model robustness under space weather disturbances, this paper proposes a combined prediction model integrating Wavelet Neural Network (WNN), Long Short-Term Memory network (LSTM), and a composite attention mechanism (WNN-LSTM-Attention). The model uses WNN to extract local multi-scale features of TEC sequences, LSTM to capture long-term temporal dependencies, and a composite attention mechanism (temporal, feature, wavelet attention) to adaptively weight key information, achieving feature complementarity and optimization. Experiments were conducted based on TEC data from seven GNSS observation stations in China from 2016 to 2018, along with Dst and Kp indices. The results show that the overall root mean square error (RMSE) of the combined model is 1.19 TECu, reducing by 48.7% and 36.3% compared with single LSTM and WNN models, respectively; under weak, moderate, and strong geomagnetic storm conditions, its mean absolute error (MAE) decreased by an average of 21.1% compared to LSTM and 12.0% compared to WNN; it also demonstrates the best stability and accuracy in seasonal forecasting. The proposed model provides an effective method to improve the prediction accuracy and robustness of ionospheric TEC under extreme space weather conditions.