Abstract: Aiming at the nonlinear and non-stationary characteristics of ionospheric total electron content (TEC) and the problems that a single LSTM model has in prediction, such as low accuracy and easy to fall into local optimality,On the basis of improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and sample entropy (SE) algorithms, combined with sparrow search algorithm (SSA) and long short-term memory (LSTM) neural network, a combined prediction model for TEC in the ionosphere was constructed, and analyzes the prediction accuracy of ionospheric TEC during the low year calm period of solar activity and the high year disturbance period of solar activity for 5 consecutive days. The experimental results show that compared with the single LSTM model and the SSA-LSTM model, the root mean square error of the combined model in this paper is reduced by 1.06 TECU and 2.25 TECU respectively under different latitude and longitude of the low solar activity quiet period and the high solar activity disturbance period. The average absolute error decreased by 0.74 TECU and 1.68 TECU respectively, and the average relative accuracy increased by 7.63% and 8.97% respectively. The prediction effect of the combined model was significantly better than that of the single LSTM model and the SSA-LSTM model.