Abstract: Ionospheric delay is one of the most important sources of error in global satellite navigation and positioning. Improving the prediction accuracy of ionospheric Total Electron Content (TEC) is very important to improve the positioning accuracy of satellite navigation. In this paper, we combine sliding window and Long-Short-Term Memory (LSTM) neural network, and use sliding window algorithm to continuously update the input time series data set. We tested the accuracy of the models corresponding to different input sequence lengths and recorded them, and found that the accuracy of the last 10% of the input data series was the best when the predicted value was updated. Finally, we used the sliding window method to update the last 10% of the input data series with the predicted value to build the TEC prediction model. The newly constructed model, traditional LSTM model and BP model are used to predict the same TEC time series data, and Root Mean Square Error (RMSE), absolute residual error and Mean Absolute Error (MAE) are used to evaluate the accuracy of the model prediction results, and verify the prediction performance of the new model. The experimental results show that the proportion of residual absolute value less than 5 TECU predicted by the newly constructed model SLSTM (Sliding Window on Long-Short-Term Memory) in both the calm period and the magnetic storm period exceeds 85%, and the proportion of predicted residual absolute value less than 5 TECU corresponding to the traditional LSTM model increases by 49% and 71%. On the other hand, compared with the traditional LSTM model, the root-mean-square error of the new model is reduced by 31% and 35%, respectively, and the average absolute error is reduced by 25% and 32%, respectively. In addition, we can also see that the RMSE mean values and MAE mean values of SLSTM model are smaller than those of traditional LSTM model and BP model.