Abstract:Aiming at the high noise, nonlinear and non-stationary dynamic characteristics of ionospheric total electron content(TEC) time series, we construct an improved short-term ionospheric combined prediction model based on singular spectrum analysis(SSA) and long short term memory(LSTM) neural network model, to realize the model’s ionospheric TEC prediction during magnetic storms and magnetic quiet periods and analyze its accuracy. The results show that the relative accuracy of model is 91.17% and 95.46% respectively, which is 4.92 percent and 3.17 percent higher than that of single LSTM model, during the period of magnetic explosion and magnetic calm.