<p>The Total Electron Content (TEC) is an important parameter that describes the morphology and structure of the ionosphere. Deep learning is an important and effective tool for forecasting TEC, but the role of different solar activity indices and geomagnetic indices in TEC prediction remains unclear. The Long Short-Term Memory (LSTM) network has special structure design and good generalization ability, which is capable of learning the features of long-term sequence data and has been widely applied in the research of ionosphere prediction. Therefore, in this study, the LSTM network is used to achieve short-term forecasting of low, middle, and high latitudes TEC during geomagnetic storms that occurred in 2016. At the same time, the effects of four different index combinations, F10.7, Kp, Dst, and AE indices, on the prediction results at different latitudes were analyzed. The results show that the appropriate combination of index inputs effectively improves the prediction performance of the model. At low latitudes, the model incorporating Kp, Dst and F10.7 indices performed best, with a 51.3% average decrease in RMSE compared to the model without any additional indices. The best model is one that uses Kp and F10.7 indices at middle latitudes, compared to model without any indices, its average RMSE decreased by 57.0%. At high latitudes, the model using Kp, Dst, and AE indices performed best, with a 43.2% average decrease in RMSE compared to the model without any indices. However, more indices do not necessarily improve prediction accuracy.</p>