In order to solve the problem of high miss rate and high false alarm rate in the online grade prediction of military vocational education, this paper proposes a two-stage stage long short memory (TS-LSTM) grade prediction model based on online learning grade analysis. The weighted loss (WL) function assigns higher weights to a few categories to achieve a high detection rate and a low false alarm rate; the LSTM is modified to a short-term gate-long short-term memory network (STG-LSTM) to simulate short-term changes in online learning performance, which effectively suppresses a large number of false alarms associated with the prediction of risk category labels. The experimental results show that the model can highlight the changes in the course of students' reexamination, and has high accuracy; compared with other models, the F1 score has increased by 28.8%.