Abstract: The life management and evaluation during the start-up process of the turbine rotor pays much attention to the rotor surface stress, while the time cost of using finite element method to calculate stress is relatively large. Recurrent neural network can quickly and accurately calculate and predict the rotor surface stress during the start-up process of the turbine rotor which greatly improves the online evaluation efficiency. However, the traditional recurrent neural network(RNN) has the problems of gradient disappearance or explosion and cannot solve the long-term dependence which affects its practical application. To solve this problem, a turbine rotor surface stress prediction model based on LSTM(long short-term memory) recurrent neural network was established. In multiple sets of experiments with hyperparameters, it is found that the prediction accuracy is optimal when the time step is 3, the dropout ratio is 0.2, and the number of hidden layer units is 6, which verifies the rationality of the prediction model. By comparing the time cost of the prediction model with that of the finite element calculation model, it is found that the time cost of the prediction model is greatly shortened, which verifies the high efficiency of the prediction model. The universality analysis of the prediction model shows that the prediction model can be extended and applied to the prediction of similar time series problems.