Abstract: The prediction of outlet-water temperature difference of stator bars (outlet-water temperature difference for short) is of great significance to ensure the safe operation of turbo-generator. However, the working conditions of turbo-generator are variable, resulting in the complex pattern of temperature difference sequence, which makes it difficult to predict the trend accurately. In this paper, long short-term memory (LSTM) neural network is used to learn complex patterns. And Autoregressive Integrated Moving Average Model (ARIMA) is further combined to make up for the lack of training caused by variable working conditions, so as to correct the prediction of LSTM. Experiments are carried out on the QFSN-660-2-22 turbo-generator. Results show that this combined method performs better than the single LSTM or ARIMA. This method can be applied to short-term early warning, of which the accuracy is higher than 95%.