The rapid development of AI technology brings new opportunities for reactor safety. Many scholars have begun to use AI models to predict the parameter changes under reactor accident conditions, especially the long short-term memory network (LSTM). However, some AI models may not perform well due to the change of prediction target. On these bases, this paper newly establishes the gas-liquid two-phase counter-flow limitation (CCFL) prediction model by using LSTM network and Bayesian optimization. This model uses four reactor measurement parameters to determine the surge line CCFL status in different SBLOCA situations. After the network hyperparameter optimization, the model can predict CCFL well with the long-term input dataset. However, this paper also find that the ordinary LSTM model cannot accurately multi-step predict CCFL when there are only a few input points in the initial stage of SBLOCA. Therefore, this paper improves the LSTM model by using the static and rolling mechanisms. The test result shows that the former mechanism can provide faster and more accurate prediction results. Besides, this paper also tests the difference between the normal dataset and the diff dataset. The results prove that the diff dataset is suitable for data where the constant part is larger than the changing part.

LSTM,CCFL,Multi-Step prediction