This article introduced a hybrid neural network method (CNN + BiLSTM) to predict the long-term status of a nuclear power plant (NPP) under nuclear accident conditions, which is an innovative analytical method for the safety tolerance estimation in the accident management. Considering the lack of real nuclear accident dataset, we firstly use the severe accident analysis program to simulate a large number of accident sequences, which are referred to the Final Safety Analysis Report (FASR). Then the data-driven method in deep learning has been adept at discovering patterns from nuclear accident data and making inferential predictions. A method combining accident mechanism and sensitivity analysis has been applied for the model features engineering. Through this predictive method, we estimated the long-term connections of the parameters in the time series data for nuclear accidents, which are selected to forecast the progression of accident deterioration with limited historical observations. Finally, we applied the prediction model to the general second-generation NPPs of China. Moreover, the results have shown that the method provides an efficient way to predict the future deterioration status of the NPP under accidents, and gives timely warnings of risks to assist accident management.