Host: The Japan Society of Mechanical Engineers
Name : The 30th International Conference on Nuclear Engineering (ICONE30)
Date : May 21, 2023 - May 26, 2023
Although the prediction of system behavior plays a crucial role in accident management, there is very limited research on this subject in the nuclear industry. Accident process prediction methods are usually implemented as data-based methods due to the fast inference speed. However, the existing methods used for reactor accidents often ignore the predictive capability of the forecasting models over long-term patterns, or do not analyze the confidence of the prediction.
In this paper, we have proposed a method for accident process prediction based on a Temporal Fusion Transformer (TFT) model. On the one hand, the method leverages multiple types of covariates as auxiliary, such as static labels of accidents and other monitoring data, to improve the prediction accuracy; on the other hand, it carries out uncertainty assessment of the predicted sequences.
The method proposed in this paper is applied to MBLOCA post-accident prediction of HPR1000. Extensive results show that the method significantly outperforms the NiHiTS, Nbeats, Transformer, LSTM, GRU and RNN etc. models in terms of prediction accuracy on the test dataset. Furthermore, the uncertainty analysis of the prediction results allows the method to be applied to production scenarios of nuclear reactors, which is of high safety demands.
