Using Deep Learning for Nuclear Reactor Fault Diagnosis
Using Deep Learning for Nuclear Reactor Fault Diagnosis
-
DOI: https://doi.org/10.22533/at.ed.1317672622065
-
Palavras-chave: LOCA, IFA-650-10, CNN, GRU, LSTM.
-
Keywords: LOCA, IFA-650-10, CNN, GRU, LSTM.
-
Abstract: Loss-of-Coolant Accidents (LOCA) are among the most important design-basis accidents in light-water reactors and have been the focus of nuclear safety research. Accurate predictions of loss-of-coolant accidents (LOCAs) are critical to continued safe and efficient operation. Predicting LOCA is very challenging due to the dynamics and intricate relations involved. Data-driven predictive methods offer considerable promise for improving accident management and operator support in Nuclear Power Plants (NPPs). Machine Learning (ML)- based models can detect early signs of accidents and set initial conditions that enable better responses and reduce manual intervention. Currently, several predictive models are inaccurate, hindering effective decision-making and the prevention of severe accidents. Among the solutions adopted, exploring various ML methods can enable robust, real-time fault diagnosis. ML-based models can detect the earliest signs of accidents and establish initial conditions that enable better responses and reduce the need for manual intervention. In LOCA scenarios, the fuel cladding suffers a series of physical processes, including high-temperature creep and oxidation. Under transients, a significant amount of radioactive fission products is released. Cladding leads to ballooning and, in some cases, rupture. Therefore, accurate analysis and simulation of cladding deformation, followed by ballooning and fuel relocation, are significant for reactor safety assessment. Based on the dataset produced in the IFA-650-10 from the Halden experiment.
- Daniel de Souza Gomes