Predicting fuel rod failures and their contributing factors with surrogate support vector machines in statistical LOCA analysis

In this study, a surrogate support vector machine (SVM) model is used to predict fuel cladding failures and their causes in a large break loss-of-coolant accident (LB-LOCA). The training data of the surrogate model originates from a statistical fuel behaviour analysis of a LB-LOCA scenario in EPR-type nuclear power plant (Arkoma et al., Nuclear Engineering and Design Vol. 285, 2015). In the preceding analysis, the accident was simulated 59 times, as stated by the first order Wilks’ formula, with variation in the thermal hydraulic and power boundary conditions and model parameters between each of these global scenarios. A thousand rods were random sampled from the loading pattern, with fuel manufacturing parameter values varied, and then simulated with coupled fuel performance - thermal hydraulics subchannel code FRAPTRAN-GENFLO. The worst global scenario was this way identified.

In order to obtain an estimation of the number of failing rods in the whole reactor scale, an SVM is now trained using the data from the worst scenario. A surrogate model is needed in order to replace the computationally costly fuel performance code simulations, especially when the analysis is eventually broadened to consider other global scenarios and global parameters. The problem is that the number of failing rods is very limited in the subset of thousand simulated rods. Consequently, surrogate models produce inaccurate predictions, as demonstrated with neural networks (Arkoma et al., 2015). To cure the problem in rod failure recognition caused by the imbalance between the number failing and surviving rods, an iterative solution is applied. An SVM is trained with the existing small number of simulations, and then it is used in predicting a new set of rods which are susceptible to failure in LOCA. These rods are simulated with the fuel behavior code, and the resulting additional data is included in the training set of the SVM. It is demonstrated that this kind of procedure improves the predictions of the surrogate model with moderate increase in the number of additional fuel code simulations.

To discover the most significant varied parameters with respect to the fuel failures, a sensitivity analysis is performed. SVMs are applied in the sensitivity analysis, with comparisons to the sensitivity analysis results of actual fuel performance code simulations (Arkoma and Ikonen, Nuclear Engineering and Design Vol. 305, 2016).