Power distribution reconstruction from in-core detector measurements using gaussian process regression

In order to safely and efficiently operate pressurized water reactors, it is essential to monitor neutron flux and power distributions throughout the operation cycle. Most of currently operated reactors have in-core instrumentation systems capable of taking assembly-wise axially integrated flux measurements. However, commercial reactors typically possess few fuel assemblies that contain guide tubes for in-core instrumentation and it is necessary to estimate power values at the remaining fuel assemblies. This work employs gaussian process regression, a non-parametric supervised learning method, to predict the power distribution over the entire core using only measured data. The OpenMC Monte Carlo code was employed to emulate detector signals from instrumented assemblies as well as to produce reference values at non-instrumented assemblies against which the regression method was validated. The mean and maximum relative discrepancies were below 0.8% and 2.5%, respectively, in a scenario with full detector availability. Considering the failure of a single detector, the maximum relative discrepancy did not exceed 5.3%, showing the feasibility of the model as an operation monitoring system capable of real-time reconstruction of the core power distribution.