An approach based on Support Vector Machines and a K-D Tree search algorithm for identification of the failure domain and safest operating conditions in nuclear systems

The safety of a Nuclear Power Plant (NPP) is verified by analyzing the system responses under normal and accidental conditions. This is done by resorting to a Best-Estimate (BE) Thermal-Hydraulic (TH) code, whose outcomes are compared to given safety thresholds enforced by regulation. This allows identifying the limit-state function that separates the failure domain from the safe domain. In practice, the TH model response is affected by uncertainties (both epistemic and aleatory), which make the limit-state function and the failure domain probabilistic. The present paper sets forth an innovative approach to identify the failure domain together with the safest plant operating conditions. The approach relies on the use of Reduced Order Models (ROMs) and K-D Tree. The model failure boundary is approximated by Support Vector Machines (SVMs) and, then, projected onto the space of the controllable variables (i.e., the model inputs that can be manipulated by the plant operator, such as reactor control-rods position, feed-water flow-rate through the plant primary loops, accumulator water temperature and pressure, repair times, etc.). The farthest point from the failure boundary is, then, computed by means of a K-D Tree-based nearest neighbor algorithm; this point represents the combination of input values corresponding to the safest operating conditions. The approach is shown to give satisfactory results with reference to one analytical example and one real case study regarding the Peak Cladding Temperature (PCT) reached in a Boiling Water Reactor (BWR) during a Station-Black-Out (SBO), simulated using RELAP5-3D.