Research on Process Diagnosis of Severe Accidents Based on Deep Learning and Probabilistic Safety Analysis

Abstract

Severe accident process diagnosis provides data basis for severe accident prognosis, positive and negative effect evaluation of Severe Accident Management Guidelines (SAMGs), especially to quickly diagnose Plant Damage State (PDS) for operators in the main control room or personnel in the Technical Support Center (TSC) based on historic data of the limited number of instruments during the operation transition from Emergency Operation Procedures (EOPs) to SAMGs. This diagnosis methodology is based on tens of thousands of simulations of severe accidents using the integrated analysis program MAAP. The simulation process is organized in reference to Level 1 Probabilistic Safety Analysis (L1 PSA) and EOPs. According to L1 PSA, the initial event of accidents and scenarios from the initial event to core damage are presented in Event Trees (ET), which include operator actions following up EOPs. During simulation, the time uncertainty of operations in scenarios is considered. Besides the big data collection of simulations, a deep learning algorithm, Convolutional Neural Network (CNN), has been used in this severe accident diagnosis methodology, to diagnose the type of severe accident initiation event, the breach size, breach location, and occurrence time of the initial event of LOCA, and action time by operators following up EOPs intending to take Nuclear Power Plant (NPP) back to safety state. These algorithms train classification and regression models with ET-based numerical simulations, such as the classification model of sequence number, break location, and regression model of the break size and occurrence time of initial event MBLOCA. Then these trained models take advantage of historic data from instruments in NPP to generate a diagnosis conclusion, which is automatically written into an input deck file of MAAP. This input deck originated from previous traceback efforts and provides a numerical analysis basis for predicting the follow-up process of a severe accident, which is conducive to severe accident management. Results of this paper show a theoretical possibility that under limited available instruments, this traceback and diagnosis method can automatically and quickly diagnose PDS when operation transit from EOPs to SAMGs and provide numerical analysis basis for severe accident process prognosis.