Reliability Analysis Of Ap1000Passive System Based On Artificial Neural Networks

The public have had a higher standards on the security and cost-effective of nuclearpower plants these years, because the cost uncertainty of the disabled nuclear powerplants and the fear of nuclear leakage accident. The passive safety design has provided asolution. The passive safety design, by simplifying the device and reducing dependenceon external inputs, has improved the safety and economic efficiency of nuclear powerplants. But for the significant difference of action, the Probabilistic Safety Assessment(PSA) methods are not apply to the passive nuclear power plants, which is widely usedin safety analysis of traditional nuclear power plants. It is of great importance toconduct specialized research on reliability of passive nuclear system, for it is quiteinadequately.In this paper, we have studied the safety assessment methods of Passive ResidualHeat Removal system (PRHRs) of AP1000, which is one of the most developed passivenuclear plants. We have analyzed the framework of passive reliability analysis ofcurrently research, and innovative apply the fast-running surrogate of thethermal-hydraulic (T-H) system code to deal with the prohibitive computational burdenin estimating the failure probability of a T-H passive system. We used the artificialneural networks (ANNs) as a replace of the T-H code, and the quadratic responsesurface (RS) empirical regression as compared.Firstly, according to the data from AP1000Design Control Document (DCD), weestablished the T-H model of Relap5code, and debugged the steady-state. The modelwas appropriately simplified to the analysis of PRHRs. Parameters was set for the studyof loss of feedwater flow.Secondly, the relevant parameters PRHRs were analyzed using the analytichierarchy process (AHP) method to determine the most important structure parametersand thermal parameters, determine the distribution of the input parameters and sampling.The input parameters were computed by T-H model to obtain initial data sample.Finally, the initial sample data was used to train the ANNs and RS model, replacingthe original best-estimate T-H code. We also studied the impact of the neural networkapproach network structure (activation function, number of hidden layer, hidden layer nodes, etc.), and compared the size of training data sample on the simulation results.The bootstrap method was also employed to correct the output parameters and provideconfidence intervals. In the last part of the paper, the time efficiency of differentmethods was compared.