基于贝叶斯神经网络的冷却剂丧失事故不确定性预测方法OA

For traditional best estimate plus uncertainty(BEPU)methods are constrained by long computational times and limited sampling sizes.In this regard,a simulation model of AP1000 under cold leg break accident conditions is used by RELAP5 to generate training data.A Bayesian neural network(BNN)is established as a surrogate model for predicting uncertainties in peak cladding temperature(PCT)and core water level variations during loss of coolant accident(LOCA)accidents.Additionally,DAKOTA coupled with RELAP5 is employed for uncertainty analysis to validate the feasibility and advantages of the BNN model in this context.The analytical results demonstrate that:the error between the PCT pre-dicted by the BNN model and the actual value is within 6%,while the error in core water level predictions is below 1%;the training and computational time of the BNN model(5 minutes)is significantly shorter than that of traditional BEPU methods(17 hours),indicating high computational efficiency;the uncertainty analysis results obtained from BNN predic-tions are in close agreement with those from DAKOTA-RELAP5 coupled calculations,confirming the reliability of BNN in uncertainty prediction;using standard deviation to quantify output uncertainty,the magnitude of uncertainty predicted by BNN does not exceed 5%of the mean value,indicating low uncertainty in the established simulation model.The findings of this study provide a reference for subsequent direct application of neural network models in uncertainty prediction.