Optimization Of Probabilistic Safety Margin Analysis Method For Pressurized Water Reactor Nuclear Power Plant

It is a major scientific issue for the sustainable development of nuclear energy to understand the factors and ways that lead to nuclear accidents scientifically,and to take design,operation and management measures to protect workers,the public and the environment under improper radiation hazards.Traditional nuclear power plant safety analysis methods include deterministic safety analysis and probabilistic safety analysis,but both of them have certain limitations in the probabilistic safety margin(PSM)analysis of small design changes of nuclear power plant.The risk-informed safety margin characterization(RISMC)method effectively combines the two methods,makes up for the respective defects of the two methods,and can achieve accurate analysis of the PSM of nuclear power plant.Currently,RISMC has been successfully applied to nuclear power plant safety analysis.However,the overall efficiency of RISMC analysis is low,which requires a lot of computing resources and computing time.In this thesis,an optimized probabilistic safety margin analysis method(OPSM)based on optimization algorithm and deep learning is proposed to solve the problem of low efficiency of RISMC analysis method,which significantly improves the analysis efficiency of PSM.Firstly,this thesis systematically summarizes the theoretical framework and mathematical model of the traditional RISMC method.On this basis,the PSM analysis model of small break loss of coolant accident(SBLOCA)at 100% and 105% power of typical double-loop pressurized water reactor is modeled,and the key problems in the modeling process are analyzed deeply.According to the RISMC mathematical model,the three key factors affecting the efficiency of PSM analysis are determined accurately: the number of branches of accident sequence,the sample size of single accident sequence analysis and the time of accident simulation calculation.On this basis,a general OPSM analysis method is proposed,which provides an optimization scheme for the above three factors.Secondly,in order to achieve accurate and efficient branch simplification of accident sequence,this thesis proposes a method of branch number simplification of accident sequence.In this method,the key safety parameters of all accident sequences with design parameters are analyzed and judged by the expert knowledge,and the accident sequence pre-classification is realized.Aiming at the accident sequences which are difficult to be determined by qualitative analysis,the limit state of the accident sequences is accurately solved by quantitative calculation method.Due to the complex nonlinear problems difficult to be solved by traditional numerical methods,this thesis uses hybrid particle swarm optimization algorithm to find the optimal solution.Finally,according to the probability cutoff method,all the accident sequences that may cause core damage are further screened.This method can effectively reduce the number of accident sequence branches that need to be calculated by best estimate plus uncertainty(BEPU),concentrate the computing resources on the probabilistic failure sequence,and greatly reduce the computation amount while ensuring the accuracy.Finally,in order to reduce the calculation time of accident simulation,the surrogate model based on deep learning algorithm is used to replace the deterministic safety analysis program.This model integrates convolutional neural network and long-short term memory neural network,which can extract local features of data and learn time series information well at the same time,and has good prediction performance for long time series accident process.In order to determine a surrogate model with higher relative accuracy,the key hyperparameters are analyzed by combining random search and grid method.In order to accurately predict the failure probability of the accident sequence,the adaptive sampling method is used instead of the traditional sampling method.This sampling method can better concentrate the learning data near the safety boundary and reduce the sample size of single accident sequence analysis while ensuring the accuracy of the model.In order to improve the modeling efficiency of all accident sequences,a model-based transfer learning model is integrated in this thesis.This model can reduce the sample size required by the surrogate model of accident sequence with high similarity to a certain extent and improve the overall modeling efficiency.Based on the above work content,the OPSM analysis platform system is designed to realize the rapid and accurate analysis of PSM.Based on this system,the probabilistic safety margin of SBLOCA under 100% and 105% power of a typical two-loop pressurized water reactor is analyzed.By branch simplification based on expert knowledge and hybrid particle swarm optimization algorithm,the accident sequence under two conditions is reduced to 3 and5.On this basis,a hybrid neural network is used to model and predict 8 accident sequences.Finally,the accuracy and efficiency of traditional probabilistic safety margin analysis method and OPSM analysis method are compared.The results show that a small power increase has an obvious effect on the reactor PSM.OPSM analysis method not only has high analysis accuracy,but also significantly improves the analysis efficiency.