Research On Optimization Of Nuclear Accident Source Term Inversion Model Based On Bp Neural Network

Three serious accidents in the history of nuclear power development have alarmed people.Rapid and accurate prediction and evaluation of accident consequences can provide scientific basis for decision-making of nuclear emergency measures.Among them,the estimation of accident source term under reactor accident condition is the core link of nuclear accident consequence evaluation.It is a common method to estimate the source term based on the readings of reactor monitoring instruments or accident safety analysis reports.However,serious accidents will affect the accuracy of in-reactor instrumentation,and the preset types of accidents cannot cover all the real situations.It is an assistant way to estimate the internal source term of the reactor under accident conditions to invert the source term of the accident by combining the outfield monitoring data with the atmospheric diffusion model of radioactive materials and using appropriate numerical algorithm.BP neural network with good self-adaptive,self-learning and non-linear mapping ability is an effective algorithm to solve this problem.In this paper,the problems of BP neural network inversion model of nuclear accident source terms,such as poor generalization ability caused by strong dependence on input samples,slow learning speed,and easy to fall into local minimum,are optimized to provide reference for the evaluation and decision-making of nuclear accident consequences.There are many factors affecting the consequences of accidents under accident conditions.Therefore,in the process of source term inversion of nuclear accidents,release height,mixed layer height,wind speed,wind direction,precipitation type,atmospheric stability,environmental dose rate at different distances from the core and other factors must be taken as input samples of BP neural network training.In order to solve the problem that the training result of neural network is closely dependent on the input samples and the approximation and generalization ability of the network model largely depends on the typicality of the samples,three feature processing methods,principal component analysis,average impact value algorithm and random forest algorithm,are used to pretreat the influence factors of the source term respectively.BP neural network is constructed with MATLAB.Source term inversion model of collateral nuclear accident.The processed data are used as input samples of BP neural network and the release rate of I-131 is used as target output of BP neural network.39200 sets of data were obtained by Inter RAS,and 30 000 of them were selected for neural network training.The remaining 9200 sets of data were used for testing.The results show that compared with the traditional BP neural network,the BP neural network inversion model optimized by feature processing method has less training time,15% higher inversion speed on average,and 11% lower test relative error,and the model has good stability.BP neural network model is sensitive to initial weights and threshold parameters,easy to fall into local minimum and long convergence time.In this study,we choose evolutionary thinking algorithm and quantum genetic algorithm to improve the initial weights and thresholds of the network,which have both the ability of global search and local optimization,and the ability of fast convergence and global optimization,so as to improve the shortcomings of BP neural network algorithm.The results show that compared with BP neural network,the average inversion speed is increased by 16%,the relative error of the test is reduced by 24%,and the stability and accuracy of the inversion model are improved.When an accident occurs,the model can be used as a way to obtain the source information of the accident and provide decision support for nuclear emergency response.