Research On Source Terms Inversion Of Nuclear Accident Based On Deep Learning

After the nuclear accident,the diffusion of air-borne radionuclides will results in the most rapid and serious radiation consequences.The evaluation of the consequences is an important task of the emergency response and decision support system.Generally,the radiation consequences are estimated by the atmospheric diffusion model,which is based on the source terms and meteorological information.The facts of the serious nuclear accident in Chernobyl and Fukushima show that the explosion of the reactor plant and the power outage will lead to the unavailability of the monitoring instrument data,thus there is great uncertainty in the application of forward source estimation method to nuclear accident.Therefore.source inversion based on environmental radioactive monitoring data and meteorological information has been widely concerned and studied.In this paper,the release rate of Iodine 131(I-131)is selected as the object of the source terms inversion,and the radionuclide atmospheric diffusion codes RADC developed by our research team is used to generate a large amount of data required for neural network training and testing.On this basis.this paper does research into the source terms inversion capabilities of different neural networks.First of all,this paper studies the effect of BP neural network for source terms inversion,and uses genetic algorithm to optimize the initial parameters of BP neural network to improve its learning ability.Then,this paper establishes an deep feedforward neural network with 5 hidden layer based on open source deep learning framework TensorFlow,its optimal hyperparameters values are determined through many experiments.Finally,the uncertainty of the predicted value caused by the error of the neural network input data is analyzed using the Bayesian MCMC method.The research results show that:The average error of BP neural network for predicting nuclide release rate is 12.47%.which can be reduced to 5.23%after genetic algorithm optimization.The deep feedforward neural network used in the source term inversion shows that the average prediction error is about 2.04%and the prediction errors corresponding to more than 80%of the test data are less than 2.0%.Compared with BP neural network.the prediction error of deep feedforward neural network is much smaller,indicating that it has stronger learning ability and generalization ability.The uncertainty analysis can obtain the confidence interval and risk curve of the predicted value corresponding to different input error ranges,in order to provide more reliable source terms information for accident emergency and dicision-making.