Comparison of Intelligent Optimization Algorithms in Neural Network Model for Nuclear Accident Source Term Evaluation

Abstract

Abstract After a severe nuclear accident, the source term cannot be obtained by the forward method due to the power outage of the whole plant. But the source term needs to be estimated quickly and accurately to protect the public and the environment. The source term inversion method based on the backpropagation neural network can effectively estimate the source term, but it is prone to fall into local minimum. To solve this problem, many intelligent optimization algorithms have been proposed to optimize the weights and thresholds of neural networks. In order to compare the effectiveness of different intelligent optimization algorithms, this paper builds neural network models based on Whale Optimization Algorithm (WOA), Genetic Algorithm (GA), Particle Swarm Optimization Algorithm (PSO), Simulated Annealing-Genetic Algorithm (SAGA) and Mindset Evolution Algorithm (MEA). And the Wind diagnose model and air dispersion model (CALMET-LAPMOD) was used to generates a large amount of data for neural network training near Sanmen Nuclear Power Plant. In addition, this study uses kernel principal component analysis (KPCA) to reduce data dimensionality and training time. The results show that the model using the WOA has the best source term estimate ability, and the relative error of test data is only 4.31%. Therefore, the WOA-KPCA neural network is recommended in similar nuclear accident emergency or consequence evaluation scenarios by this study.