Research on core thermal hydraulic parameters prediction based on the improved GAN method and combined ANN model

Abstract

This research presents advanced methods of data processing as tools applicable in nuclear engineering. Three methods-autoencoder, random forest and multilayer perceptron were considered for the testing. The multi-layer perceptron (MLP) method preserved best the structure of the data. A better generative adversarial network (GAN) based on the Wasserstein distance was implemented for data generation, which overcomes the issues of gradient vanishing and mode collapse prevailing in common GANs. While examining the generated data, advanced statistical and machine learning techniques were applied to minutely compare the generated and original data. Data forecasting applied a combined model of MLP, convolutional neural network (CNN), and recurrent neural networks (RNN). With limited-memory broyden-fletcher-goldfarb-shanno (LBFGS) optimization algorithm being combined with bayesian optimization, there was significant improved prediction of core thermal hydraulic parameters. Meanwhile, this research provides important techniques to deal with challenges of nuclear engineering data which further impacts the field of nuclear engineering.