Probabilistic deep learning model as a tool for supporting the fast simulation of a thermal–hydraulic code

Abstract

Following the Fukushima Daiichi accident, enhancing the safety of nuclear power plants has become the priority mission for the future of nuclear energy. Probabilistic safety assessment (PSA) is a well-known technique to quantify the anticipated risk of nuclear power plants according to the accident scenario. One approach to strengthening nuclear safety is reducing the uncertainty in PSA by analyzing a wide spectrum of accident scenarios. In doing this, massive simulations of thermal–hydraulic (TH) dynamics are required by running TH code. As the computation time for such large-scale simulation is a heavy burden, it is necessary to develop a fast simulation model, but related research has recently begun. For doing this, conditional autoencoder was firstly introduced, however, prediction accuracy can be further improved by exploiting temporal characteristics of simulation data. In this paper, we formalize deep learning-based fast simulation model of TH code, and propose a novel deep learning model, namely ensemble quantile recurrent neural network (eQRNN). By leveraging bi-directional long short-term memory, concatenative positional encoding, quantile regression, and model ensemble, the proposed eQRNN can provide a more accurate prediction on the simulation results and uncertain boundaries of its prediction. Compared to the base RNN model, the proposed eQRNN shows 39% and 28% lower error overall in terms of mean absolute percentage error (MAPE) and mean squared error (MSE). Finally, eQRNN achieved 75%, 79% lower MAPE and MSE than the existing conditional autoencoder-based model.