Data-driven surrogate modeling for thermal-hydraulic codes via hybrid deep neural networks and quantile learning

Hyojun Yi, Hyeonmin Kim, Seunghyoung Ryu

Abstract Nuclear energy is a clean, reliable power source, but realizing its potential requires strict safety measures in nuclear power plants. Thermal-hydraulic (TH) codes are used to simulate potential accident scenarios in probabilistic safety assessment (PSA). Their high computational cost becomes particularly problematic in dynamic PSA, which requires repeated simulations across large numbers of accident scenarios. Consequently, deep learning-based surrogate models (e.g. conditional autoencoders and recurrent neural networks) have been proposed to accelerate TH simulations. However, there remains substantial room for improvement in predictive accuracy, particularly in capturing abrupt fluctuations and complex temporal patterns. To address these limitations, we propose a novel surrogate model named QLCT that combines long short-term memory, one-dimensional convolutional neural networks, and transformer with multiple quantile regression, enhanced by temporal differential supervision. We compared QLCT against various deep learning models, and the results demonstrate that the proposed model achieves substantial improvements: 22.0%–40.0% reduction in mean absolute error, 12.4%–34.9% reduction in root mean squared error, and 27.9%–49.0% reduction in normalized mean absolute error. In addition, QLCT also shows superior performance in probabilistic forecasting compared to the state-of-the-art model, achieving improvements of 25.8%, 21.9%, and 15.2% in the winkler score, pinball loss, and coverage width-based criterion, respectively.