ScholarWorks@중앙대학교

초록

Fast and reliable simulation of thermal-hydraulic (TH) dynamics is crucial for probabilistic safety assessment of nuclear power plants. However, conventional high-fidelity TH codes are computationally expensive, which limits their use for large-scale scenario analysis. Surrogate models provide an efficient alternative by approximating simulation outputs at a fraction of the computational cost. In this study, we propose PQNet, a deep learning-based surrogate that integrates the MLP-Mixer architecture with a random quantile regression (RQR) strategy. By treating the quantile level as an input, the model enables multivariate and multi-quantile forecasting without requiring separate models for each target variable. PQNet was validated using MAAP simulation data for three representative initiating events—station blackout, small-loss-of-coolant accident, and medium-loss-of-coolant accident. It consistently outperformed state-of-the-art models including TCN, and QRNN, achieving 35~58% improvement in normalized mean absolute error. Overall, PQNet provides a unified network that predicts arbitrary quantile values with better accuracy on multivariate TH output.

키워드