Accident severity assessment method for nuclear power plants based on Temporal Convolutional Network (TCN) and Bayesian optimization

Abstract

This study proposes a novel severity assessment model for nuclear power plant (NPP) accidents based on the Temporal Convolutional Network (TCN) and Bayesian optimization. The model assesses the severity of NPP accidents by using deviations in key operational parameters during the initial phases as input. The model’s hyperparameters were optimized through the Bayesian optimization method. The proposed severity assessment model was developed using data from four types of NPP accidents: small break hot-leg LOCA, small break cold-leg LOCA, SGTR, and steam line break inside containment (SLBIC). The severity levels of these accident scenarios were determined by the pipe break size or the tube rupture degree. The model was trained and tested using simulated operation data generated by the PCTRAN simulator. For comparative analysis, models based on LSTM, BiLSTM, GRU, and CNN-LSTM were trained under identical conditions to the TCN model. Simulation experiments revealed that the proposed TCN model outperforms the other models, particularly for extrapolative testing dataset beyond the training dataset’s severity range. Furthermore, the robustness to noise of the models was evaluated using simulated data with added artificial Gaussian noise, demonstrating that the TCN-based model maintains superior performance in most cases, even in the presence of data noise.